WO2022266890A1 - 一种确定失效原因的方法及装置 - Google Patents

一种确定失效原因的方法及装置 Download PDF

Info

Publication number
WO2022266890A1
WO2022266890A1 PCT/CN2021/101866 CN2021101866W WO2022266890A1 WO 2022266890 A1 WO2022266890 A1 WO 2022266890A1 CN 2021101866 W CN2021101866 W CN 2021101866W WO 2022266890 A1 WO2022266890 A1 WO 2022266890A1
Authority
WO
WIPO (PCT)
Prior art keywords
root cause
failure
analysis model
cause analysis
root
Prior art date
Application number
PCT/CN2021/101866
Other languages
English (en)
French (fr)
Inventor
黄鑫
秦旻
黄宇
许若圣
丁智浩
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to CN202180099751.1A priority Critical patent/CN117561502A/zh
Priority to PCT/CN2021/101866 priority patent/WO2022266890A1/zh
Publication of WO2022266890A1 publication Critical patent/WO2022266890A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/30—Monitoring
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • This application is mainly concerned with the field of failure analysis. More specifically, it relates to a method, apparatus, device, computer-readable storage medium, and computer program product for determining a failure cause.
  • the physical root cause also referred to as "root cause”
  • the analysis results obtained by the layout-aware analysis technology have relatively large noises. For example, a certain analysis result may point to multiple suspected areas, and a suspected area may contain multiple root causes, etc. This makes further analysis to determine the true root cause difficult. In such cases, further screening is performed by using root cause analysis (RCA) techniques.
  • RCA root cause analysis
  • the embodiments disclosed in this application provide a solution for determining the cause of failure.
  • an apparatus for determining a cause of a failure includes an inference module configured to identify a root cause of failure of the component under analysis based on a root cause analysis model based on a plurality of quantified representations respectively associated with failures of portions of the component under analysis.
  • the apparatus also includes a learning module coupled to the reasoning module.
  • the learning module includes: an input sample receiving unit configured to receive a set of input samples for the root cause analysis model; a feature sample acquisition unit configured to acquire a set of feature samples related to the historical training of the root cause analysis model, a set of Each sample in the input sample and a set of feature samples includes a root cause of failure of the reference element, and a plurality of reference quantified representations respectively associated with failure of a plurality of reference portions of the reference element; and a model update unit configured to based on the reference The root cause of component failure and multiple references are quantified to train the root cause analysis model.
  • the reasoning module can use the root cause analysis model to quickly determine the root cause of the failure of the analyzed component.
  • the learning module can update the root cause analysis model with additional samples from new processes or new designs and feature samples related to historical training. In this way, the root cause analysis model can be adapted for different designs and processes. This can improve the identification accuracy of the root cause analysis model for new processes and new designs, while ensuring stable performance for all historical tasks, thereby providing more accurate root cause determination results.
  • the update of the root cause analysis model can be done automatically, thereby saving manpower, material and time costs and improving efficiency.
  • the characteristic sample acquisition unit is further configured to: generate a set of characteristic samples according to the sample generator, and the sample generator is configured to reproduce the characteristics of the historical samples used to train the root cause analysis model, And a set of feature samples reflects the characteristics of historical samples.
  • the sample generator uses the sample generator, the storage space consumed for storing historical samples can be reduced. In this way, updating of the root cause analysis model can be achieved in a resource-efficient manner.
  • the learning module further includes: a generator updating unit configured to update network parameters of the sample generator based on a set of input samples.
  • a generator updating unit configured to update network parameters of the sample generator based on a set of input samples.
  • the updated sample generator is able to reproduce the characteristics of a set of input samples.
  • the updated sample generator can be used for the next update of the root cause analysis model, thereby achieving stable continuous learning and updating of the root cause analysis model.
  • the model updating unit is further configured to: determine potential causes of reference element failures identified by the root cause analysis model by applying the plurality of reference quantified representations to the root cause analysis model; and by Minimize the difference between the root cause and the potential cause of the reference component failure, updating the network parameters of the root cause analysis model.
  • the parameters of the root cause analysis model can be re-determined using the supervised training process to ensure the stable performance of the root cause analysis model for all historical tasks.
  • the reasoning module includes: a diagnosis report receiving unit configured to receive a plurality of diagnosis reports on the element to be analyzed, each diagnosis report corresponding to one of the plurality of parts of the element to be analyzed , and indicating potential root causes, physical defects, and logical errors of a part failure; a quantitative representation generation unit configured to generate a plurality of quantitative representations respectively related to failures of the plurality of parts from a plurality of diagnostic reports; and a root cause identification A unit configured to identify a root cause of failure of the component under analysis by applying the plurality of quantitative representations to the root cause analysis model. In this way, the root cause of failure of the component under analysis can be quickly determined using an end-to-end root cause analysis model.
  • the root cause identification unit is further configured to: generate a plurality of local failure features from the plurality of quantitative representations by applying the plurality of quantitative representations to a feature extractor in the root cause analysis model, each a local failure feature corresponding to one of the multiple parts; combining the multiple local failure features into a global failure feature for the component to be analyzed; and determining by applying the global failure feature to a classifier in the root cause analysis model The probability that different root causes will cause failure of the component under analysis. In this way, a local-to-global integration is achieved, which facilitates a more accurate determination of the root cause of failure of the component under analysis.
  • the quantitative representation generating unit is further configured to: for a given diagnostic report of the plurality of diagnostic reports, combine the quantitative representations of the plurality of potential root causes indicated in the given diagnostic report into a combination with a quantitative representation of a first physical defect associated with a plurality of potential root causes; combining the quantitative representation of the first physical defect and the quantitative representation of at least a second physical defect into a first physical defect associated with the first physical defect and the second physical defect a quantified representation of a logical error; and generating a quantified representation related to a failure of a part based at least on the quantified representation of the first logical fault.
  • the quantitative representation related to the failure is constructed in a bottom-up manner, so that the constructed quantitative representation can reflect the local failure of the component to be analyzed.
  • the reasoning module further includes: a quantitative representation adjustment unit configured to, before applying the multiple quantitative representations to the root cause analysis model, based on The area of , to adjust multiple quantized representations. In this way, the influence of component area corresponding to different root causes can be reduced or even eliminated, which helps to further improve the accuracy of root cause determination.
  • At least one input sample of the set of input samples includes a root cause determined from a root cause analysis model and verified via testing.
  • root cause analysis models can be updated with real samples from new designs or processes. This allows updated root cause analysis models to better accommodate new designs or processes.
  • the element to be analyzed includes a wafer to be analyzed, and the plurality of portions includes a plurality of dies of the wafer to be analyzed.
  • a method for determining a failure cause includes receiving a set of input samples for a root cause analysis model configured to identify root causes of failure of the component under analysis based on a plurality of quantitative representations respectively associated with failures of portions of the component under analysis.
  • Cause obtain a set of feature samples related to the historical training of the root cause analysis model, each sample in the set of input samples and the set of feature samples includes the root cause of the failure of the reference element, and a plurality of reference parts respectively related to the reference element multiple reference quantified representations related to the failure of the reference component; and training a root cause analysis model based on the root cause of the failure of the reference component and the multiple reference quantified representations.
  • the root cause of the failure of the analyzed component can be quickly determined by using the root cause analysis model. Updating the root cause analysis model with additional samples from new processes or new designs and feature samples related to historical training allows the root cause analysis model to adapt to different designs and processes. This can improve the identification accuracy of the root cause analysis model for new processes and new designs, while ensuring stable performance for all historical tasks, thereby providing more accurate root cause determination results. In addition, the update of the root cause analysis model can be done automatically, thereby saving manpower, material and time costs and improving efficiency.
