WO2023121798A1

WO2023121798A1 - Maintenance of modules for light sources in semiconductor photolithography

Info

Publication number: WO2023121798A1
Application number: PCT/US2022/050261
Authority: WO
Inventors: Nathan Gibson WELLS; Christopher James STEVENS; Deepthi MYSORE NAGARAJ
Original assignee: Cymer, Llc
Priority date: 2021-12-23
Filing date: 2022-11-17
Publication date: 2023-06-29

Abstract

A light source includes one or more modules that together are configured to produce a light beam for semiconductor photolithography. An apparatus for maintaining the light source includes: a prediction unit and an ensemble unit. The prediction unit is configured to: receive a plurality of sub-feature datasets, each sub-feature dataset associated with a unique performance criterion of the light source; and for each sub-feature dataset, evaluate whether a failure mode is detected in the light source in a prediction increment. The ensemble unit is configured to: receive the plurality of evaluations from the prediction unit; and for each failure mode, determine which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode.

Description

MAINTENANCE OF MODULES FOR LIGHT SOURCES IN SEMICONDUCTOR PHOTOLITHOGRAPHY

CROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Application No. 63/293,453, filed December 23, 2021, titled MAINTENANCE OF MODULES FOR LIGHT SOURCES IN SEMICONDUCTOR PHOTOLITHOGRAPHY, which is incorporated herein in its entirety by reference.

TECHNICAL FIELD

[0002] The disclosed subject matter relates to maintenance of light sources such as those used for integrated circuit photolithographic manufacturing processes.

BACKGROUND

[0003] Light, which can be laser radiation, that is used for semiconductor photolithography is typically supplied by a system referred to as a light source. These light sources produce radiation as a series of pulses at specified repetition rates, for example, in the range of about 500 Hz to about 6 kHz. Additionally, such light sources conventionally have expected useful lifetimes measured in terms of the number of pulses they are projected to be able to produce before requiring repair or replacement, typically expressed as billions of pulses.

[0004] One system for generating light such as laser radiation at frequencies useful for semiconductor photolithography (such as at deep-ultraviolet (DUV) wavelengths) involves use of a master oscillator power amplifier (MOP A) dual-gas-discharge-chamber configuration. This configuration has two chambers, a master oscillator chamber (MO chamber) and a power amplifier chamber (PA chamber). These chambers and many other system components can be regarded as being modules, and the light source overall can be regarded as an ensemble of modules. Each module in general has a lifetime that is shorter than the lifetime of the overall light source. Thus, over the course of the lifetime of the light source, the health of individual modules is evaluated to determine whether the modules should be repaired or replaced.

SUMMARY

[0005] In some general aspects, an apparatus maintains a light source including one or more modules that together are configured to produce a light beam for semiconductor photolithography. The apparatus includes a prediction unit and an ensemble unit. The prediction unit is configured to: receive a plurality of sub-feature datasets, each sub-feature dataset associated with a unique performance criterion of the light source; and for each sub-feature dataset, evaluate whether a failure mode is detected in the light source in a prediction increment. The ensemble unit is configured to: receive the plurality of evaluations from the prediction unit; and, for each failure mode, determine which one or more performance criterion is related to the failure mode based on the associated subfeature dataset of that failure mode.

[0006] Implementations can include one or more of the following features. For example, the prediction unit can include or access a plurality of models, each model being generated by machine learning using a unique sub-feature dataset. Each of the models in the prediction unit can have the same fitted structure. Each of the models in the prediction unit can have a different fitted structure. At least one of the models in the prediction unit can have a fitted structure that is distinct from the fitted structure of another model in the prediction unit.

[0007] The plurality of sub-feature datasets can include a sub-feature dataset relating to bandwidth of the light beam, a sub-feature dataset relating to wavelength of the light beam, and a sub-feature dataset relating to energy of the light beam. The plurality of sub-feature datasets can include a subfeature dataset relating to long term bandwidth of the light beam, a sub-feature dataset relating to short term bandwidth of the light beam, a sub-feature dataset relating to bandwidth error events, a sub-feature dataset relating to long term wavelength of the light beam, a sub-feature dataset relating to short term wavelength of the light beam, a sub-feature dataset relating to wavelength error events, a sub-feature dataset relating to long term energy of the light beam, a sub-feature dataset relating to short term energy of the light beam, and a sub-feature dataset relating to energy error events.

[0008] The ensemble unit can be configured to determine which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset by aggregating the evaluations from the prediction unit, and prioritizing reasons for failure that are associated with each detected failure mode. The ensemble unit can be configured to aggregate evaluations from the prediction unit based on the performance criterion associated with each sub-feature dataset. The ensemble unit can be configured to label a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected. The ensemble unit can be configured to label the performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.

[0009] The prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment can include evaluating whether the failure mode is detected in a specific module of the light source based on a model that is trained for the specific module. The specific module of the light source can include a master oscillator module. The prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment can include evaluating whether the failure mode is detected in one or more specific modules of the light source based on models that are each trained for a specific module. The one or more specific modules of the light source can include one or more of: a master oscillator module, a power amplifier module, a line narrowing module, a spectral feature analysis module, and a pulse stretcher module. The prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment can include: separately evaluating each subfeature dataset according to a model that is generated using machine learning using that sub-feature dataset; and, for each sub-feature dataset evaluation, outputting a score related to a probability of failure. The score can correspond to a certainty score, a binary output, or a function in a range between 0 and 1.

[0010] The prediction unit can be configured to, for each sub-feature dataset, evaluate whether a failure mode is detected in the light source in the prediction increment by performing an evaluation each time an evaluation increment has passed. An evaluation increment can be 100 million pulses. [0011] In other general aspects, an apparatus maintains a light source include one or more modules that together are configured to produce a light beam for semiconductor photolithography. The apparatus includes a pipeline unit, a prediction unit, and an ensemble unit. The pipeline unit is configured to: receive a plurality of sub-feature datasets, each sub-feature dataset relating to a unique performance criterion of the light source during operation; and categorize each of the received subfeature datasets into a feature category based on its related unique performance criterion. The prediction unit is configured to: receive the plurality of sub-feature datasets and assigned categories for each sub-feature dataset; and evaluate whether a failure mode is detected in the light source in a prediction increment. The ensemble unit is configured to determine which performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset. [0012] Implementations can include one or more of the following features. For example, the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment can include evaluating whether the failure mode is detected in a specific module of the light source based on a plurality of models that are trained for the specific module.

Each model can be developed through machine learning by supplying a respective sub-feature dataset to train the model. The apparatus can also include an alert generating unit configured to instruct a maintenance operation on the specific module if the evaluation determines that the failure mode is detected in the specific module.

[0013] The prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment can include: separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and, for each sub-feature dataset evaluation, outputting a score related to a probability of failure. The ensemble unit being configured to determine which performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset can include analyzing the score from the sub-feature datasets related to the same assigned category. The score can be a certainty score, a binary output, or a function in a range between 0 and 1.

[0014] The prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment can include determining whether at least one module within the light source does not have at least the minimum probability of operating without a failure in the prediction increment. The prediction increment can be measured as a number of pulses of the light beam.

[0015] Prior to each evaluation, the pipeline unit can be further configured to aggregate, in usage, data of at least one sub-feature dataset over one or more look back increments that precede the evaluation. The prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment can include separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset, and, if the sub-feature dataset is aggregated, in usage, over one or more look back increments, then using the aggregated sub-feature dataset. Each look back increment can be selected from a set of possible look back increments that relate to how performance criteria change over usage. The set of possible look back increments can include a first look back increment of about 2 billion pulses, a second look back increment of about 1 billion pulses, and a third look back increment of about 100 million pulses. One or more of the sub-feature datasets can relate to a performance criterion that is tracked in a relatively short look back increment and one or more of the sub-feature datasets can relate to a performance criterion that is tracked in a relatively long look back increment.

[0016] The pipeline unit can be configured to aggregate in usage one or more sub-feature datasets based on the same look back increment, and the prediction unit can be configured to receive the plurality of sub-feature datasets by receiving an aggregation of sub-feature datasets that are tracked using the same look back increment.

[0017] In other general aspects, a method is performed for maintaining a light source including one or more modules that together are configured to produce a light beam for semiconductor photolithography. The method includes: receiving a plurality of sub-feature datasets, each sub-feature dataset associated with a unique performance criterion of the light source; for each sub-feature dataset, evaluating whether a failure mode is detected in the light source in a prediction increment; and, for each detected failure mode, determining which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode.

[0018] Implementations can include one or more of the following features. For example, the method can further include, after receiving the plurality of sub-feature datasets, categorizing each of the received sub-feature datasets into a feature category based on its related unique performance criterion. And, determining which one or more performance criterion is related to the failure mode can include determining which performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset.

[0019] Evaluating whether a failure mode is detected in the light source in a prediction increment can include: separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and, for each sub-feature dataset evaluation, outputting a score related to a probability of failure. Evaluating whether a failure mode is detected in the light source in the prediction increment can include performing an evaluation each time an evaluation increment has passed.

[0020] Determining which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode can include aggregating the evaluations, and prioritizing reasons for failure that are associated with each detected failure mode. Evaluations can be aggregated by aggregating the evaluations based on the performance criterion associated with each sub-feature dataset. The method can further include labelling a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected. The method can also include labelling the performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.

[0021] Evaluating whether the failure mode is detected in the light source in the prediction increment can include evaluating whether the failure mode is detected in a specific module of the light source based on a model that is trained for the specific module. Evaluating whether the failure mode is detected in the light source in the prediction increment can include evaluating whether the failure mode is detected in one or more specific modules of the light source based on models that are each trained for a specific module.

[0022] The method can also include, prior to each evaluation, aggregating in usage data of at least one sub-feature dataset over one or more look back increments that precede the evaluation. Aggregating in usage data of at least one sub-feature dataset over one or more look back increments can include aggregating in usage one or more sub-feature datasets based on the same look back increment, and receiving the plurality of sub-feature datasets can include receiving an aggregation of sub-feature datasets that are tracked using the same look back increment.

[0023] In other general aspects, a method is performed for maintaining a light source including one or more modules that together are configured to produce a light beam for semiconductor photolithography. The method includes: receiving a plurality of sub-feature datasets, each sub-feature dataset associated with a unique performance criterion of the light source; for each sub-feature dataset, evaluating whether a failure mode is detected in the light source in a prediction increment including separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and, for each detected failure mode, determining which one or more performance criterion is related to the failure mode based on the associated subfeature dataset of that failure mode. Determining includes aggregating the evaluations, and prioritizing reasons for failure that are associated with each detected failure mode.

[0024] Implementations can include one or more of the following features. For example, evaluations can be aggregated by aggregating the evaluations based on the performance criterion associated with each sub-feature dataset. The method can further include labelling a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.

DESCRIPTION OF DRAWINGS

[0025] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate the present invention and, together with the description, further serve to explain the principles of the present invention and to enable a person skilled in the relevant art(s) to make and use the present invention. One or more of the drawings may not be to scale.