  • obtaining a set of characteristic samples includes: generating a set of characteristic samples according to a sample generator configured to reproduce the characteristics of historical samples used to train the root cause analysis model, and a Group characteristic samples reflect the characteristics of historical samples.
  • a sample generator configured to reproduce the characteristics of historical samples used to train the root cause analysis model
  • a Group characteristic samples reflect the characteristics of historical samples.
  • the method further includes updating network parameters of the sample generator based on the set of input samples.
  • the updated sample generator is able to reproduce the characteristics of a set of input samples.
  • the updated sample generator can be used for the next update of the root cause analysis model, thereby achieving stable continuous learning and updating of the root cause analysis model.
  • training the root cause analysis model includes: determining potential causes of reference element failures identified by the root cause analysis model by applying a plurality of reference quantified representations to the root cause analysis model; The network parameters of the root cause analysis model are updated by optimizing the difference between the root cause and the potential cause of the reference component failure. In this way, the parameters of the root cause analysis model can be re-determined using the supervised training process to ensure the stable performance of the root cause analysis model for all historical tasks.
  • the method further comprises: receiving a plurality of diagnostic reports on the element to be analyzed, each diagnostic report corresponding to one of the plurality of parts of the element to be analyzed and indicating a failure of one part.
  • Potential root causes, physical defects, and logical errors from multiple diagnostic reports, generate multiple quantified representations, each associated with the failure of multiple parts; and identify component failures under analysis by applying multiple quantified representations to a root cause analysis model root cause. In this way, the root cause of failure of the component under analysis can be quickly determined using an end-to-end root cause analysis model.
  • identifying the root cause of failure of the component to be analyzed comprises: generating a plurality of local failure features from the plurality of quantitative representations by applying the plurality of quantitative representations to a feature extractor in the root cause analysis model, each local failure feature corresponds to one of the plurality of parts; combining the plurality of local failure features into a global failure feature for the component to be analyzed; and by applying the global failure feature to a classifier in the root cause analysis model, Determine the probability that different root causes will cause the component under analysis to fail. In this way, a local-to-global integration is achieved, which facilitates a more accurate determination of the root cause of failure of the component under analysis.
  • generating the plurality of quantitative representations includes: for a given diagnostic report of the plurality of diagnostic reports, combining quantitative representations of the plurality of potential root causes indicated in the given diagnostic report into a combination of multiple a quantitative representation of a first physical defect associated with a potential root cause; combining the quantitative representation of the first physical defect and the quantitative representation of at least a second physical defect into a first logical error associated with the first physical defect and the second physical defect and generating a quantified representation related to a failure of a part based at least on the quantified representation of the first logic error.
  • the quantitative representation related to the failure is constructed in a bottom-up manner, so that the constructed quantitative representation can reflect the local failure of the component to be analyzed.
  • the method further comprises: prior to applying the plurality of quantified representations to the root cause analysis model, adjusting the plurality of quantified representations based on areas associated with potential root causes respectively indicated by the plurality of diagnostic reports express. In this way, the influence of component area corresponding to different root causes can be reduced or even eliminated, which helps to further improve the accuracy of root cause determination.
  • At least one input sample of the set of input samples includes a root cause determined from a root cause analysis model and verified via testing.
  • root cause analysis models can be updated with real samples from new designs or processes. This allows updated root cause analysis models to better accommodate new designs or processes.
  • the element to be analyzed comprises a wafer to be analyzed
  • the plurality of portions comprises a plurality of dies of the wafer to be analyzed.
  • an electronic device including: at least one processor; at least one memory, the at least one memory is coupled to the at least one processor and stores instructions for execution by the at least one processor, The instructions, when executed by at least one processor, cause the device to implement the method in any one of the implementation manners in the second aspect.
  • a computer-readable storage medium on which one or more computer instructions are stored, wherein one or more computer instructions are executed by a processor to implement any one of the second aspects. method in the implementation.
  • a computer program product is provided.
  • the computer program product is run on a computer, the computer is instructed to execute some or all steps of the method in any implementation manner in the second aspect.
  • the electronic device of the third aspect, the computer storage medium of the fourth aspect, or the computer program product of the fifth aspect provided above are all used to execute the method provided in the second aspect. Therefore, the explanations or explanations about the second aspect are also applicable to the third aspect, the fourth aspect and the fifth aspect.
  • the beneficial effects achieved by the third aspect, the fourth aspect, and the fifth aspect reference may be made to the beneficial effects in the corresponding methods, which will not be repeated here.
  • Figure 1 shows a schematic diagram of an example environment in which various embodiments of the present application can be implemented
  • Figure 2 shows a schematic diagram of constructing a quantitative representation from a diagnostic report according to some embodiments of the present application
  • Fig. 3 shows a schematic diagram of the architecture of the root cause analysis model according to some embodiments of the present application
  • Fig. 4 shows a schematic diagram of updating a root cause analysis model according to some embodiments of the present application
  • Fig. 5 shows a schematic diagram of an extended feature library according to some embodiments of the present application.
  • FIG. 6 shows a flowchart of a process for determining a failure cause according to some embodiments of the present application
  • Fig. 7 shows a schematic block diagram of an apparatus for determining a failure cause according to some embodiments of the present application.
  • Figure 8 shows a schematic block diagram of a computing device capable of implementing various embodiments of the present application.
  • a "neural network” is capable of processing input and providing a corresponding output, which generally includes an input layer and an output layer and one or more hidden layers between the input layer and the output layer.
  • Neural networks used in deep learning applications often include many hidden layers, extending the depth of the network. The layers of the neural network are connected in sequence so that the output of the previous layer is provided as the input of the subsequent layer, where the input layer receives the input of the neural network, and the output of the output layer serves as the final output of the neural network.
  • the terms "neural network”, “network”, “neural network model” and “model” are used interchangeably.
  • root cause analysis techniques need to be used to determine the root cause of a problem.
  • root cause analysis refers to the process of identifying the root cause of a problem or failure.
  • root cause analysis can be used to find the root cause of chip failure, thereby effectively ensuring the high productivity of chips in the mass production process.
  • chip testing is usually performed first.
  • the main process of chip testing is: first, test the chip or a part of the chip according to the test vector generated by the automated test equipment (ATE), and find the error type from the error dictionary according to the output result of the logic circuit, so as to obtain the ATE failure data; Netlist pattern layout, conduct layout-aware diagnosis of failed pins in ATE failure data, and generate a diagnosis report based on prior knowledge.
  • the diagnostic report mainly includes key areas, suspected areas, etc.
  • the root cause analysis mainly analyzes the generated diagnostic report to obtain the root cause of chip failure. Specifically, using the root cause analysis, the probability distribution of various root causes can be inferred from the diagnosis report, so as to locate the main root cause and solve the problem of chip failure.
  • Bayesian models are used to calculate probability distributions for individual root causes. Based on the analysis of real physical relationships, corresponding probabilistic links can be established between root causes, physical defects and logical errors.
  • the corresponding data can be extracted from the diagnostic report generated by the failure and the corresponding transition probability can be calculated.
  • the unsupervised learning-based Expectation-Maximization (EM) algorithm is used to solve for the root-cause distribution that best explains a series of diagnostic reports.
  • the solution process of the EM algorithm includes an iterative expectation step and a maximization step.