[0026] Fig. 1A is a block diagram of a maintenance apparatus configured to maintain modules of a light source that produces a light beam for use by an output apparatus, and to provide advance notice of module failure, the maintenance apparatus including a prediction unit and an ensemble unit;

[0027] Fig. IB is a block diagram of an implementation of the maintenance apparatus that includes a pipeline unit;

[0028] Fig. 2 A is a block diagram of a training pipeline that shows development of machine learning models that are accessed by and used by the maintenance apparatus of Figs. 1 A and IB;

[0029] Fig. 2B is a block diagram of an operating pipeline that shows how the machine learning models are accessed by the prediction unit during operation of the light source of Figs. 1A and IB;

[0030] Fig. 3A is a diagram illustrating a maintenance procedure in terms of usage according to some implementations ;

[0031] Fig. 3B is a diagram illustrating an example of three look back increments at an evaluation point for three sub-feature datasets according to some implementations;

[0032] Fig. 4 is a block diagram of an implementation of the light source of Fig. 1 A;

[0033] Fig. 5 is a block diagram of an implementation of the output apparatus of Fig. 1 A;

[0034] Fig. 6 is a flow chart of a procedure performed by the maintenance apparatus of Fig. 1 A or IB for maintaining modules of the light source;

[0035] Fig. 7A is a table showing operational data used by the maintenance apparatus that is parsed into nine sub-feature datasets, with each sub-feature dataset being associated with a unique performance criterion of the light source and being associated with a machine learning model according to some implementations; and

[0036] Fig. 7B is a table showing a set of evaluations taken at evaluation points for each machine learning model of Fig. 7A.

DESCRIPTION

[0037] Referring to Fig. 1 A, a light source 100 generates an amplified light beam 105 for use by an output apparatus 110, which can be a photolithography apparatus. The light source 100 includes a plurality of modules 115-m, where m is a set of numbers from 1 to a positive plural integer. In the example shown in Fig. 1 A, m = 1, 2, 3, 4, 5, the plurality of modules 115-m includes modules 115-1, 115-2, 115-3, 115-4, 115-5, and the light source 100 overall can be regarded as an ensemble of such modules 115-1, 115-2, 115-3, 115-4, 115-5. While five modules are shown in Fig. 1A (namely, 115-1, 115-2, 115-3, 115-4, 115-5), it is possible for the light source 100 to have fewer than five or more than five modules. Each module 115-m in general has a lifetime that is shorter than the lifetime of the overall light source 100. Thus, over the course of the lifetime of the light source 100, the health of individual modules 115-m is evaluated to determine whether and which of the modules 115-m should be repaired or replaced. The timing for maintenance of modules 115-m is determined according to a maintenance strategy implemented by a maintenance apparatus 120.

[0038] In one example in which the light source 100 is a dual-stage deep ultraviolet light source (as discussed below with reference to Fig. 4), modules 115-m include a master oscillator (MO) chamber module, a power amplifier (PA) chamber module, a line narrowing module (LNM), a line analysis module (LAM), a bandwidth analysis module (BAM), and an optical pulse stretcher (OPuS) module. In general, any module for which data are collected and related to performance of the module can be considered a module 115-m.

[0039] The maintenance strategy implemented by the maintenance apparatus 120 is a predictive maintenance strategy designed to monitor the condition of in-service equipment (that is, the modules 115-m) to predict when a module 115-m will fail and also predict which of the modules 115-m will fail. The future behavior/condition of machine components is approximated, which makes it possible to optimize maintenance tasks (for example, prognostic health monitoring). Accordingly, machine downtime and maintenance costs can be reduced significantly while undertaking maintenance as infrequently as possible. The maintenance strategy implemented by the maintenance apparatus allows advance detection of pending failures and enables timely pre-failure interventions, utilizing prediction tools based on historical data.

[0040] The maintenance apparatus 120 employs machine learning (ML) models that utilize sensical, logical, and physics-based datasets that together form an entire feature dataset 140 to help derive information relating to causality of failure and thus determine which of the modules will fail in the future. Each dataset within the feature dataset 140 is referred to as a sub-feature dataset 140-sf, where sf denotes a unique criterion of performance (aka performance criterion)) of the light source 100. Thus, in the example of Fig. 1A, there are 8 sub-feature datasets 140-1, 140-2, 140-3, ... 140-8. It is possible for the feature dataset 140 to include fewer than 8 or more than 8 sub-feature datasets. Each sub-feature dataset 140-sf targets a specific aspect of the performance (that is, the unique performance criterion) of the light source 100. To put it another way, each sub-feature dataset 140-sf is curated to model a specific aspect of the light source 100. Each sub-feature dataset 140-sf corresponds to a set of data relating to operation of the light source 100. Moreover, each machine learning model is generated using machine learning and based on a sub-feature training set that corresponds to a specific subfeature dataset 140-sf. Thus, each model is individually and separately trained with a unique subfeature training set corresponding to the specific sub-feature dataset 140-sf that is piped into that particular model to evaluate the light source 100 during standard operation of the light source 100 (that is, while the light source 100 is producing the amplified light beam 105 for use by the output apparatus 110).

[0041] During standard operation, the maintenance apparatus 120 evaluates or calculates a probability for each model, and can use that information to separately determine whether a specific module 115-m in the light source 100 will fail in the future. The maintenance apparatus 120 can make this determination regarding a specific module 115-m because of the pipeline that is formed between each performance aspect of the light source and each evaluation by way of the sub-feature training set that is used to train the model. In this way, the maintenance apparatus 120 is able to determine not only that there will be a failure somewhere in the light source 100 in the future, but is also able to indicate how to adjust the light source 100 to prevent the failure, and specifically can target which module 115-m needs to be acted on. The maintenance apparatus 120 can pinpoint the problem within the light source 100 by relying on this pipeline in which each machine learning model is trained on a unique sub-feature training set and then using that machine learning model for the sub-feature dataset 140-sf that corresponds to that unique sub-feature training set that was used to train that model.

[0042] In operation, some modules 115-m might influence particular performance aspects of the light source 100 more than other modules 115-m. For example, using the example of the dual-stage deep ultraviolet light source 400 (as discussed below with reference to Fig. 4), operation of the MO chamber module 461 influences several performance aspects of the light source 100, such as a bandwidth, a wavelength, and an energy of the amplified light beam 105. This is because the MO chamber module 461 is the first module in the light source 400 that generates optical energy, and the other modules in the light source are optically downstream of the MO chamber module 461. For example, if the amplified light beam 405 output from the light source 400 is not producing enough optical energy, and the output light beam from the MO chamber module 461 has an energy that is below a threshold energy, then it can be assumed that the MO chamber module 461 is not operating at an optimum or acceptable level. If, however, both the output light beam from the MO chamber module 461 and the output light beam from the PRA chamber module 466 have acceptable energy levels, but the amplified light beam 405 output from the light source 400 is not producing enough energy, then it can be assumed that there is a problem with a module (such as one of the modules 476, 477, or 478) that follows (optically) the PRA chamber module 466. On the other hand, operation of the OpuS module 477 has very little influence on the energy of the amplified light beam 105.

[0043] To this end, a particular ML model can be trained for specific modules 115-m of the light source 100. For example, the ML model associated with wavelength can be trained for the MO chamber module 461 (Fig. 4). As another example, the ML model associated with bandwidth can be trained for the MO chamber module 461 (Fig. 4). And, the ML model associated with energy (of the amplified light beam 405) can be trained for the MO chamber module 461 (Fig. 4). [0044] The maintenance apparatus 120 evaluates the light source 100 using a plurality of machine learning models, as opposed to a single machine learning model. And, moreover, a pipeline is created between training of each machine learning model and use of each trained machine learning model, as discussed above. In this way, average prediction accuracy (in the evaluations) is improved and performance metrics that are required by the output apparatus 110 can be better met. Moreover, by using a plurality of machine learning models, bias toward dominant features and results can be reduced and there is no need to down select features, change underlying data, or model parameters to capture less prominent features. Additionally, the use of plural machine learning models in this manner results in a reduction in error in the evaluations and a reduction in variation whenever new data is evaluated by the maintenance apparatus 120.

[0045] The maintenance apparatus 120 includes, among other features, a prediction unit 122 and an ensemble unit 124. The prediction unit 122 accesses the plurality of machine learning models (which are either stored in the prediction unit 122 or stored within the maintenance apparatus 120). The prediction unit 122 evaluates each of the sub-feature datasets 140-sf. Specifically, the prediction unit 122 evaluates a particular sub-feature dataset 140-sf using the machine learning model that was trained based on the sub-feature training set that corresponds to that particular sub-feature dataset 140- sf. The ensemble unit 124 receives the evaluations from the prediction unit 122 and interprets these evaluations to determine a root cause of failure by aggregating the underlying model predictions and prioritizing failure reasons.

[0046] The maintenance apparatus 120 includes one or more of digital electronic circuitry, computer hardware, firmware, and software. The maintenance apparatus 120 also includes memory 128 that can be read-only memory and/or random access memory. Storage devices suitable for tangibly embodying computer program instructions and data include all forms of non-volatile memory, including, by way of example, semiconductor memory devices, such as EPROM, EEPROM, and flash memory devices; magnetic disks such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM disks. The maintenance apparatus 120 can also include one or more input devices 121 (such as a keyboard, touch screen, microphone, mouse, hand-held input device, etc.) and one or more output devices 123 (such as a speaker or a monitor). The maintenance apparatus also can include components to enable wireless communication including Bluetooth, NFC, and Wi-Fi. The maintenance apparatus 120 also includes one or more programmable processors, and one or more computer program products tangibly embodied in a machine-readable storage device for execution by one or more programmable processors. The one or more programmable processors can each execute a program of instructions to perform desired functions by operating on input data and generating appropriate outputs. Generally, the processors receive instructions and data from the memory. Any of the foregoing may be supplemented by, or incorporated in, especially designed ASICs (application-specific integrated circuits). The maintenance apparatus 120 can be centralized or be partially or wholly distributed throughout the light source 100, and it can be in communication with other controllers for controlling other aspects of the light source 100.

[0047] In further implementations, as shown in Fig. IB, an implementation 120B of the maintenance apparatus 120 includes a pipeline unit 126. The pipeline unit 126 is configured to receive the feature set 140 (which includes the plurality of sub-feature datasets 140-sf), and categorizes each sub-feature dataset 140-sf into a feature category based on its related unique performance criterion. For example, if a performance criterion is the long term bandwidth of the amplified light beam 105, then the feature category is bandwidth of the amplified light beam 105, if the performance criterion is the short term wavelength of the amplified light beam 105, then the feature category is wavelength of the amplified light beam 105, and if the performance criterion is errors in energy of the amplified light beam 105, then the feature category is energy of the amplified light beam 105.

[0048] The pipeline unit 126 communicates the sub-feature datasets 140-sf along with their related feature categories to the prediction unit 122. Each of the prediction unit 122, the ensemble unit 124, and the pipeline unit 126 can communicate with each other. Each of the prediction unit 122, the ensemble unit 124, and the pipeline unit 126 can access memory stored within the maintenance apparatus 120B. Moreover, each of the prediction unit 122, the ensemble unit 124, and the pipeline unit 126 can include one or more of digital electronic circuitry, computer hardware, firmware, and software. Each of the prediction unit 122, the ensemble unit 124, and the pipeline unit 126 can also include one or more programmable processors, and one or more computer program products tangibly embodied in a machine-readable storage device for execution by one or more programmable processors, and the one or more programmable processors can each execute a program of instructions to perform desired functions by operating on input data and generating appropriate outputs. Generally, these processors receive instructions and data from the memory 128.

[0049] The memory that is accessed by the units 122, 124, 126 can store the plurality of machine learning models, and can store the training datasets (the sub-feature training sets).

[0050] The maintenance apparatus 120B can also include an alert generating unit 125 configured to instruct a maintenance operation on a specific module 115-m if the ensemble unit 124 determines that one or more evaluations from the prediction unit 122 are related to that specific module 115-m.

[0051] A training pipeline 250A is shown in Fig. 2A and an operating pipeline 250B is shown in Fig. 2B. The training pipeline 250A shows development of the ML models that are accessed by and used by the maintenance apparatus 120 to provide advance notice of module failure for modules 115-m that are currently installed in the light source 100. The training pipeline 250 A includes training data 251 A acquired during data acquisition by a metrology apparatus, which includes a plurality of metrology units, each configured to acquire information about prior light sources 100P. For example, over time, many instances of the same module have been installed and subsequently deinstalled in prior light sources 100P deployed at global semiconductor fabs, and data collected during the lifetime of individual modules have been centrally stored. These historical data for deinstalled modules can be used for the data 251 A.

[0052] In general, each ML model (ML-sf in block 254A) is the output of a machine learning algorithm run on the training data 251 A and it represents what was learned by a machine learning algorithm. The prediction unit 122 can perform these ML algorithms, or another unit separate from the maintenance apparatus 120 can be used to perform these ML algorithms for training. Each ML model ML-sf is saved into memory 128 after running the machine learning algorithm on the training data 251 A. The ML model ML-sf represents algorithm-specific data structures (which can include rules and/or numbers) required for the maintenance apparatus 120 to eventually make predictions (operating pipeline 250B) on the light source 100 during operation to produce the amplified light beam 105. For example, the data structures of the ML model ML-sf can be made up of a vector or matrix of coefficients with specific values; a tree of if-then statements with specific values; or a graph structure with vectors or matrices of weights with specific values. Each of these types of ML models ML-sf have their own fitted data structures, where they are fit based on the algorithm that is used in training. Each ML model ML-sf that is saved to memory 128 includes both data and a procedure for using the data to recognize certain types of patterns and to make the prediction. Once each ML model ML-sf is trained, it is saved to memory 128, and the maintenance apparatus 120 uses each ML model ML-sf (in operating pipeline 250B) to reason over the data (the feature set 140) that it has not seen before, and to make predictions about those data (the feature set 140). Because each ML model ML-sf includes both the fitted data structure and the procedure for how to use this data, the maintenance apparatus 120 can make the prediction on the new data (for each sub-feature dataset 140-sf).

[0053] Examples of suitable ML models ML-sf that can be trained include Random Forest, AdaBoost, support vector machine (SVM), and Artificial Neural Network (ANN). Other ML models ML-sf can be used. In some implementations, each and every ML model ML-sf that is saved to memory 128 and used in the operating pipeline 250B has the same fitted data structure. In this case, for example, all of the ML models ML-sf could be Random Forest models. In other implementations, one or more of the ML models ML-sf that is saved to memory 128 and used in the operating pipeline 250B can have a unique or different fitted data structure from other ML models ML-sf that are saved to memory 128. Thus, for example, ML model ML-1 can be a Random Forest model while ML models ML-2 and ML-3 can be ANN models.

[0054] Also, in some implementations, the trained ML models ML-sf use many instances for every module in the population of deinstalled modules. For a strategy that evaluates modules starting at 10 billion light beam pulses to end of life in 0.1 billion pulse increments, for example, a module that fails at 32.55 billion pulses contributes 226 instances (of data) that can be used to train the models. Accordingly, a population of a few hundred modules can create a training dataset with tens of thousands of instances, which is favorable from a machine learning point of view in the sense that any of many different machine learning models can be used for implementation. It is also favorable because of enhanced accuracy for a larger training dataset.

[0055] The raw training data 251 A can be preprocessed to bring it into a form from which condition indicators can be extracted. Thus, block 252A corresponds to feature identification or extraction, which includes the identification of features (or metrics) whose behavior changes in a predictable way as the module ages and degrades.

[0056] Block 253 A corresponds to the identified sub-feature training sets 241-sf (where sf denotes each unique performance criterion of the prior light source 100P and sf has integer values between 1 and N). Each sub-feature training set 241-sf is used to train its respective machine learning model ML-sf, which provides a prediction increment estimate. The set of machine learning models ML-sf is represented by block 254A. In operational use (such as in the operating pipeline 250B), each machine learning model ML-sf forecasts an outcome based on that model that is prepared and trained on this past or historical input data (in training pipeline 250A) and its output behavior (the sub-feature training set 241-sf). Here, the training involves aligning estimates with real- world data as described below.

[0057] Each sub-feature training set 241-sf targets a specific aspect of performance (the unique performance criterion) of the prior light sources 100P. Each sub-feature training set 241-sf is curated to model a specific aspect of the prior light sources 100P. Examples of sub-feature training sets 241-sf are discussed next.

[0058] In one example, one particular performance criterion of prior light sources 100P is a longterm bandwidth of the amplified light beam 105 that is produced by the prior light sources 100P. The bandwidth of the amplified light beam 105 is the width of the intensity spectrum of the light beam 105, and this width can be given in terms of wavelength or frequency of the light beam 105. Any suitable mathematical construction (that is, metric) related to the details of the spectrum of the light beam 105 can be used to estimate the bandwidth of the light beam 105. For example, the full width of the spectrum at a fraction (X) of the maximum peak intensity (referred to as FWXM) can be used to estimate the light beam bandwidth. As another example, a width of the spectrum that contains a fraction (Y) of the integrated spectral intensity (referred to as EY) can be used to estimate the light beam bandwidth. Other metrics for estimating a bandwidth are suitable. Moreover, the bandwidth is measured on a “long term” usage scale. This means that each of the metrics are evaluated looking back for a longer usage (and then averaged using any suitable calculation). For example, a long-term bandwidth can be measured or estimated using one particular mathematical construct based on evaluations that look back a large number (for example, 2 billion) pulses of the light beam 105 produced by the prior light sources 100P. Thus, one sub-feature training set 241-sf (for example, training set 241-1) can include a set of metrics relating to the long-term bandwidth of the amplified light beam 105 produced by the prior light sources 100P such as the set [EYn, FWXMn, .. . MEn], where ME is any other metric that can be used and “It” denotes the look back evaluation is a longterm usage scale.

[0059] As another example, another performance criterion of prior light sources 100P is a short-term bandwidth of the amplified light beam 105 that is produced by the prior light sources 100P. In this particular criterion, the bandwidth is measured on a “short term” usage scale. This means that each of the metrics are evaluated looking back for a shorter usage (and then averaged using any suitable calculation). For example, a short -term bandwidth can be measured or estimated using one particular mathematical construct based on evaluations that look back a small number (for example, 100 million) pulses of the light beam 105 produced by the prior light sources 100P. Thus, one sub-feature training set 241-sf (for example, training set 241-2) can include a set of metrics relating to the shortterm bandwidth of the amplified light beam 105 produced by the prior light sources 100P such as the set [EY_st, FWXM_st, .. . ME_st], where ME is any other metric that can be used and “st” denotes the look back evaluation is a short-term usage scale.

[0060] Another possible performance criterion of the prior light sources 100P that can form a subfeature training set 241-sf (such as sub-feature training set 241-3) includes data relating to the number of error events (BQs) relating to bandwidth, and this data can be measured in the short-term usage scale or the long-term usage scale.

[0061] Other possible criteria of performance of prior light sources 100P that can form respective sub-feature training sets 241-sf (such as could be identified as sub-feature training sets 241-4, 241-5, 241-6, not explicitly shown in Fig. 2 A) include data relating to long-term energy (of the amplified light beam 105 output from the prior light sources 100P), short-term energy (of the amplified light beam 105 output from the prior light sources 100P), and data relating to the number of error events (BQs) relating to energy. Another possible performance criterion of prior light sources 100P that can form a sub-feature training set 241-sf (such as sub-feature training set 241-7, not explicitly shown in Fig. 2A) can correspond to a number of detections of very low energy of the amplified light beam 105 output from the prior light sources 100P, causing a system (light source) shutdown in a previous usage environment.

[0062] Other possible performance criterion of prior light sources 100P that can form respective subfeature training sets 241-sf (such as could be identified as sub-feature training sets 241-8, 241-9, 241- 10, not explicitly shown in Fig. 2A) include data relating to long-term wavelength (of the amplified light beam 105 output from the prior light sources 100P), short-term wavelength (of the amplified light beam 105 output from the prior light sources 100P), and data relating to the number of error events (BQs) relating to wavelength.

[0063] Other possible performance criterion of prior light sources 100P that can form respective subfeature training sets 241-sf (such as could be identified as sub-feature training sets 241-11 and 241-12, not explicitly shown in Fig. 2A) include data relating to long-term voltage and power amplifier signatures, and data relating to the number of error events (BQs) relating to such signatures. [0064] Still further possible performance criteria of prior light sources 100P that can form respective sub-feature training sets 241-sf include data relating to specific modules 115-m within the light source 100. For example, with additional reference to Fig. 4, other possible criteria of performance of prior light sources that are designed like the dual-stage light source 400 that can form respective subfeature training sets 241-sf include data relating to energy, wavelength, and bandwidth of the output light beam from the MO system 460 or data relating to energy, wavelength, and bandwidth of the output light beam from the PRA system 465. Other possible performance criteria of prior light sources that are designed like the dual-stage light source 400 that can form respective sub-feature training sets 241-sf include data relating to a pressure of the MO chamber module 461 or data relating to a pressure of the PRA chamber module 466. The specific modules 115-m within the light source 100 and the specific modules within the light source 400 can be inter-related or coupled together so that operation of one of the modules can impact operation of one or more other modules of that light source. As an example, if the MO chamber module 461 is not operating at an optimum or acceptable level, this can impact operation of the PRA chamber module 466, which can in turn operate outside its acceptable level to compensate for poor performance from the MO chamber module 461. In this way, some data in the sub-feature training sets 241-sf that is specifically associated with a particular module can impact operation of other modules. For example, the pressure of the PRA chamber module 466 can change if the MO chamber module 461 is operating outside its acceptable level. As another example, an energy of the output light beam from the PRA chamber module 466 can affect operation of the MO chamber module 461. As another example, failures associated with the PRA WEB module 467 are not associated with bandwidth performance of the light beam and thus it does not make sense to train a ML model using a bandwidth sub-feature training set 241-sf using data from PRA WEB modules 467. [0065] As mentioned above, the training data 251 A obtained for the sub-feature training sets 241-sf is obtained from metrology apparatuses 202P, each associated a prior light source 100P. Such metrology apparatus 202P includes a plurality of metrology units, each configured to acquire information about the prior light source 100P. Thus, in the examples above, the wavelength data can be obtained from a line-center analysis module (“LAM”), which can include, for example, an etalon spectrometer for fine wavelength measurement and a coarser resolution grating spectrometer through which the amplified light beam 105 (or a pre-cursor light beam that forms the light beam 105) is passed.