  • the probability distribution P(c k )' of the kth root cause c k in the current iteration is used to estimate the posterior probability P(c k
  • the probability distribution P(c k )' of the root cause is updated by maximizing the sum of the posterior probabilities P(c k
  • This conventional method has great shortcomings, because different processes have a great influence on the accuracy of root cause analysis. Especially with the continuous evolution of technology, the process size is getting smaller and smaller, and the root cause of the middle process layer (that is, the lower metal layer) is more difficult to distinguish. Each process update needs to rely on the cooperation and communication of various parties to improve, which consumes a lot of manpower and time.
  • RCA is also required to determine the root cause of failure.
  • a root cause analysis model is first obtained.
  • the root cause analysis model is configured to identify a root cause of failure of the component under analysis based on a plurality of quantitative representations respectively related to failure of portions of the component under analysis. If a set of input samples for the root cause analysis model is received, a set of feature samples related to the historical training of the root cause analysis model is obtained. Each sample includes, respectively, a root cause of failure of the reference element, and a plurality of reference quantified representations related to the failure of the plurality of reference portions of the reference element.
  • input samples and feature samples are labeled samples.
  • the root cause analysis model is trained again to update the network parameters of the root cause analysis model based on the multiple reference quantitative representations included in the input sample and the feature sample and the root cause of failure of the reference components.
  • the root cause of failure of the analyzed component can be quickly determined by using the root cause analysis model. Utilizing newly added samples of new processes or new designs and characteristic samples related to historical training to update the root cause analysis model can realize the continuous learning of the root cause analysis model and make the root cause analysis model adaptable to different designs and processes. This can improve the identification accuracy of the root cause analysis model for new processes and new designs, while ensuring stable performance for all historical tasks, thereby providing more accurate root cause determination results. In addition, the update of the root cause analysis model can be done automatically, thereby saving manpower, material and time costs and improving efficiency.
  • FIG. 1 shows a schematic diagram of an example environment 100 in which various embodiments of the present application can be implemented.
  • the example environment 100 includes a root cause analysis engine 105 and a plurality of diagnostic reports 120-1, 120-2, . is the diagnosis report 120, wherein N is an integer greater than or equal to 1.
  • N is an integer greater than or equal to 1.
  • the component to be analyzed 101 includes N parts 102 - 1 , 102 - 2 , .
  • the component to be analyzed 101 is a wafer, and the plurality of parts 102 are a plurality of dies included in the wafer.
  • the element to be analyzed 101 may be a spacecraft, and the parts may be parts included in the spacecraft.
  • Each diagnostic report 120 corresponds to a portion of the component 101 to be analyzed, and indicates diagnostic information related to the failure of that portion, such as potential root causes of failure, physical defects, logical errors, and the like.
  • diagnostic report 120-2 indicates diagnostic information related to failure of part 102-2, such as potential root causes of failure of part 102-2, physical defects, logical errors, and the like. Diagnostic report 120 may be generated in any suitable manner, as embodiments of the present application are not limited in this regard.
  • the root cause analysis engine 105 includes a reasoning module 110 and a continuous learning module 160 (also referred to simply as a "learning module").
  • the reasoning module 110 is used to identify the root cause of failure of the component to be analyzed 101 according to a plurality of diagnostic reports 120 on the component to be analyzed 101 .
  • the reasoning module 110 receives a plurality of diagnostic reports 120 and converts diagnostic information about failures in the diagnostic reports into a quantitative representation.
  • the reasoning module 110 generates N quantified representations 130-1, 130-2, ..., 130-N respectively related to the failures of the plurality of parts 102 from the plurality of diagnostic reports 120, which are also collectively referred to as a plurality of quantified Representation 130 or quantized representation 130 alone.
  • Each quantitative representation 130 is derived from a diagnostic report 120 .
  • This quantitative representation 130 derived from the diagnostic report 120 can be regarded as a local raw signature of the failure of the component 101 to be analyzed.
  • the reasoning module 110 may also include an adjustment operation 140 .
  • An adjustment operation 140 adjusts each quantized representation 130 with a weight to remove the effect of the device area corresponding to the different root causes.
  • Quantitative representation 130 or adjusted quantitative representation 130 is applied to root cause analysis model 150 .
  • the root cause analysis model 150 generates an analysis result 155 based on the received quantitative representation 130 that identifies the root cause of the failure of the component 101 under analysis.
  • analysis results 155 may include probabilities for different root causes.
  • the analysis results may include a predetermined number (eg, one) of the most likely root causes.
  • different root causes may include, but are not limited to, metal layer open circuit (such as M1 open circuit, M2 open circuit, etc.), metal layer short circuit (such as M1 short circuit, M2 short circuit, etc.), Via (VIA) shorts (such as VIA1 short, VIA2 short, etc.), larger devices have problems (referred to as “device problems” in this article), such as shorts to surrounding parts, etc.
  • the continuous learning module 160 is used for training the root cause analysis model 150 to update network parameters of the root cause analysis model 150 .
  • the continuous learning module 160 receives a set of input samples 180 .
  • the input samples 180 may be or may include samples that have not been used to train the root cause analysis model 150 .
  • input samples 180 may be new samples produced in a new process or a new design.
  • the input samples 180 are labeled samples that include: a root cause of the failure of the reference element, and a plurality of reference quantified representations respectively associated with the failure of the plurality of reference portions of the reference element.
  • input samples 180 may be samples analyzed using root cause analysis model 150 .
  • the flags included in the input sample 180 ie, the root cause of failure of the reference component
  • the reasoning module 110 identifies the main root cause of the failure of the wafer based on the diagnostic reports of multiple dies of the failed wafer.
  • a Physical Validation Analysis (PFA) can then be performed experimentally against this primary root cause. If the primary root cause identified by the root cause analysis model 150 passes the PFA verification, the data related to that wafer may be used by the continuous learning module 160 as an input sample 180 .
  • PFA Physical Validation Analysis
  • the input samples 180 may be samples generated by error injection simulation. For example, after developing a new process or a new design that has not been learned in the historical training of the root cause analysis model 150, new samples for the new process or new design can be generated through error injection simulation as a method for updating the root cause analysis. An input sample 180 to the model 150 . In this way, after a new process or a new design appears, additional samples can be quickly generated, so that the updated root cause analysis model 150 can be quickly applied to the new process or new design.
  • the continuous learning module 160 includes a feature library 190 .
  • the signature library 190 is used to reproduce (eg, store or generate) samples related to the historical training of the root cause analysis model 150, which are also referred to as "signature samples.”
  • the feature samples from the feature library 190 together with the input samples 180 form a full sample set.
  • the continuous learning module 160 further utilizes the continuous learning algorithm 170 to retrain the root cause analysis model 150 based on the full sample set.
  • the input samples 180 may be used to expand the feature library 190 such that the expanded feature library 190 reproduces the input samples 180 in subsequent updates.
  • the feature library 190 may include storage means for storing historical samples used for training the root cause analysis model 150 in historical training tasks. For example, when the number of historical samples is small, the historical samples may be stored in the feature library 190 after the training task ends. In such an embodiment, after receiving the input samples 180 , historical samples can be directly retrieved from the feature library 190 as feature samples, and together with the input samples, form a full sample set. If the number of input samples 180 is small, the input samples 180 may be directly stored in the feature library 190 to expand the feature library 190 .
  • feature library 190 may include a sample generator.
  • the sample generator is configured to reproduce characteristics of the historical samples used to train the root cause analysis model in the historical training task.
  • feature samples that can reflect the characteristics of historical samples can be generated according to the sample generator, and together with the input samples, a full sample set can be formed.
  • the sample generator can be retrained based on the input samples 180 to expand the feature library 190 . Such an embodiment will be described in detail below with reference to FIGS. 4 and 5 .
  • root cause analysis engine 105 may be divided in any suitable manner.
  • the functionality of deriving quantitative representation 130 from diagnostic report 120 may be implemented independently of reasoning module 110 . That is, in some embodiments, inference module 110 may receive quantitative representation 130 and apply quantitative representation 130 or adjusted quantitative representation 130 to root cause analysis model 150 .
  • the reasoning module 110 and the continuous learning module 160 may be implemented on the same computing device or computing system. In other embodiments, the reasoning module 110 and the continuous learning module 160 may be implemented on different computing devices or computing systems that communicate with each other.
  • FIG. 2 shows a schematic diagram of constructing a quantitative representation 130 from a diagnostic report 120 according to some embodiments of the present application.
  • the diagnosis report 120 may be any one of the plurality of diagnosis reports 120-1, 120-2, . . . , 120-N shown in FIG. 1 .
  • Failure-related information in diagnostic report 120 may be organized in a recursive tree structure 200 as shown in FIG. 2 .
  • the recursive tree includes bottom-up multiple levels such as root cause level, physical defect level and logical defect level.
  • the root cause hierarchy includes potential multiple root causes 210-1, 210-2, 210-3 indicated by the diagnostic report 120 that caused the failure of the corresponding part 102, also referred to collectively as multiple root causes 210 or individually as root causes 210 .
  • Information related to root cause 210 may include the name and critical area of the root cause.
  • the critical area refers to the area of the device involved by the corresponding root cause.
  • the physical defect level above the root cause level includes a plurality of sub-suspect areas 220 - 1 , 220 - 2 , 220 - 3 , which are also collectively referred to as a plurality of sub-suspect areas 220 or individually referred to as a suspected area 220 .
  • Each sub-region of suspicion 220 may be a physical defect and is associated with at least one root cause of the root cause hierarchy.
  • sub-region of suspicion 220-1 is associated with root causes 210-1, 210-2, 210-3, etc.
  • Information about the sub-suspect area 220 may include the type of defect and the score given by the diagnostic report 120 . This score can represent the probability of occurrence of a defect.
  • the level of physical defects is the level of logical errors, which in turn can be divided into two sub-levels.
  • the lower sub-level includes a plurality of suspected regions 230 - 1 , 230 - 2 , 230 - 3 , which are collectively referred to as a plurality of suspected regions 230 or individually referred to as a suspected region 230 .
  • Each suspected area 230 is associated with multiple suspected sub-areas, for example, may include multiple suspected sub-areas.
  • suspected region 230-1 is associated with sub-suspected regions 220-1, 220-2, 220-3, and so on.
  • a higher sub-level includes a plurality of symptoms 240 - 1 , 240 - 2 , 240 - 3 , which are also collectively referred to as a plurality of symptoms 240 or individually as symptoms 240 .
  • Symptoms 240 may refer to errors manifested directly in the test.
  • Each symptom 240 is associated with multiple suspected areas. For example, symptom 240-1 is associated with suspected areas 230-1, 230-2, 230-3.
  • the reasoning module 110 can construct the quantitative representation 130 in a bottom-up recursive manner based on the recursive tree structure 200 . Specifically, for each root cause 210 , the reasoning module 110 can extract information related to the root cause from the diagnosis report 120 , such as the name of the root cause (such as M1 open circuit, VIA2 short circuit, etc.) and critical area. The extracted information is then quantized to generate a quantized representation (eg, vector) of the root cause. The quantified representation of the root cause and the quantified representations of other root causes are combined into a quantified representation of the associated sub-suspect region 220 of the upper level. In the example of FIG. 2, the quantified representations of root causes 210-1, 210-2, 210-3, etc. are combined into a quantized representation of sub-region of suspicion 220-1.
  • the quantified representations of root causes 210-1, 210-2, 210-3, etc. are combined into a quantized representation of sub-region of suspicion 220-1.
  • the reasoning module 110 can extract information related to the corresponding suspected sub-region 220 from the diagnosis report 120 , such as type and score. Based on the extracted information, the quantified representation of this sub-suspect region and the quantized representations of other sub-suspect regions are combined into a quantized representation of the associated suspected region 230 of the upper level. In the example of Fig. 2, the quantified representations of the sub-suspect regions 220-1, 220-2, 220-3 are combined into a quantized representation of the suspected region 230-1.
  • quantified representation of each suspected region 230 and the quantified representations of the other suspected regions are combined into a quantified representation of the associated symptom 240 of the next higher level.
  • the quantified representations of suspect regions 230-1, 230-2, 230-3 are combined into a quantified representation of symptom 240-1.
  • the quantitative representations of the plurality of symptoms 240 - 1 , 240 - 2 , 240 - 3 are combined into a quantitative representation 130 corresponding to the diagnostic report 120 .
  • Quantized representation 130 may be implemented as a one-dimensional vector, but is not limited thereto.
  • Each quantitative representation 130 corresponds to a portion 102 of the component to be analyzed 101 and is a local raw signature of the failure of the component to be analyzed 101 .
  • multiple quantized representations of the lower level may be combined into a quantized representation of the upper level in any suitable manner.
  • multiple quantized representations at a lower level may be averaged or maximized to generate a quantized representation at a higher level.
  • the quantitative representation of the sub-region of suspicion 220-1 may be generated by averaging the quantitative representations of the root causes 210-1, 210-2, 210-3.
  • multiple quantized representations at a lower level may be concatenated as different channels, and multiple quantized representations concatenated together may be used as quantized representations at an upper level.
  • the quantified representations of the root causes 210-1, 210-2, 210-3 linked as different channels can be used as the quantified representation of the sub-region of suspicion 220-1.
  • the reasoning module 110 may also include an adjustment operation 140 .
  • the adjustment operation 140 may adjust each quantized representation 130 according to the following equation:
  • ReportFeat i represents the quantified representation extracted from the i-th diagnostic report
  • w i represents the adjustment weight for the i-th diagnostic report
  • the value of i ranges from 1 to N.
  • the weight w i is used to reduce or even eliminate the influence of the device area corresponding to different root causes.
  • the weight w i may be determined based on the critical areas of the different root causes indicated by the diagnostic report 120 (as described above with reference to FIG. 2 ).
  • an RCA constant extracted from a plurality of diagnostic reports 120 on the component to be analyzed 101 may be used as the weight w i .
  • the weight w i is the same for all diagnostic reports. It can be understood that the weight w i has the same dimension as ReportFeat i , and the calculation of formula (1) is element-wise. In such an embodiment, by reducing the influence of the areas associated with different root causes, the root causes ultimately identified by the root cause analysis model 150 can be made more accurate.
  • the root cause analysis model 150 After obtaining the quantitative representation 130 , the root cause analysis model 150 identifies the root cause of the failure of the component 101 to be analyzed based on the quantitative representation 130 .
  • FIG. 3 shows a schematic diagram of the architecture of the root cause analysis model 150 according to some embodiments of the present application.
  • the root cause analysis model 150 generally includes a feature extractor 310 and a classifier 320 .
  • the feature extractor 310 generates N corresponding local failure features 301-1, 301-2, ..., 301-N based on the N quantized representations 130-1, 130-2, ..., 130-N, which are also collectively referred to as There are a plurality of partial failure features 301 or individually referred to as partial failure features 301 . Since each quantitative representation 130 corresponds to a portion 102 of the component 101 to be analyzed, each local failure signature 301 characterizes a failure of the corresponding portion 102 . In the context of chip failure analysis, each local failure feature 301 corresponds to a die of the wafer and characterizes the failure of that die. In this case, the partial failure feature 301 can be viewed as a die feature embedding.