[0066] As another example, the bandwidth and energy data can be obtained from a bandwidth analysis module (“BAM”) that is positioned to receive a portion of the amplified light beam 105 (or a pre-cursor light beam that forms the light beam 105). The BAM is configured to measure the bandwidth and energy of pulses of the light beam 105 and it can include a grating spectrometer in which the light beam 105 (or its pre-cursor) is directed toward an echelle grating, which separates or disperses the light beam according to its wavelength, and the light beam reflected from the grating is directed to a camera such as a charge coupled device camera, which is able to resolve the wavelength distribution of the light beam.

[0067] As another example, data relating to energy of the pulses of the amplified light beam 105 can be measured using an energy monitor that measures a pulse energy of the light beam 105 before it enters the photolithography apparatus 110. The energy monitor can include a photodiode module. With reference to Fig. 4, an energy of an output light beam from the PRA chamber module 466 can be measured with a metrology unit positioned at the output of the PRA chamber module 466 and an energy of an output light beam from the (OPuS) module 477 can be measured with a metrology unit positioned at the output of the OPuS module 477. If the energy of the pulses of the output light beam from the PRA chamber module 466 is 10 milliJoules (mJ) and the energy of the pulses of the output light beam from the OPuS module 477 is only 2 mJ, then the OPuS module 477 is not operating within an acceptable range. To track operation of the OPuS module 477, a ML model can be trained for the overall light source 400 and a ML model can be trained for just the PRA chamber module 466, and the ensemble unit 124 can look at the results from both of these models to make a determination regarding whether the OPuS module 477 needs to be repaired or replaced.

[0068] The training data obtained for the sub-feature training sets 241-sf can be obtained from deinstalled modules of prior light sources 100P. For example, there is a certain population of these modules that have been deinstalled as part of carrying out umbrella maintenance. A subset of these modules will have failed. The complementary subset of these modules will have been deinstalled not because they have failed (they will not have failed) but because they are replaced when another module is replaced. Each of these deinstalled modules has associated with it an array of stored historical data in the form of metrics collected at various pulse counts. These metrics can be extracted and categorized into the sub-feature training set 241-sf. Then, the model ML-sf is trained by evaluating how well its predictions align with actual historical field outcomes. Ideally, the model ML- sf should predict failure within the prediction increment and should not predict failure within the prediction increment otherwise. For example, each of approximately 1000 deinstalled modules may have been first evaluated at 10 Bp and then in 0.1 Bp increments to the pulse count at deinstall. Assuming an average deinstall pulse count of about 30 Bp, this means each module will on average have contributed about 200 evaluations ((30 Bp - 10Bp)Z.1) and the entire group of deinstalled modules will yield 200,000 evaluations to provide the raw data 251 A. As discussed below, the construction of increments (or usage) can be expressed in terms of time. When considering the usage in terms of time, a daily increment to the pulse count can be considered at deinstall. Each of the 1000 deinstalled modules can be deinstalled an average of 180 days from 10 Bps to the 30 Bps expected end of life.

[0069] The metrics can also include additional parameters such as laser attributes including model, region where deployed, account associated with the module, customer type, power-level employed, temperature and blower speed, bandwidth and wavelength stability, bearing hours, neon reduction settings and faults, efficiency metrics, and so on. The above data is captured in varying lookback windows or as a rate, for example, mean wavelength sigma in previous 1 Bp and previous 5 Bp, or in previous 1 Bp with respect to the first 5 Bp of life. The metrics can also include consistent feature sets such as a count of individual fault signatures in the previous 100 Mp, 1 Bp, and 2 B or data derived from individual sensors such as voltage and MO energy.

[0070] Assuming at least 4 metrics for each of the 200,000 evaluations results in approximately 800,000 total metrics.

[0071] These metrics are supplied as the sub-feature training set 241-sf (block 253 A) to the respective model ML-sf in block 254A. In training, the model ML-sf makes predictions, compares the predictions to field outcomes, and then tunes its application of the metrics and compares again, converging on a prediction logic that aligns with the field outcomes. For example, the prediction logic can be executed to derive approximately 200,000 predictions. Then these 200,000 predictions can be compared to a prediction objective, for example, to predict failure within 2 Bp or less with respect to the deinstall pulse count for modules that actually failed, and do not predict failure otherwise. According to an aspect, outcome of the training results in predicting binary module failure in a prediction increment, for example, to predict module failure within a prediction increment prior to deinstall for failure while never predicting module failure prior to module deinstall for modules causing no technical issues. The system can work for any module and is designed and tested using historical data.

[0072] The set of machine learning models ML-sf (block 254 A) is stored in memory 128 within the maintenance apparatus 120. The machine learning models ML-sf are accessed by the prediction unit 122 in the operating pipeline 250B (Fig. 2B) during operation of the light source 100. The operating pipeline 250B includes data 25 IB that is acquired during data acquisition by a metrology apparatus 202, which includes a plurality of metrology units, each configured to acquire information about the light source 100 during operation of the light source 100. Examples of metrology units within the metrology apparatuses 202 are discussed with reference to the metrology apparatus 202P.

[0073] The raw data 25 IB can be preprocessed to bring it into a form from which condition indicators can be extracted. The pipeline unit 126 receives the raw data 25 IB (which includes the subfeature datasets 140-sf) and identifies, extracts, and categorizes these sub-feature datasets 140-sf from the data 251B. The sub-feature datasets 140-sf mirror the sub-feature training sets 241-sf. Thus, subfeature dataset 140-1 corresponds to the real-time data that is received relating to a unique performance criterion of the light source 100 during standard operation, while the sub-feature training set 241-1 corresponds to the training data associated with that unique performance criterion of the prior light sources 100P. Thus, in the example given above, if the sub-feature training set 241-1 includes the set of metrics relating to the long-term bandwidth of the amplified light beam 105 produced by the prior light sources 100P, then the sub-feature dataset 140-1 includes the set of metrics relating to the long-term bandwidth of the amplified light beam 105 produced by the light source 100. As another example, if the sub-feature training sets 241-8, 241-9, 241-10 include the metrics relating to, respectively, long-term wavelength (of the amplified light beam 105 output from the prior light sources 100P), short-term wavelength (of the amplified light beam 105 output from the prior light sources 100P), and data relating to the number of error events (BQs) relating to wavelength of the light beam 105 output from the prior light sources 100P, then the sub-feature datasets 140-8, 140-9, 140-10 include the metrics relating to, respectively, long-term wavelength, short-term wavelength, and data relating to number of error events (BQs) relating to wavelength of the light beam 105 output from the light source 100.

[0074] Both the preprocessing of the raw data 25 IB and the feature identification and extraction of the sub-feature datasets 140-sf from the data 251B is performed by the pipeline unit 126. In particular, the pipeline unit 126 receives this data 251B that includes all of the sub-feature datasets 140-sf and categorizes each sub-feature dataset 140-sf into a feature category based on its unique performance criteria. These identified and categorized sub-feature datasets 140-sf and assigned categories are provided to the prediction unit 122. Moreover, because the pipeline unit 126 maintains the pipeline between each sub-feature dataset 140-sf and its respective machine learning model ML-sf, which is trained based on the sub-feature training set 241-sf that corresponds to the sub-feature dataset 140-sf, the prediction unit 122 applies a particular sub-feature dataset 140-sf to the machine learning model ML-sf that was trained (Fig. 2A) based on the assigned category of the particular sub-feature training set 241-sf. Each machine learning model ML-sf is applied to its respective sub-feature dataset 140-sf and a prediction increment estimate Ev-sf is output from each machine learning model ML-sf. In particular, each machine learning model ML-sf forecasts an outcome based on that model that is prepared and trained on the past or historical input data (from data 251 A and sub-feature training sets 241-sf). Like its respective training set 241-sf, each sub-feature dataset 140-sf targets a specific aspect of performance (the unique performance criterion) of the light source 100, and each sub-feature dataset 140-sf is curated to model a specific aspect of the light source 100. The individual and separate estimates Ev-sf are provided to the ensemble unit 124.

[0075] Referring to Fig. 3A, a usage diagram 355 shows the terms that are used herein to describe operation of the maintenance apparatus 120 based on the usage of the light source 100. The diagram 355 extends along an axis of increasing usage 356 (denoted by the right-facing arrow) of the light source 100. The diagram 355 is not necessarily drawn to scale in order to show the concepts on a single page. Usage 356 of the light source 100 can be measured by any suitable construct. For example, the usage 356 of the light source 100 can be measured in terms of a number of pulses of the amplified light beam 105 that are produced. Thus, in the diagram 355, the number of pulses of the amplified light beam 105 increases in the direction of the arrow. In other implementations, it may be suitable to measure the usage 356 of the light source 100 in terms of time, for example, days, hours, minutes, etc. [0076] The maintenance apparatus 120 performs an evaluation at each evaluation moment or point Ei (where i = 0, 1, 2, 3, etc.) along the axis of usage 356. Evaluation points E0, El, E2, E3 are shown in Fig. 3 but many more evaluation points can persist after E3 even though they are not shown in Fig. 3. A first evaluation E0 is performance after a usage U0. An evaluation increment Elj is a usage gap between adjacent evaluations Ei and Ei+1, where j is indexed by 0, 1, 2, etc. Only three evaluation increments Elj are shown in Fig. 3A, namely, EI0, Ell, EI2, where EI0 is the usage gap between E0 and El; Ell is the usage gap between El and E2; and EI2 is the usage gap between E2 and E3. A prediction increment Pli is shown as extending forward in increasing usage from its respective evaluation point Ei. Thus, prediction increment PIO extends from evaluation point E0, prediction increment PI1 extends from evaluation point El, etc.

[0077] At each evaluation point Ei, there is at least one look back increment LBIi, which looks back in usage from its evaluation point Ei. As shown in the implementation or example of Fig. 3A, there are three look back increments LBIi associated with each evaluation point Ei. In other implementations, there could be fewer than three or more than three look back increments LBIi associated with each evaluation point Ei. In this example, the look back increments LB 10 are associated with the evaluation point E0, the look back increments LBII are associated with the evaluation point El, the look back increments LB 12 are associated with the evaluation point E2, and the look back increments LBI3 are associated with the evaluation point E3.

[0078] Specifically, in operation (such as when implementing the operating pipeline 250B), at each evaluation point Ei, the maintenance apparatus 120 evaluates, for each sub-feature dataset 140-sf, whether a failure mode is detected in the light source 100 (in a prediction increment Pi), and determines which performance criterion of the light source 100 is related to that failure mode based on the associated sub-feature dataset 140-sf of that failure mode. The maintenance apparatus 120 can determine that there is a failure mode for more than one sub-feature dataset 140-sf. The maintenance apparatus 120 can moreover determine that there are a plurality of performance criteria related to a particular failure mode. At each evaluation point Ei, the maintenance apparatus 120 looks forward in the prediction increment Pi to determine whether a failure mode will occur in that prediction increment Pi. On the other hand, at each evaluation point Ei, the maintenance apparatus 120 considers an aggregation in usage of each sub-feature dataset 140-sf over the look back increments LBIi that precede the evaluation point Ei. Thus, for example, at the evaluation point E0, the data of a particular sub-feature dataset 140-sf is taken from an aggregation in usage of data taken at three different look back increments LBI0_l, LBI0_2, LBI0_3. A look back increment LBIi can include one last usage step backward from the evaluation point Ei in which case aggregation merely involves using that last sub-feature dataset 140-sf at that last usage step backward.

[0079] The aggregation occurs while maintaining the pipeline. That is, the many sub-feature datasets 140-1 in the look back increment LBIi are aggregated with each other (but not with other sub-feature datasets 140-2, 140-3, etc.); the many sub-feature datasets 140-2 in the look back increment LBIi are aggregated with each other (but not with other sub-feature datasets 140-1, 140-3, etc.); and the many sub-feature datasets 140-3 in the look back increment LBIi are aggregated with each other (but not with other sub-feature datasets 140-1, 140-2, etc.).

[0080] With reference to Fig. 3B, an example of three look back increments LB 10 at evaluation point E0 are shown for three sub-feature datasets 140-1, 140-2, 140-3. In this example, the third look back increment LBI0 (labeled as the arrow with 3) includes one last usage step backward, and the data that is evaluated at the evaluation point E0 (referred to as 140-1 (AG3), 140-2 (AG3), 140-3 (AG3)) for the third look back increment LB 10 is simply the value of each of these three sub-feature datasets 140- 1, 140-2, 140-3 as they existed at the last usage step backward. The first look back increment LBII (labeled as the arrow with 1) includes several usage steps backward. For example, the first look back increment LBII can correspond to 100 million pulses backward. Thus, for the first look back increment LBII, data for the sub-feature dataset 140-1 is aggregated over 100 million pulses backward to form the aggregated sub-feature dataset 140-1 (AG1); data for the sub-feature dataset 140-2 is aggregated over 100 million pulses backward to form the aggregated sub-feature dataset 140- 2 (AG1); and data for the sub-feature dataset 140-3 is aggregated over 100 million pulses backward to form the aggregated sub-feature dataset 140-3 (AG1). The second look back increment LBII (labeled as the arrow with 2) includes several usage steps backward and is greater than the first look back increment LBII and the third look back increment LBII. For example, the second look back increment LBII can correspond to 1 billion pulses backward. Thus, for the second look back increment LBII, data for the sub-feature dataset 140-1 is aggregated over 1 billion pulses backward to form the aggregated sub-feature dataset 140-1 (AG2); data for the sub-feature dataset 140-2 is aggregated over 1 billion pulses backward to form the aggregated sub-feature dataset 140-2 (AG2); and data for the sub-feature dataset 140-3 is aggregated over 1 billion pulses backward to form the aggregated sub-feature dataset 140-3 (AG2).

[0081] The data can be aggregated using any suitable mathematical construct. For example, an aggregation of the data for a sub-feature dataset 140-sf can correspond to just a clumping together of the values according to some aggregation principle, such as using addition, or finding a mean value, or computing the averages for the values.

[0082] In some implementations, each look back increment LBIi is selected from a set of possible look back increments that relate to how the performance criteria change over usage. In some implementations, a set of look back increments LBIi includes a first look back increment (labeled as arrow 1) of about 2 billion pulses, a second look back increment (labeled as arrow 2) of about 1 billion pulses, and a third look back increment (labeled as arrow 3) of about 100 million pulses.

[0083] Referring again to Fig. 3A, one specific implementation is described next. The usage U0 is 10 billion pulses, which means that the maintenance apparatus 120 performs the first evaluation E0 after the light source 100 has produced 10 billion pulses. The evaluation increment Elj is 100 million pulses. Thus, the maintenance apparatus 120 performs an evaluation at each evaluation point Ei every 100 million pulses. There are three look back increments LBIi (thus there are three look back arrows such as shown in Figs. 3A and 3B) and the look back increments LBIi are 100 million pulses, 1 billion pulses, and 2 billion pulses. In this particular implementation, a module 115-m within the light source 100 that operates for 40 billion pulses, the evaluation point E0 is at 10 billion pulses, the evaluation point El is at 10.1 billion pulses, the evaluation point E2 is at 10.2 billion pulses, the evaluation point E3 is at 10.3 billion pulses, and so on up to the evaluation point at 40 billion pulses. Moreover, in this particular implementation, at the evaluation point E0 (which is 10 billion pulses), the first look back increment LBI0 of 100 million pulses includes data from 9.9 billion pulses to 10 billion pulses, the second look back increment LB 10 of 1 billion pulses includes data from 9 billion pulses to 10 billion pulses, and the third look back increment LB 10 of 2 billion pulses includes data from 8 billion pulses to 10 billion pulses. At the evaluation point El (which is 10.1 billion pulses), the first look back increment LBII of 100 million pulses includes data from 10 billion pulses to 10.1 billion pulses, the second look back increment LBII of 1 billion pulses includes data from 9.1 billion pulses to 10.1 billion pulses, and the third look back increment LBII of 2 billion pulses includes data from 8.1 billion pulses to 10.1 billion pulses.

[0084] The size of the evaluation increment Elj can be adjusted and can be configurable. For example, the evaluation increment can be 50 million pulses or 1.5 billion pulses.

[0085] In some implementations, at each evaluation point Ei, the maintenance apparatus 120 looks forward in a short term prediction increment Pi such as ahead 100 million pulses to 1 day ahead. In other implementations, at each evaluation point Ei, the maintenance apparatus 120 looks forward in a long term prediction increment Pi such as ahead 2 billion pulses to 20 days ahead.

[0086] As mentioned above, some of the sub-feature datasets 140-sf (such as the short-term bandwidth, the short-term wavelength, and the short-term energy) can relate to a performance criterion that is tracked in a relatively short look back increment LBIi. Using the example above, the short look back increment LBIi can be 100 million pulses. Some of the sub-feature datasets 140-sf (such as long-term bandwidth, long-term wavelength, and long-term energy) can relate to a performance criterion that is tracked in a relatively long look back increment LBIi. Using the example above, the long look back increment LBIi can be 1 billion pulses. Moreover, in such implementations in which there are long look back increments LBIi, the pipeline unit 126 is configured to aggregate in usage the sub-feature datasets 140-sf based on the same look back increment, and the prediction unit 122 is configured to receive the plurality of sub-feature datasets 140-sf by receiving an aggregation of sub-feature datasets 140-sf that are tracked using the same look back increment LBIi.

[0087] Referring to Fig. 4, an implementation 400 of the light source 100 is a dual-stage pulsed light source that produces a pulsed amplified beam 405 as a light beam 105. The light source 400 includes a solid state or gas discharge master oscillator (MO) system 460, a power amplification (PA) system such as a power ring amplifier (PRA) system 465, relay optics 470, and an optical output subsystem 475. [0088] The MO system 460 can include, for example, an MO chamber module 461, in which electrical discharges between electrodes (not shown) can cause lasing gas discharges in a lasing gas to create an inverted population of high energy molecules, such as including argon, krypton, or xenon to produce relatively broad band radiation that is line narrowed to a relatively very narrow bandwidth and center wavelength selected in a line narrowing module (‘LNM’) 462. The MO system 460 can also include an MO output coupler (MO OC) 462, which can include a partially reflective mirror, forming, with a reflective grating (not shown) in the LNM 462, an oscillator cavity in which the MO system 460 oscillates to form the seed output pulse thereby forming a master oscillator. The MO system 460 can also include a line-center analysis module (LAM) 463. As mentioned above, the LAM 180 can include, for example, an etalon spectrometer for fine wavelength measurement and a coarser resolution grating spectrometer.

[0089] The relay optics 470 can include an MO wavefront engineering box (WEB) 471 that serves to redirect the output of the MO system 460 toward the PA system 465, and can include, for example, beam expansion with, for example, a multi prism beam expander (not shown) and coherence busting, for example, in the form of an optical delay path (not shown).

[0090] The PA system 465 includes a PRA chamber module 466, which is also an oscillator, for example, formed by injection of the output light beam from the MO system 460 and output coupling optics (not shown) that can be incorporated into a PRA WEB 467 and can be redirected back through a gain medium in the chamber 466 by way of a beam reverser 468. The PRA WEB 467 can incorporate a partially reflective input/output coupler (not shown) and a maximally reflective mirror for the nominal operating wavelength (which can be at around 193 nm for an ArF system) and one or more prisms. The PA system 465 optically amplifies the output light beam from the MO system 460. [0091] The optical output subsystem 475 can include a bandwidth analysis module (BAM) 476 at the output of the PA system 465 that receive the output light beam of pulses from the PA system 465 and picks off a portion of the light beam for metrology purposes, for example, to measure the output bandwidth and pulse energy. The output light beam of pulses then passes through an optical pulse stretcher module (OPuS) 477 and an output combined autoshutter metrology module (CASMM) 478, which can also be the location of a pulse energy meter. One purpose of the OPuS 477 can be to convert a single output pulse into a pulse train. Secondary pulses created from the original single output pulse can be delayed with respect to each other. By distributing the original laser pulse energy into a train of secondary pulses, the effective pulse length of the light beam can be expanded and at the same time the peak pulse intensity reduced.

[0092] The light source 400 is made up of modules 115-m. Each of the components (such as the MO chamber 461, the LNM 462, the MO WEB 471, the PRA chamber 466, the PRA WEB 467, the OPuS 477, the BAM 476) of the light source 400 are modules 115-m. The overall availability of the light source 400 is the direct result of the respective availabilities of these individual modules 115-m making up the light source 400. In other words, the light source 400 cannot be available unless all of these modules making up the light source 400 are available. The maintenance apparatus 120 enables these modules to be monitored and replaced before they fail to maintain the operation of the light source 400 and optimize and improve productivity of the output apparatus 110. The maintenance apparatus 120 provides validated failure alerts that can be used to pinpoint which modules 115-m will fail.