  • the plurality of local failure features 301 are combined into a global failure feature 302 for the component 101 to be analyzed.
  • multiple local failure features 301 are aggregated into a global failure feature 302 .
  • global failure features 302 can be regarded as wafer feature embeddings.
  • multiple partial failure features 301 may be combined in any suitable manner.
  • the global failure feature 302 may be generated by averaging a plurality of local failure features 301 .
  • multiple local failure features 301 may be linked together as a global failure feature 302 .
  • Global failure features 302 are applied to classifier 320 .
  • the classifier 320 determines the probability that different root causes cause the component 101 to be analyzed to fail based on the global failure signature 302 . In other words, classifier 320 generates a root cause probability distribution based on global failure signature 302 .
  • different root causes or categories classified by the classifier 320 may include metal layer open circuit, metal layer short circuit, via short circuit, device problem, and the like.
  • the metal layer disconnection may include ten different root causes such as M0 disconnection, M1 disconnection, M2 disconnection, M3 disconnection, M4 disconnection, M5 disconnection, M6 disconnection, M7 disconnection, M8 disconnection, and M9 disconnection.
  • the metal layer short circuit can include ten different root causes such as M0 short circuit, M1 short circuit, M2 short circuit, M3 short circuit, M4 short circuit, M5 short circuit, M6 short circuit, M7 short circuit, M8 short circuit, and M9 short circuit.
  • Via short circuit can include ten different root causes such as VIA0 short circuit, VIA1 short circuit, VIA2 short circuit, VIA3 short circuit, VIA4 short circuit, VIA5 short circuit, VIA6 short circuit, VIA7 short circuit, VIA8 short circuit, VIA9 short circuit and so on.
  • VIA0 short circuit VIA0 short circuit
  • VIA1 short circuit VIA2 short circuit
  • VIA3 short circuit VIA4 short circuit
  • VIA5 short circuit VIA6 short circuit
  • VIA7 short circuit VIA8 short circuit
  • VIA9 short circuit and so on.
  • any suitable type and number of root causes may be set for the classifier 320 to generate a root cause probability distribution.
  • a counting device (Count Cell) category is introduced for the classifier 320 .
  • the significance of the counted device category is that the more times a certain device appears in the diagnosis report 120 , the higher the probability that the short circuit of the device is the real root cause.
  • count device inferencer 330 may map and recall each count device class.
  • counting device 0 corresponds to the device with the largest value obtained through the recursive tree structure 200
  • counting device 1 corresponds to the device with the second largest value obtained through the recursive tree structure 200
  • the device name of the device with the largest value, the device name of the device with the second largest value, and so on may be extracted from the diagnostic report 120 .
  • the counted device names and corresponding probabilities are output as part of the analysis results 155 .
  • the root cause analysis model 150 is an end-to-end architecture. Therefore, in the application phase, the root cause of failure of the component to be analyzed can be quickly determined by using the root cause analysis model 150 .
  • root cause analysis model 150 described above with reference to FIG. 3 is exemplary only, and is not intended to limit the scope of the present application.
  • the root cause analysis model 150 can also be implemented using other architectures, for example, the operations of the feature extractor and the classifier can be implemented as a whole, that is, implemented by the same network.
  • labeled training samples may be generated by way of error annotation simulation to train the root cause analysis model 150 .
  • the training samples correspond to reference components (e.g., wafers in a mis-insertion simulation), and include quantitative representations related to failure of portions of the reference component (e.g., multiple dies) as well as root causes of reference component failures (e.g., , note the root cause injected by the error simulation).
  • a transferability enhancement operation 340 may also be implemented, which is used to eliminate differences between different processes or designs involved in training. Taking chip failure as an example, if the training involves both a-nanometer process and b-nanometer process (where a is not equal to b), considering that the areas of the same type of devices in these two different processes are different, in order to improve the root cause analysis The accuracy of the model needs to be enhanced with migration capabilities.
  • the transfer capability enhancement operation 340 may perform normalization processing on the features of each reference component (eg, the wafer corresponding to the training sample) involved in the training, so as to eliminate the differences between different processes or designs.
  • the feature extractor 310 and the classifier 320 can be implemented based on any suitable neural network structure.
  • the feature extractor 310 and the classifier 320 can be constructed based on a deep neural network (DNN).
  • DNN deep neural network
  • the feature extractor 310 and the classifier 320 can be constructed based on other types of neural networks, such as convolutional neural networks, recurrent neural networks, and transformers. It should be understood that the scope of the present application is not limited in this respect.
  • the continuous learning module 160 is used to update the trained root cause analysis model 150 as described above with reference to FIG. 1 .
  • the input samples 180 that have not been used to train the root cause analysis model 150 can form a full sample set with the feature samples from the feature library 190 .
  • the continuous learning module 160 can further utilize the full sample set to retrain the root cause analysis model 150 .
  • the input samples 180 and the signature samples from the signature library 190 are labeled samples including the root cause of the failure of the reference element, and a plurality of reference quantified representations respectively associated with the failure of the plurality of reference portions of the reference element. Accordingly, the training of the root cause analysis model 150 using the full sample set is a supervised learning process.
  • continuous learning module 160 may apply multiple reference quantified representations to root cause analysis model 150 to determine potential causes of reference component failures identified by root cause analysis model 150 .
  • the network parameters of the root cause analysis model 150 may be updated by minimizing the difference between the root cause (ie, true value) of the reference component failure and the identified potential cause (ie, predicted value). For example, a loss function can be calculated based on the difference, and the network parameters of the root cause analysis model 150 can be updated by minimizing the loss function.
  • the feature library 190 may include a sample generator.
  • FIG. 4 shows a schematic diagram of updating the root cause analysis model 150 using the sample generator 410 according to some embodiments of the present application.
  • the sample generator 410 is configured to reproduce the characteristics of the historical samples used to train the root cause analysis model in the historical training tasks. In other words, the feature samples for each historical training task generated by the sample generator 410 have the same distribution characteristics as the historical samples in this historical training task. For this reason, the sample generator 410 needs to be trained with historical samples in the historical training task.
  • t is used to denote the current training task corresponding to the input sample 180
  • 1 ⁇ t-1 is used to denote the historical training tasks.
  • the input samples 180 of the current training task can be denoted as D t
  • the feature samples related to the historical training tasks can be denoted as D 1 ⁇ t-1 .
  • the sample generator 410 may use a random number z to generate feature samples for each historical training task, so as to obtain feature samples D 1 ⁇ t ⁇ 1 related to historical training tasks 1 ⁇ t ⁇ 1 . It can be understood that the feature samples D 1 ⁇ t ⁇ 1 are not original diagnosis reports, but marked data, including multiple quantitative representations and corresponding failure root causes.
  • the feature samples D 1 ⁇ t ⁇ 1 and the sample D t of the current task form a full sample set 401 , that is, D 1 ⁇ t .
  • the continuous learning algorithm 170 further trains the root cause analysis model 150 based on the full sample set 401 to update the network parameters of the model.
  • Sample generator 410 may be implemented using any suitable neural network.
  • the sample generator 410 may be implemented using a conditional generative adversarial network (GAN).
  • GAN conditional generative adversarial network
  • the sample generator 410 may also be implemented based on other neural networks, for example, the sample generator 410 may be implemented by using a variational autoencoder (VAE) or a conditional VAE.