[0093] Referring to Fig. 5, in some implementations, the output apparatus 110 is a photolithography exposure apparatus 510. The photolithography exposure apparatus 510 uses the amplified light beam 105 to pattern microelectronic features on a substrate or wafer 511. The wafer 511 is placed on a wafer table 512 constructed to hold the wafer 511 and connected to a positioner configured to position the wafer 511 accurately in accordance with certain parameters. The photolithography exposure apparatus 510 can use a light beam 505 (output from the light source 100) having a wavelength in the deep ultraviolet (DUV) range, which can include wavelengths from, for example, about 100 nanometers (nm) to about 400 nm. For example, the light source 100 that produces such a light beam 505 can be a gas discharge light source such as an excimer light source, or excimer laser that uses a combination of one or more noble gases, which can include argon, krypton, or xenon, and a reactive gas, which can include fluorine or chlorine as the gain medium. The light source 400 can be an excimer light source. Thus, for example, the gain medium can include argon fluoride (ArF), krypton fluoride (KrF), or xenon chloride (XeCl). If the gain medium includes argon fluoride, then the wavelength of the amplified light beam 505 is about 193 nm and if the gain medium includes krypton fluoride, then the wavelength of the amplified light beam 505 is about 248 nm. The size of the microelectronic features patterned on the wafer 511 depends on the wavelength of the light beam 505, with a lower wavelength resulting in a smaller minimum feature size. When the wavelength of the light beam 505 is 248 nm or 193 nm, the minimum size of the microelectronic features can be, for example, 50 nm or less. The bandwidth of the light beam 505 can be the actual, instantaneous bandwidth of its optical spectrum (or emission spectrum), which contains information on how the optical energy of the light beam 505 is distributed over different wavelengths.

[0094] The photolithography exposure apparatus 510 includes an optical arrangement having, for example, one or more condenser lenses, a mask, and an objective arrangement. The mask is movable along one or more directions, such as along an optical axis of the light beam 505 or in a plane that is perpendicular to the optical axis. The objective arrangement includes a projection lens and enables an image transfer to occur from the mask to the photoresist on the wafer 511. The photolithography exposure apparatus 510 also includes an illumination system that adjusts the range of angles for the light beam 505 impinging on the mask. The illumination system also homogenizes (makes uniform) the intensity distribution of the light beam 505 across the mask.

[0095] The photolithography exposure apparatus 510 can also include, among other features, a lithography controller 513, air conditioning devices, and power supplies for the various electrical components. The lithography controller 513 controls how layers are printed on the wafer 511. The lithography controller 513 includes a memory that stores information such as process recipes. A process program or recipe determines the length of the exposure on the wafer 511, the mask used, and other factors that affect the exposure. During lithography, a plurality of pulses of the light beam 505 illuminates the same area of the wafer 511 to together constitute an illumination dose.

[0096] Referring to Fig. 6, once the ML models ML-sf have been trained and saved to memory 128 (as shown in the training pipeline 250A of Fig. 2A), the maintenance apparatus 120 performs a procedure 680 for maintaining the light source 100 by evaluating whether and which of its modules 115-m should be repaired or replaced. Initially, the maintenance apparatus 120 receives the subfeature datasets 140 in the form of operational data 251B (see also Fig. 2B) (681). Such data 251B is obtained from the various metrology units of the metrology apparatus 202. With additional reference to Fig. 3A, the maintenance apparatus 120 receives the operational data 251B just prior to an evaluation point Ei as the light source 100 is being used (and usage increases). And, after the first evaluation E0, the maintenance apparatus 120 receives each set of data 25 IB following an evaluation increment Elj. Additionally, the data 25 IB can be aggregated according to one or more look back increments LBIi. Thus, the data 25 IB can be stored in memory 128 and when an evaluation point Ei is reached, the past data 25 IB can be aggregated according to a look back increment LBIi at the evaluation point Ei.

[0097] As discussed above, the operational data 25 IB is parsed into the plurality of sub-feature datasets 140-sf, with each sub-feature dataset 140-sf being associated with a unique performance criterion of the light source 100. An example of nine sub-feature datasets 140-sf (where sf = 1, 2, ... 9), with each representing one of three performance criteria bandwidth, wavelength, and energy, is shown in the table of Fig. 7A. Sub-feature datasets 140-1, 140-2, 140-3 are all associated with a bandwidth of the amplified light beam 105 that is output from the light source 100; sub-feature datasets 140-4, 140-5, 140-6 are all associated with a wavelength of the amplified light beam 105 that is output from the light source 100; and sub-feature datasets 140-7, 140-8, 140-9 are all associated with an energy of the amplified light beam 105 that is output from the light source 100. Thus, the ML models ML-1, ML-2, ML-3 were trained, respectively, on a long term bandwidth training set, a short term bandwidth training set, and a bandwidth errors training set; the ML models ML-4, ML-5, ML-6 were trained, respectively, on a long term wavelength training set, a short term wavelength training set, and a wavelength errors training set; and the ML models ML-7, ML-8, ML-9 were trained, respectively, on a long term energy training set, a short term energy training set, and an energy errors training set. As also discussed above, the long term training sets are data sets that are aggregated over longer look back increments LBIi while the short term training sets are data sets that are aggregated over shorter look back increments LBIi.

[0098] Referring again to Fig. 6, the maintenance apparatus 120 performs an evaluation (682) based on the received sub-feature datasets 140-sf. Specifically, at step 682, the maintenance apparatus 120 evaluates 682-sf, for each sub-feature dataset 140-sf, whether a failure mode is detected in the light source 100. In particular, the prediction unit 122 performs each evaluation 682-sf using the ML model ML-sf that is associated with the sub-feature dataset 140-sf, in accordance with the pipeline discussed above. The prediction unit 122 performs the evaluation at the evaluation point Ei (Fig. 3A). In evaluating whether a failure mode is detected (682-sf) for a particular sub-feature dataset 140-sf, the prediction unit 122 is evaluating whether a failure will occur in the light source 100 at a future usage, that is, in the forward usage prediction increment Pli (Fig. 3A).

[0099] The evaluations 682-sf are scores based on the outcome of the respective ML model ML-sf. The score is related to a probability of failure in a module 115-m of the light source 100 in the prediction increment Pli. The score can be a certainty score, a binary output, or a function in a range between 0 and 1. For example, if the ML model ML-sf is a Random Forest, then the score is based on the majority rule voting since the Random Forest model is a collection of uniformed Decision Trees. As another example, in ML model ML-sf that is a single Decision Tree with two classes, a score can correspond to a terminal node. For example, if a Tree ended in a terminal node (leaf) with 100 samples, 80 of which are class 1 and 20 are of class 2, then the certainty score can be defined as 80/100 = .8 for class 1 and 20/100 = .2 of class 2. And, different terminal nodes and tree paths yield other certainty scores based on the data. The Tree is motivated to produce pure terminal nodes. As a further example, an ANN ML model ML-sf can convert results into a two-class output using an activation function like sigmoid or logistic or some other function where the range is between 0 and 1. Thus a score from an evaluation 682-sf can vary depending on which ML model ML-sf is being used for the particular evaluation 682-sf.

[0100] Referring to the table of Fig. 7B, a set of evaluations 682-1, 682-2, 682-3, 682-4, 682-5, 682- 6, 682-7, 682-8, 682-9 are shown with the output score for each of the respective nine sub-feature datasets 140-1, 140-2, 140-3, 140-4, 140-5, 140-6, 140-7, 140-8, 140-9 of Fig. 7A, with each subfeature dataset 140-sf representing one of the three performance criteria categories of bandwidth, wavelength, and energy. As discussed above with reference to Fig. 7 A, the ML models ML-1, ML-2, ML-3 were trained, respectively, on a long term bandwidth training set, a short term bandwidth training set, and a bandwidth errors training set; the ML models ML-4, ML-5, ML-6 were trained, respectively, on a long term wavelength training set, a short term wavelength training set, and a wavelength errors training set; and the ML models ML-7, ML-8, ML-9 were trained, respectively, on a long term energy training set, a short term energy training set, and an energy errors training set. The evaluations 682-sf are shown for three evaluation points El, E2, and E3. Many more evaluation points Ei can follow what is shown in Fig. 7B. For example, at evaluation point El, the output of the evaluation 682-1 (which corresponds to evaluating the sub-feature dataset 140-1 with the ML model ML-1) is 0.16 and the output of evaluation 682-8 (which corresponds to evaluating the sub-feature dataset 140-8 with the ML model ML-8) is 0.89. As a further example, at evaluation point E2, the output of evaluation 682-3 (which corresponds to evaluating the sub-feature dataset 140-3 with the ML model ML-3) is 0.21; the output of evaluation 682-4 (which corresponds to evaluating the sub- feature dataset 140-4 with the ML model ML-4) is 0.92; the output of evaluation 682-5 (which corresponds to evaluating the sub-feature dataset 140-4 with the ML model ML-4) is 0.85; and the output of evaluation 682-8 (which corresponds to evaluating the sub-feature dataset 140-8 with the ML model ML-8) is 0.28. The prediction unit 122 can determine that a failure mode is detected if a particular score is greater than a minimum value. For example, the prediction unit 122 can determine that a failure mode is detected if a score is greater than 0.8. Such scores are bolded in Fig. 7B.

[0101] Next, the maintenance apparatus 120 determines which performance criterion or criteria are related to any failure modes that were detected in step 682 (683). In particular, at step 683, the ensemble unit 124 first determines whether any of the evaluations 682-sf (at step 682) resulted in a detected failure mode (683-sf), and if there is a detected failure mode (683-sf), then the ensemble unit 124 determines which performance criterion or criteria is related to that failure mode 683-sf (684-sf) . This determination at 684-sf is based on the sub-feature dataset 140-sf that is associated with that failure mode 683-sf.

[0102] The ensemble unit 124 can determine which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset (685) by aggregating the evaluations from the prediction unit 122, and prioritizing reasons for failure that are associated with each detected failure mode.

[0103] In this example, the ensemble unit 124 is configured to categorize an evaluation as a “hard failure” if a single evaluation of a particular performance criterion is greater than 0.95; a “soft failure” if two separate evaluations from two separate sub-feature datasets within a particular performance criterion are between 0.85 and 0.95; and no failure mode otherwise. Based on this configuration, the following evaluation 682-sf results in a hard failure relating to short term energy: evaluation 682-8 performed at evaluation point El (for the sub-feature dataset 140-8 that is short term energy). Additionally, the following evaluations result in soft failures: the evaluation 682-4 performed at evaluation point E2 (for the sub-feature dataset 140-4 that is long term wavelength) and the evaluation 682-6 performed at evaluation point E2 (for the sub-feature dataset 140-6 that is wavelength errors). In this particular example, at step 683, the ensemble unit 124 determines that: the failure mode detected during the evaluation 682-8 at evaluation point El is related to the energy of the light beam 105 output from the light source 100; the failure mode detected during the evaluation 682-4 at evaluation point E2 is related to the long term wavelength of the light beam 105 output from the light source 100; and the failure mode detected during the evaluation 682-6 at evaluation point E2 is related to the wavelength errors of the light beam 105 output from the light source 100. A hard failure can indicate that a particular module 115-m of the light source 100 needs to be entirely replaced. A soft failure can indicate that a particular module 115-m of the light source 100 needs to be watched or repaired, or the performance criterion needs to be labeled for further evaluation. Other hard failure and soft failure values may be used in other examples. [0104] As mentioned above, some modules 115-m of the light source 100 can be much more likely to affect a certain performance criterion or criteria than other modules 115-m of the light source 100. For example, operation of the MO chamber module 461 has a greater impact on the bandwidth, the wavelength, and the energy of the amplified light beam 405 from the light source 400. Thus, because of this, a hard failure detected for short term energy (682-8 at El) can indicate that the MO chamber module 451 will fail during the prediction increment PI1. The ensemble unit 124 can analyze those scores 682-sf that are related to a particular assigned performance criterion or category in making the determination of whether a module 115-m needs to be replaced or watched. For example, the ensemble unit 124 can analyze the scores 682-7, 682-8, 682-9, which all relate to energy of the amplified light beam 105, to make the determination regarding whether a particular module 115-m (such as the MO chamber module 461) needs to be watched or replaced. The maintenance apparatus 120 is able to pinpoint the problem within the light source 100 (and specifically the particular module 115-m) based on which sub-feature dataset 140-sf resulted in a failure mode.