  • VAE variational autoencoder
  • the current feature library 190 can only reproduce the sample features of 1-t-1 historical training tasks.
  • the feature library 190 may be expanded based on the input sample D t of the current task t.
  • the sample generator 410 can be retrained using the full sample set D 1 ⁇ t to update the network parameters of the sample generator 410, so that the updated sample generator 410 can reproduce the characteristics of the samples in the 1 ⁇ t training task.
  • FIG. 5 shows a schematic diagram of an extended feature library according to some embodiments of the present application.
  • the sample generator 410 is implemented using a conditional GAN.
  • the training of the sample generator 410 needs the aid of the sample discriminator 520 .
  • the sample generator 410 generates a sample of the training task, and inputs it to the sample discriminator 520 .
  • the samples of the training task in the full sample set 401 are also input to the sample discriminator 520 as the true value.
  • the sample discriminator 520 judges whether the sample from the sample generator 410 is a real sample or a fake sample, and obtains a generative adversarial network loss function L GAN and a cross-entropy loss function L CE .
  • the loss functions L GAN and L CE are used to train both the sample generator 410 and the sample discriminator 520 until the sample discriminator 520 cannot judge whether the samples from the sample generator 410 are real samples or fake samples.
  • the sample generator 410 is updated, and the updated sample generator 410 can reproduce the characteristics of the samples in the 1st to t training tasks. That is, the feature library 190 is expanded.
  • the sample generator 410 is implemented based on a memory replay (Memory Replay) GAN, and can implement generative feature replay. In this way, continuous learning of the root cause analysis model can be realized while saving storage space without forgetting historical tasks.
  • a memory replay Memory Replay
  • FIG. 6 shows a flowchart of a process 600 of determining a failure cause according to some embodiments of the present application.
  • Process 600 may be implemented, for example, by root cause analysis engine 105 in FIG. 1 .
  • the process 600 is described below with reference to FIGS. 1-5 .
  • root cause analysis engine 105 receives a set of input samples for root cause analysis model 150 .
  • a set of input samples may be, for example, samples that have not been used to train root cause analysis model 150 .
  • the root cause analysis model 150 is configured to identify a root cause of failure of the component under analysis 101 based on a plurality of quantitative representations 130 respectively related to failures of portions of the component under analysis 101 .
  • At least one input sample in a set of input samples includes a root cause determined from the root cause analysis model 150 and verified via testing.
  • the component to be analyzed 101 includes a wafer to be analyzed, and the plurality of portions includes a plurality of dies of the wafer to be analyzed.
  • the root cause analysis engine 105 obtains a set of feature samples related to the historical training of the root cause analysis model 150 .
  • Each of the set of input samples and the set of feature samples includes a root cause of failure of the reference element and a plurality of reference quantified representations respectively associated with failure of a plurality of reference portions of the reference element.
  • root cause analysis engine 105 generates a set of characteristic samples according to a sample generator.
  • the sample generator 410 is configured to reproduce the characteristics of the historical samples used to train the root cause analysis model 150 .
  • the set of characteristic samples reflects the characteristics of historical samples.
  • the root cause analysis engine 105 updates the network parameters of the sample generator 410 based on a set of input samples. In this way, the updated sample generator 410 is able to reproduce the characteristics of a set of input samples.
  • the root cause analysis engine 105 trains the root cause analysis model 150 based on the root cause of the reference component failure and the plurality of reference quantified representations.
  • the root cause analysis model 150 is retrained based on a set of input samples and a set of feature samples.
  • the root cause analysis engine 105 determines the potential for reference component failure identified by the root cause analysis model 150 by applying a plurality of reference quantified representations to the root cause analysis model 150. cause; and updating the network parameters of the root cause analysis model 150 by minimizing the difference between the root cause and the potential cause of the reference component failure.
  • the process 600 further includes the root cause analysis engine 105 receiving a plurality of diagnostic reports 120 on the component to be analyzed 101, each diagnostic report corresponding to one of the plurality of parts of the component to be analyzed 101, and indicating The potential root causes of the failure of the part, physical defects and logical errors; the root cause analysis engine 105 generates a plurality of quantitative representations 130 from the plurality of diagnostic reports 120, each related to the failure of the plurality of parts; A number of quantitative representations 130 are applied to a root cause analysis model 150 to identify the root cause of failure of the component 101 under analysis.
  • the root cause analysis engine 105 in order to identify the root cause of failure of the component under analysis 101, the root cause analysis engine 105 generates a plurality of local failure signatures 301, each local failure signature corresponding to one of the plurality of parts; the root cause analysis engine 105 combining the plurality of local failure signatures 301 into a global failure signature 302 for the component to be analyzed 101; The causal analysis engine 105 determines the probability that different root causes cause failure of the component under analysis by applying the global failure feature 302 to the classifier 320 in the root cause analysis model 150 .
  • root cause analysis engine 105 combines quantitative representations of multiple potential root causes indicated in a given diagnostic report into A quantitative representation of a first physical defect associated with a plurality of potential root causes; the root cause analysis engine 105 combines the quantitative representation of the first physical defect and the quantitative representation of at least a second physical defect with the first physical defect and the second physical defect a quantified representation of a first logical error associated with the defect; and root cause analysis engine 105 generating a quantified representation associated with a failure of a part based at least on the quantified representation of the first logical error.
  • the process 600 further includes: before applying the plurality of quantitative representations 130 to the root cause analysis model 150, the root cause analysis engine 105 adjusts the A plurality of quantized representations 130 .
  • Fig. 7 shows a schematic block diagram of an apparatus 700 for determining a failure cause according to some embodiments of the present application.
  • Apparatus 700 may be used to implement root cause analysis engine 105 shown in FIG. 1 .
  • the apparatus 700 includes an inference module 701, such as the inference module 110 shown in FIG. 1 .
  • the reasoning module 701 is configured to identify the root cause of the failure of the component to be analyzed based on the multiple quantitative representations respectively related to the failure of the multiple parts of the component to be analyzed according to the root cause analysis model.
  • the apparatus 700 also includes a learning module 702 , such as the continuous learning module 160 shown in FIG. 1 .
  • the learning module 702 includes an input sample receiving unit 710 configured to receive a set of input samples for the root cause analysis model.
  • the learning module 702 also includes a feature sample acquisition unit 720, which is configured to acquire a set of feature samples related to the historical training of the root cause analysis model, each sample in a set of input samples and a set of feature samples includes a reference component failure A root cause, and a plurality of reference quantifications respectively associated with failures of the plurality of reference portions of the reference element.
  • the learning module 702 further includes a model update unit 730 configured to train a root cause analysis model based on the root cause of failure of the reference component and the plurality of reference quantified representations.
  • the feature sample acquisition unit 720 is further configured to: generate a set of feature samples according to the sample generator.
  • the sample generator is configured to reproduce the characteristics of the historical samples used to train the root cause analysis model, and the set of characteristic samples reflect the characteristics of the historical samples.
  • the learning module 702 further includes: a generator updating unit configured to update network parameters of the sample generator based on a set of input samples.
  • a generator updating unit configured to update network parameters of the sample generator based on a set of input samples.
  • An updated sample generator capable of reproducing the characteristics of a set of input samples.
  • the model update unit 730 is further configured to: determine potential causes of reference element failures identified by the root cause analysis model by applying a plurality of reference quantified representations to the root cause analysis model; and by minimizing the reference The difference between the root cause and the potential cause of component failure updates the network parameters of the root cause analysis model.