[0105] In one example, the ensemble unit 124 is configured to label a performance criterion (such as the energy) for further evaluation if at least one failure mode associated with that performance criterion is detected, and such label can be independent of whether another failure mode associated with another performance criterion (such as bandwidth and wavelength) is also detected. Indeed, in Fig. 7B, the ensemble unit 124 labels the energy for further evaluation at evaluation point El even though there is no failure mode for either the bandwidth or the wavelength at evaluation point El. [0106] The ensemble unit 124 can determine which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset 140-sf by aggregating the evaluations 682-sf from the prediction unit 122, and prioritizing reasons for failure that are associated with each detected failure mode.

[0107] For example, if one of the bandwidth models ML-1, ML-2, ML-3 reports a score 682-sf of .8 but the wavelength and energy models ML-4, ML-5, ML-6, ML-7, ML-8, ML-9 for the MO chamber module 461 are reporting a score 682-sf of .1 and .15, respectively, then the ensemble unit 124 can deduce that the bandwidth of the light beam is out of range, and there is about to be a failure in the MO chamber module 461. Moreover, as discussed above, there are three sub-feature datasets corresponding to the bandwidth, namely, the long term bandwidth (associated with model ML-1), the short term bandwidth (associated with model ML-2), and the bandwidth errors (associated with model ML-3). If the scores 682-sf of models ML-1, ML-2, ML-3 are, respectively, .7, .1, .15, and all other models ML-4 to ML-9 reported scores below .2, then the ensemble unit 124 determines that the bandwidth error is the root cause and errors are being generated that result in failure and replacement of the MO chamber module 461.

[0108] As an example, the ensemble unit 124 can instruct the following actions based on each of scores (evaluations 682-sf), where a score can be any number from 0 to 1 and each score is associated with a ML model: for a score 682-sf that is from 0-.25, the ensemble unit 124 can determine there is no issue and can ignore, and also there is no need to report this out; for a score 682-sf that is from .25-.5, the ensemble unit 124 can determine there is no issue but also can report this out; for a score 682-sf that is from .5-.75, the ensemble unit 124 can determine there is a potential issue, can report, and can instruct to monitor; and for a score 682-sf that is from .75-1, the ensemble unit 124 can determine there is an impending issue, can report, and can instruct to replace the module 115-m because there will be failure soon.

[0109] The procedure 680 can further include, prior to step 681, receiving the entire feature set 140, which includes the plurality of sub-feature datasets 140-sf, and categorizing each of these sub-feature datasets 140-sf into a feature category based on its related unique performance criterion. Specifically, the pipeline unit 126 can perform these functions, in order to prepare the data for use by the prediction unit 122, and also to maintain the link between each sub-feature dataset 140-sf and its associated ML model ML-sf. Moreover, the pipeline unit 126 can also aggregate the data from each sub-feature dataset 140-sf in usage based on the one or more look back increments LBIi, as discussed above.

[0110] The embodiments can be further described using the following clauses:

1. An apparatus for maintaining a light source comprising one or more modules that together are configured to produce a light beam for semiconductor photolithography, the apparatus comprising: a prediction unit configured to: receive a plurality of sub-feature datasets, each sub-feature dataset associated with a unique performance criterion of the light source; and for each sub-feature dataset, evaluate whether a failure mode is detected in the light source in a prediction increment; and an ensemble unit configured to: receive the plurality of evaluations from the prediction unit; and for each failure mode, determine which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode.

2. The apparatus of clause 1, wherein the prediction unit includes or accesses a plurality of models, each model being generated by machine learning using a unique sub-feature dataset.

3. The apparatus of clause 2, wherein each of the models in the prediction unit has the same fitted structure.

4. The apparatus of clause 2, wherein each of the models in the prediction unit has a different fitted structure.

5. The apparatus of clause 2, wherein at least one of the models in the prediction unit has a fitted structure that is distinct from the fitted structure of another model in the prediction unit.

6. The apparatus of clause 1, wherein the plurality of sub-feature datasets includes a sub-feature dataset relating to bandwidth of the light beam, a sub-feature dataset relating to wavelength of the light beam, and a sub-feature dataset relating to energy of the light beam. 7. The apparatus of clause 1, wherein the plurality of sub-feature datasets includes a sub-feature dataset relating to long term bandwidth of the light beam, a sub-feature dataset relating to short term bandwidth of the light beam, a sub-feature dataset relating to bandwidth error events, a sub-feature dataset relating to long term wavelength of the light beam, a sub-feature dataset relating to short term wavelength of the light beam, a sub-feature dataset relating to wavelength error events, a sub-feature dataset relating to long term energy of the light beam, a sub-feature dataset relating to short term energy of the light beam, and a sub-feature dataset relating to energy error events.

8. The apparatus of clause 1, wherein the ensemble unit is configured to determine which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset by aggregating the evaluations from the prediction unit, and prioritizing reasons for failure that are associated with each detected failure mode.

9. The apparatus of clause 8, wherein the ensemble unit is configured to aggregate evaluations from the prediction unit based on the performance criterion associated with each sub-feature dataset.

10. The apparatus of clause 9, wherein the ensemble unit is configured to label a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected.

11. The apparatus of clause 9, wherein the ensemble unit is configured to label the performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.

12. The apparatus of clause 1, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in a specific module of the light source based on a model that is trained for the specific module.

13. The apparatus of clause 12, wherein the specific module of the light source comprises a master oscillator module.

14. The apparatus of clause 1, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in one or more specific modules of the light source based on models that are each trained for a specific module.

15. The apparatus of clause 14, wherein the one or more specific modules of the light source comprise one or more of: a master oscillator module, a power amplifier module, a line narrowing module, a spectral feature analysis module, and a pulse stretcher module.

16. The apparatus of clause 1, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes: separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and for each sub-feature dataset evaluation, outputting a score related to a probability of failure.

17. The apparatus of clause 16, wherein the score is a certainty score, a binary output, or a function in a range between 0 and 1.

18. The apparatus of clause 1, wherein the prediction unit is configured to, for each sub-feature dataset, evaluate whether a failure mode is detected in the light source in the prediction increment by performing an evaluation each time an evaluation increment has passed.

19. The apparatus of clause 18, wherein an evaluation increment is 100 million pulses of the light beam.

20. An apparatus for maintaining a light source comprising one or more modules that together are configured to produce a light beam for semiconductor photolithography, the apparatus comprising: a pipeline unit configured to: receive a plurality of sub-feature datasets, each sub-feature dataset relating to a unique performance criterion of the light source during operation; and categorize each of the received sub-feature datasets into a feature category based on its related unique performance criterion; a prediction unit configured to: receive the plurality of sub-feature datasets and assigned categories for each sub-feature dataset; and evaluate whether a failure mode is detected in the light source in a prediction increment; and an ensemble unit configured to determine which of the performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset.

21. The apparatus of clause 20, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in a specific module of the light source based on a plurality of models that are trained for the specific module.

22. The apparatus of clause 21, wherein each model is developed through machine learning by supplying a respective sub-feature dataset to train the model.

23. The apparatus of clause 21, further comprising an alert generating unit configured to instruct a maintenance operation on the specific module if the evaluation determines that the failure mode is detected in the specific module.

24. The apparatus of clause 20, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes: separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and for each sub-feature dataset evaluation, outputting a score related to a probability of failure.

25. The apparatus of clause 24, wherein the ensemble unit being configured to determine which performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset includes analyzing the score from the sub-feature datasets related to the same assigned category.

26. The apparatus of clause 24, wherein the score is a certainty score, a binary output, or a function in a range between 0 and 1.

27. The apparatus of clause 20, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment comprises determining whether at least one module within the light source does not have at least the minimum probability of operating without a failure in the prediction increment.

28. The apparatus of clause 20, wherein the prediction increment is measured as a number of pulses of the light beam.

29. The apparatus of clause 20, wherein, prior to each evaluation, the pipeline unit is further configured to aggregate, in usage, data of at least one sub-feature dataset over one or more look back increments that precede the evaluation.

30. The apparatus of clause 29, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that subfeature dataset, and, if the sub-feature dataset is aggregated, in usage, over one or more look back increments, then using the aggregated sub-feature dataset.

31. The apparatus of clause 30, wherein each look back increment is selected from a set of possible look back increments that relate to how performance criteria change over usage.

32. The apparatus of clause 31, wherein the set of possible look back increments includes a first look back increment of about 2 billion pulses, a second look back increment of about 1 billion pulses, and a third look back increment of about 100 million pulses.

33. The apparatus of clause 30, wherein one or more of the sub-feature datasets relate to a performance criterion that is tracked in a relatively short look back increment and one or more of the sub-feature datasets relate to a performance criterion that is tracked in a relatively long look back increment.

34. The apparatus of clause 29, wherein the pipeline unit is configured to aggregate in usage one or more sub-feature datasets based on the same look back increment, and the prediction unit is configured to receive the plurality of sub-feature datasets by receiving an aggregation of sub-feature datasets that are tracked using the same look back increment.

35. A method for maintaining a light source comprising one or more modules that together are configured to produce a light beam for semiconductor photolithography, the method comprising: receiving a plurality of sub-feature datasets, each sub-feature dataset associated with a unique criterion of performance of the light source; for each sub-feature dataset, evaluating whether a failure mode is detected in the light source in a prediction increment; and for each detected failure mode, determining which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode.

36. The method of clause 35, further comprising, after receiving the plurality of sub-feature datasets, categorizing each of the received sub-feature datasets into a feature category based on its related unique performance criterion; wherein determining which one or more performance criterion is related to the failure mode comprises determining which performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset.

37. The method of clause 35, wherein evaluating whether a failure mode is detected in the light source in a prediction increment comprises: separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and for each sub-feature dataset evaluation, outputting a score related to a probability of failure.

38. The method of clause 35, wherein evaluating whether a failure mode is detected in the light source in the prediction increment comprises performing an evaluation each time an evaluation increment has passed.