  • the reasoning module 701 includes: a diagnosis report receiving unit configured to receive a plurality of diagnosis reports on the element to be analyzed, each diagnosis report corresponding to one of the plurality of parts of the element to be analyzed, and Indicating potential root causes, physical defects, and logical errors of a failure of a part; a quantitative representation generating unit configured to generate, from a plurality of diagnostic reports, a plurality of quantitative representations respectively associated with failures of the plurality of parts; and a root cause identification unit , configured to identify a root cause of failure of the component under analysis by applying the plurality of quantitative representations to the root cause analysis model.
  • the root cause identification unit is further configured to: generate a plurality of partial failure features from the plurality of quantitative representations by applying the plurality of quantitative representations to a feature extractor in the root cause analysis model, each local failure feature Corresponding to one of several sections; combining multiple local failure signatures into a global failure signature for the component under analysis; and identifying different root causes by applying the global failure signature to a classifier in a root cause analysis model Probability of causing failure of the component under analysis.
  • the quantitative representation generation unit is further configured to: for a given diagnostic report of the plurality of diagnostic reports, combine the quantitative representations of the plurality of potential root causes indicated in the given diagnostic report with the plurality of potential root causes a quantitative representation of the first physical defect associated with the cause; combining the quantitative representation of the first physical defect and the quantitative representation of at least a second physical defect into a quantification of a first logical error associated with the first physical defect and the second physical defect representing; and generating a quantified representation related to a failure of a portion based at least on the quantified representation of the first logic error.
  • the reasoning module 701 further includes: a quantitative representation adjustment unit configured to, before applying the multiple quantitative representations to the root cause analysis model, based on the areas related to the potential root causes respectively indicated by the multiple diagnostic reports , to adjust multiple quantized representations.
  • a quantitative representation adjustment unit configured to, before applying the multiple quantitative representations to the root cause analysis model, based on the areas related to the potential root causes respectively indicated by the multiple diagnostic reports , to adjust multiple quantized representations.
  • At least one input sample of a set of input samples includes a root cause determined from a root cause analysis model and verified via testing.
  • the components to be analyzed include a wafer to be analyzed, and the plurality of portions includes a plurality of dies of the wafer to be analyzed.
  • FIG. 8 shows a schematic block diagram of a computing device 800 capable of implementing various embodiments of the present application.
  • the device 800 can be used to implement the root cause analysis engine 105 or at least a portion thereof, such as the reasoning module 110 or the continuous learning module 160.
  • device 800 includes a computing unit 801 that may be loaded into RAM and/or ROM in accordance with computer program instructions stored in random access memory (RAM) and/or read only memory (ROM) 802 or from storage unit 807 802 to perform various appropriate actions and processes.
  • RAM and/or ROM 802 various programs and data necessary for the operation of device 800 may also be stored.
  • the computing unit 801 and the RAM and/or ROM 802 are connected to each other via a bus 803.
  • An input/output (I/O) interface 804 is also connected to the bus 803 .
  • I/O input/output
  • the I/O interface 804 includes: an input unit 805, such as a keyboard, a mouse, etc.; an output unit 806, such as various types of displays, speakers, etc.; a storage unit 807, such as a magnetic disk, an optical disk, etc. ; and a communication unit 808, such as a network card, a modem, a wireless communication transceiver, and the like.
  • the communication unit 808 allows the device 800 to exchange information/data with other devices over a computer network such as the Internet and/or various telecommunication networks.
  • the computing unit 801 may be various general-purpose and/or special-purpose processing components having processing and computing capabilities. Some examples of computing units 801 include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processing processor (DSP), and any suitable processor, controller, microcontroller, etc.
  • the computing unit 801 executes various methods and processes described above, such as the process 600 .
  • process 600 may be implemented as a computer software program tangibly embodied on a machine-readable medium, such as storage unit 807 .
  • part or all of the computer program may be loaded and/or installed onto device 800 via RAM and/or ROM and/or communication unit 808 .