39. The method of clause 35, wherein determining which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode comprises aggregating the evaluations, and prioritizing reasons for failure that are associated with each detected failure mode.

40. The method of clause 39, wherein aggregating evaluations comprises aggregating the evaluations based on the performance criterion associated with each sub-feature dataset.

41. The method of clause 40, further comprising labelling a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected.

42. The method of clause 40, further comprising labelling the performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.

43. The method of clause 35, wherein evaluating whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in a specific module of the light source based on a model that is trained for the specific module.

44. The method of clause 35, wherein evaluating whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in one or more specific modules of the light source based on models that are each trained for a specific module.

45. The method of clause 35, further comprising, prior to each evaluation, aggregating, in usage, data of at least one sub-feature dataset over one or more look back increments that precede the evaluation. 46. The method of clause 45, wherein aggregating in usage data of at least one sub-feature dataset over one or more look back increments comprises aggregating in usage one or more sub-feature datasets based on the same look back increment, and receiving the plurality of sub-feature datasets comprises receiving an aggregation of sub-feature datasets that are tracked using the same look back increment.

47. A method for maintaining a light source comprising one or more modules that together are configured to produce a light beam for semiconductor photolithography, the method comprising: receiving a plurality of sub-feature datasets, each sub-feature dataset associated with a unique performance criterion of the light source; for each sub-feature dataset, evaluating whether a failure mode is detected in the light source in a prediction increment including separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and for each detected failure mode, determining which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode, wherein determining includes aggregating the evaluations, and prioritizing reasons for failure that are associated with each detected failure mode.

48. The method of clause 47, wherein aggregating evaluations comprises aggregating the evaluations based on the performance criterion associated with each sub-feature dataset.

49. The method of clause 48, further comprising labelling a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.

[0111] The above description includes examples of multiple embodiments. It is, of course, not possible to describe every conceivable combination of components or methodologies for purposes of describing the these embodiments, but one of ordinary skill in the art may recognize that many further combinations and permutations of various embodiments are possible. Accordingly, the described embodiments are intended to embrace all such alterations, modifications, and variations that fall within the spirit and scope of the appended claims. Furthermore, to the extent that the term “includes” is used in either the detailed description or the claims, such term is intended to be inclusive in a manner similar to the term “comprising” as “comprising” is construed when employed as a transitional word in a claim. Furthermore, although elements of the described aspects and/or embodiments may be described or claimed in the singular, the plural is contemplated unless limitation to the singular is explicitly stated. Additionally, all or a portion of any aspect and/or embodiment may be utilized with all or a portion of any other aspect and/or embodiment, unless stated otherwise.

[0112] Other implementations are within the scope of the claims.

Claims

33 CLAIMS

2. The apparatus of claim 1, wherein the prediction unit includes or accesses a plurality of models, each model being generated by machine learning using a unique sub-feature dataset.

3. The apparatus of claim 2, wherein each of the models in the prediction unit has the same fitted structure.

4. The apparatus of claim 2, wherein each of the models in the prediction unit has a different fitted structure.

5. The apparatus of claim 2, wherein at least one of the models in the prediction unit has a fitted structure that is distinct from the fitted structure of another model in the prediction unit.

6. The apparatus of claim 1, wherein the plurality of sub-feature datasets includes a subfeature dataset relating to bandwidth of the light beam, a sub-feature dataset relating to wavelength of the light beam, and a sub-feature dataset relating to energy of the light beam.

7. The apparatus of claim 1, wherein the plurality of sub-feature datasets includes a subfeature dataset relating to long term bandwidth of the light beam, a sub-feature dataset relating to short term bandwidth of the light beam, a sub-feature dataset relating to bandwidth error events, a sub-feature dataset relating to long term wavelength of the light beam, a sub-feature dataset relating to short term wavelength of the light beam, a sub-feature dataset relating to wavelength error events, a sub-feature dataset relating to long term energy of the light beam, a sub-feature dataset relating to short term energy of the light beam, and a sub-feature dataset relating to energy error events. 34

8. The apparatus of claim 1, wherein the ensemble unit is configured to determine which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset by aggregating the evaluations from the prediction unit, and prioritizing reasons for failure that are associated with each detected failure mode.

9. The apparatus of claim 8, wherein the ensemble unit is configured to aggregate evaluations from the prediction unit based on the performance criterion associated with each sub-feature dataset.

10. The apparatus of claim 9, wherein the ensemble unit is configured to label a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected.

11. The apparatus of claim 9, wherein the ensemble unit is configured to label the performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.

12. The apparatus of claim 1, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in a specific module of the light source based on a model that is trained for the specific module.

13. The apparatus of claim 12, wherein the specific module of the light source comprises a master oscillator module.

14. The apparatus of claim 1, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in one or more specific modules of the light source based on models that are each trained for a specific module.

15. The apparatus of claim 14, wherein the one or more specific modules of the light source comprise one or more of: a master oscillator module, a power amplifier module, a line narrowing module, a spectral feature analysis module, and a pulse stretcher module.

16. The apparatus of claim 1, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes: separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and for each sub-feature dataset evaluation, outputting a score related to a probability of failure.

17. The apparatus of claim 16, wherein the score is a certainty score, a binary output, or a function in a range between 0 and 1.

18. The apparatus of claim 1, wherein the prediction unit is configured to, for each subfeature dataset, evaluate whether a failure mode is detected in the light source in the prediction increment by performing an evaluation each time an evaluation increment has passed.

19. The apparatus of claim 18, wherein an evaluation increment is 100 million pulses of the light beam.

20. An apparatus for maintaining a light source comprising one or more modules that together are configured to produce a light beam for semiconductor photolithography, the apparatus comprising: a pipeline unit configured to: receive a plurality of sub-feature datasets, each sub-feature dataset relating to a unique performance criterion of the light source during operation; and categorize each of the received sub-feature datasets into a feature category based on its related unique performance criterion; a prediction unit configured to: receive the plurality of sub-feature datasets and assigned categories for each subfeature dataset; and evaluate whether a failure mode is detected in the light source in a prediction increment; and an ensemble unit configured to determine which of the performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset.

21. The apparatus of claim 20, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in a specific module of the light source based on a plurality of models that are trained for the specific module.

22. The apparatus of claim 21, wherein each model is developed through machine learning by supplying a respective sub-feature dataset to train the model.

23. The apparatus of claim 21, further comprising an alert generating unit configured to instruct a maintenance operation on the specific module if the evaluation determines that the failure mode is detected in the specific module.

24. The apparatus of claim 20, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes: separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and for each sub-feature dataset evaluation, outputting a score related to a probability of failure.

25. The apparatus of claim 24, wherein the ensemble unit being configured to determine which performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset includes analyzing the score from the sub-feature datasets related to the same assigned category.

26. The apparatus of claim 24, wherein the score is a certainty score, a binary output, or a function in a range between 0 and 1.

27. The apparatus of claim 20, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment comprises determining whether at least one module within the light source does not have at least the minimum probability of operating without a failure in the prediction increment.

28. The apparatus of claim 20, wherein the prediction increment is measured as a number of pulses of the light beam.

29. The apparatus of claim 20, wherein, prior to each evaluation, the pipeline unit is further configured to aggregate, in usage, data of at least one sub-feature dataset over one or more look back increments that precede the evaluation.

30. The apparatus of claim 29, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source in the prediction increment includes separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset, and, if the sub-feature dataset is aggregated, in usage, over one or more look back increments, then using the aggregated sub-feature dataset. 37

31. The apparatus of claim 30, wherein each look back increment is selected from a set of possible look back increments that relate to how performance criteria change over usage.

32. The apparatus of claim 31, wherein the set of possible look back increments includes a first look back increment of about 2 billion pulses, a second look back increment of about 1 billion pulses, and a third look back increment of about 100 million pulses.

33. The apparatus of claim 30, wherein one or more of the sub-feature datasets relate to a performance criterion that is tracked in a relatively short look back increment and one or more of the sub-feature datasets relate to a performance criterion that is tracked in a relatively long look back increment.

34. The apparatus of claim 29, wherein the pipeline unit is configured to aggregate in usage one or more sub-feature datasets based on the same look back increment, and the prediction unit is configured to receive the plurality of sub-feature datasets by receiving an aggregation of sub-feature datasets that are tracked using the same look back increment.

35. A method for maintaining a light source comprising one or more modules that together are configured to produce a light beam for semiconductor photolithography, the method comprising: receiving a plurality of sub-feature datasets, each sub-feature dataset associated with a unique performance criterion of the light source; for each sub-feature dataset, evaluating whether a failure mode is detected in the light source in a prediction increment; and for each detected failure mode, determining which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode.

36. The method of claim 35, further comprising, after receiving the plurality of sub-feature datasets, categorizing each of the received sub-feature datasets into a feature category based on its related unique performance criterion; wherein determining which one or more performance criterion is related to the failure mode comprises determining which performance criterion or criteria is related to the failure mode based on the assigned categories for each sub-feature dataset.

37. The method of claim 35, wherein evaluating whether a failure mode is detected in the light source in a prediction increment comprises: separately evaluating each sub-feature dataset according to a model that is generated using machine learning using that sub-feature dataset; and 38 for each sub-feature dataset evaluation, outputting a score related to a probability of failure.

38. The method of claim 35, wherein evaluating whether a failure mode is detected in the light source in the prediction increment comprises performing an evaluation each time an evaluation increment has passed.

39. The method of claim 35, wherein determining which one or more performance criterion is related to the failure mode based on the associated sub-feature dataset of that failure mode comprises aggregating the evaluations, and prioritizing reasons for failure that are associated with each detected failure mode.

40. The method of claim 39, wherein aggregating evaluations comprises aggregating the evaluations based on the performance criterion associated with each sub-feature dataset.

41. The method of claim 40, further comprising labelling a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected.

42. The method of claim 40, further comprising labelling the performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.

43. The method of claim 35, wherein evaluating whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in a specific module of the light source based on a model that is trained for the specific module.

44. The method of claim 35, wherein evaluating whether the failure mode is detected in the light source in the prediction increment includes evaluating whether the failure mode is detected in one or more specific modules of the light source based on models that are each trained for a specific module.

45. The method of claim 35, further comprising, prior to each evaluation, aggregating, in usage, data of at least one sub-feature dataset over one or more look back increments that precede the evaluation.

46. The method of claim 45, wherein aggregating in usage data of at least one sub-feature dataset over one or more look back increments comprises aggregating in usage one or more sub- 39 feature datasets based on the same look back increment, and receiving the plurality of sub-feature datasets comprises receiving an aggregation of sub-feature datasets that are tracked using the same look back increment.

48. The method of claim 47, wherein aggregating evaluations comprises aggregating the evaluations based on the performance criterion associated with each sub-feature dataset.

49. The method of claim 48, further comprising labelling a performance criterion for further evaluation if at least one failure mode associated with that performance criterion is detected and independently of whether another failure mode associated with another performance criterion is detected.