  • a computer program When a computer program is loaded into RAM and/or ROM and executed by computing unit 801, one or more steps of process 600 described above may be performed.
  • the computing unit 801 may be configured to execute the process 600 in any other suitable manner (for example, by means of firmware).
  • Program codes for implementing the methods of the present application may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special purpose computer, or other programmable data processing devices, so that the program codes, when executed by the processor or controller, make the functions/functions specified in the flow diagrams and/or block diagrams Action is implemented.
  • the program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.
  • a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • a machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium.
  • a machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing.
  • machine-readable storage media would include one or more wire-based electrical connections, portable computer discs, hard drives, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.
  • RAM random access memory
  • ROM read only memory
  • EPROM or flash memory erasable programmable read only memory
  • CD-ROM compact disk read only memory
  • magnetic storage or any suitable combination of the foregoing.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Quality & Reliability (AREA)
  • Test And Diagnosis Of Digital Computers (AREA)

Abstract

本申请涉及用于确定失效原因的方法、装置、设备、存储介质和程序产品。在确定失效原因的方法中，获得根因分析模型。该根因分析模型被配置为基于分别与待分析元件的多个部分的失效有关的多个量化表示，来标识待分析元件失效的根本原因。如果接收到一组输入样本，获取与根因分析模型的历史训练有关的一组特征样本。每个样本包括分别与参考元件的多个参考部分的失效有关的多个参考量化表示和参考元件失效的根本原因。接下来，基于参考元件失效的根本原因和多个参考量化表示，更新根因分析模型。以此方式，可以实现根因分析模型的持续学习，使根因分析模型对不同设计和工艺具有稳定表现，提供高准确性的根因确定结果。

Description

一种确定失效原因的方法及装置 技术领域
本申请主要涉及失效分析领域。更具体地，涉及用于一种确定失效原因的方法、装置、设备、计算机可读存储介质以及计算机程序产品。
背景技术
在半导体制造过程中，保证集成电路芯片的稳定高产能对于质量、可靠性和利润来说都是至关重要的。随着摩尔定律的演进，半导体工艺关键尺寸不断缩小，新设计和新工艺中存在的问题越来越难以定位和分析。针对出现的问题，可以基于布局感知分析技术分析得到可能导致问题在物理上的根本原因(也简称为“根因”)。但布局感知分析技术得到的分析结果存在较大的噪声，例如某个分析结果可能会指向多个疑似区域，一个疑似区域可能包含多个根因等。这给进一步分析以确定真正的根因带来了困难。在这种情况下，通过使用根因分析(RCA)技术进行进一步筛查。
发明内容
本申请公开的实施例提供了一种确定失效原因的方案。
在本申请公开的第一方面，提供了一种用于确定失效原因的装置。该装置包括：推理模块，被配置为基于分别与待分析元件的多个部分的失效有关的多个量化表示，根据根因分析模型，来标识待分析元件失效的根本原因。该装置还包括与推理模块耦合的学习模块。学习模块包括：输入样本接收单元，被配置为接收针对根因分析模型的一组输入样本；特征样本获取单元，被配置为获取与根因分析模型的历史训练有关的一组特征样本，一组输入样本和一组特征样本中的每个样本包括参考元件失效的根本原因、和分别与参考元件的多个参考部分的失效有关的多个参考量化表示；以及模型更新单元，被配置为基于参考元件失效的根本原因和多个参考量化表示，训练根因分析模型。
推理模块可以利用根因分析模型来快速确定被分析元件失效的根因。学习模块可以利用新工艺或新设计的新增样本和与历史训练有关的特征样本，来更新根因分析模型。以此方式，可以使根因分析模型针对不同设计和工艺进行适应性调整。这可以提升根因分析模型对新工艺和新设计的识别精度，同时保证对历史全任务的稳定表现，从而提供更高精度的根因确定结果。此外，根因分析模型的更新可以自动完成，从而节约人力、物料和时间成本并提升效率。
在第一方面的一些实施例中，特征样本获取单元进一步被配置为：根据样本生成器，生成一组特征样本，样本生成器被配置为再现用于训练根因分析模型的历史样本的特征，并且一组特征样本反映历史样本的特征。在这种实施例中，利用样本生成器，可以减小存储历史样本所消耗的存储空间。以此方式，可以以节省资源的方式实现根因分析模型的更新。
在第一方面的一些实施例中，学习模块还包括：生成器更新单元，被配置为基于一组输入样本，更新样本生成器的网络参数。以此方式，经更新的样本生成器能够再现一组输入样本的特征。可以使经更新的样本生成器用于根因分析模型的下一次更新，从而实现根因分析模型稳定的持续学习和更新。
在第一方面的一些实施例中，模型更新单元进一步被配置为：通过将多个参考量化表示应用于根因分析模型，确定由根因分析模型标识的、参考元件失效的潜在原因；以及通过最小化参考元件失效的根本原因与潜在原因之间的差异，更新根因分析模型的网络参数。以此方式，可以利用监督训练过程来重新确定根因分析模型的参数，以保证根因分析模型对历史全任务的稳定表现。
在第一方面的一些实施例中，推理模块包括：诊断报告接收单元，被配置为接收关于待分析元件的多个诊断报告，每个诊断报告对应于待分析元件的多个部分中的一个部分，并且指示一个部分失效的潜在根本原因、物理缺陷和逻辑错误；量化表示生成单元，被配置为从多个诊断报告，生成分别与多个部分的失效有关的多个量化表示；以及根本原因标识单元，被配置为通过将多个量化表示应用于根因分析模型，标识待分析元件失效的根本原因。以此方式，利用端到端的根因分析模型可以快速确定待分析元件失效的根本原因。
在第一方面的一些实施例中，根本原因标识单元进一步被配置为：通过将多个量化表示应用于根因分析模型中的特征提取器，从多个量化表示生成多个局部失效特征，每个局部失效特征与多个部分中的一个部分相对应；将多个局部失效特征组合为针对待分析元件的全局失效特征；以及通过将全局失效特征应用于根因分析模型中的分类器，确定不同的根本原因引起待分析元件失效的概率。以此方式，实现了从局部到全局的整合，这有利于更准确地确定待分析元件失效的根本原因。
在第一方面的一些实施例中，量化表示生成单元进一步被配置为：针对多个诊断报告中的给定诊断报告，将给定诊断报告中指示的多个潜在根本原因的量化表示组合成与多个潜在根本原因相关联的第一物理缺陷的量化表示；将第一物理缺陷的量化表示和至少第二物理缺陷的量化表示组合成与第一物理缺陷和第二物理缺陷相关联的第一逻辑错误的量化表示；以及至少基于第一逻辑错误的量化表示，生成与一个部分的失效有关的量化表示。在这种实施例中，以从底向上的方式来构建与失效有关的量化表示，使得所构建的量化表示能够反映待分析元件的局部失效。
在第一方面的一些实施例中，推理模块还包括：量化表示调整单元，被配置为在将多个量化表示应用于根因分析模型之前，基于与多个诊断报告分别指示的潜在根本原因有关的面积，来调整多个量化表示。以此方式，可以减小或甚至消除不同的根因所对应的元件面积的影响，这有助于进一步提高根因确定的准确性。
在第一方面的一些实施例中，一组输入样本中的至少一个输入样本所包括的根本原因是根据根因分析模型确定并且经由测试验证的。以此方式，可以利用新设计或新工艺中的真实样本来更新根因分析模型。这使得经更新的根因分析模型更好地适应新设计或新工艺。
在第一方面的一些实施例中，待分析元件包括待分析晶圆，并且多个部分包括待分析晶圆的多个裸片。半导体工艺的发展迅速，新设计和新工艺不断出现。将可持续学习的根因分析模型应用于晶圆，能够有效节约资源和成本，提高效率。
在本申请公开的第二方面，提供了一种确定失效原因的方法。该方法包括：接收针对根因分析模型的一组输入样本，根因分析模型被配置为基于分别与待分析元件的多个部分的失效有关的多个量化表示，来标识待分析元件失效的根本原因；获取与根因分析模型的历史训练有关的一组特征样本，一组输入样本和一组特征样本中的每个样本包括参考元件失效的根本原因、和分别与参考元件的多个参考部分的失效有关的多个参考量化表示；以及基于参考元件失效的根本原因和多个参考量化表示，训练根因分析模型。
利用根因分析模型可以快速确定被分析元件失效的根因。利用新工艺或新设计的新增样本和与历史训练有关的特征样本来更新根因分析模型，可以使根因分析模型针对不同设计和工艺进行适应性调整。这可以提升根因分析模型对新工艺和新设计的识别精度，同时保证对历史全任务的稳定表现，从而提供更高准确性的根因确定结果。此外，根因分析模型的更新可以自动完成，从而节约人力、物料和时间成本并提升效率。
在第二方面的一些实施例中，获取一组特征样本包括：根据样本生成器，生成一组特征样本，样本生成器被配置为再现用于训练根因分析模型的历史样本的特征，并且一组特征样本反映历史样本的特征。在这种实施例中，利用样本生成器，可以减小存储历史样本所消耗的存储空间。以此方式，可以以节省资源的方式实现根因分析模型的更新。
在第二方面的一些实施例中，该方法还包括：基于一组输入样本，更新样本生成器的网络参数。以此方式，经更新的样本生成器能够再现一组输入样本的特征。可以使经更新的样本生成器用于根因分析模型的下一次更新，从而实现根因分析模型稳定的持续学习和更新。
在第二方面的一些实施例中，训练根因分析模型包括：通过将多个参考量化表示应用于根因分析模型，确定由根因分析模型标识的、参考元件失效的潜在原因；以及通过最小化参考元件失效的根本原因与潜在原因之间的差异，更新根因分析模型的网络参数。以此方式，可以利用监督训练过程来重新确定根因分析模型的参数，以保证根因分析模型对历史全任务的稳定表现。
在第二方面的一些实施例中，该方法还包括：接收关于待分析元件的多个诊断报告，每个诊断报告对应于待分析元件的多个部分中的一个部分，并且指示一个部分失效的潜在根本原因、物理缺陷和逻辑错误；从多个诊断报告，生成分别与多个部分的失效有关的多个量化表示；以及通过将多个量化表示应用于根因分析模型，标识待分析元件失效的根本原因。以此方式，利用端到端的根因分析模型可以快速确定待分析元件失效的根本原因。
在第二方面的一些实施例中，标识待分析元件失效的根本原因包括：通过将多个量化表示应用于根因分析模型中的特征提取器，从多个量化表示生成多个局部失效特征，每个局部失效特征与多个部分中的一个部分相对应；将多个局部失效特征组合为针对待分析元件的全局失效特征；以及通过将全局失效特征应用于根因分析模型中的分类器，确定不同的根本原因引起待分析元件失效的概率。以此方式，实现了从局部到全局的整合，这有利于更准确地确定待分析元件失效的根本原因。
在第二方面的一些实施例中，生成多个量化表示包括：针对多个诊断报告中的给定诊断报告，将给定诊断报告中指示的多个潜在根本原因的量化表示组合成与多个潜在根本原因相关联的第一物理缺陷的量化表示；将第一物理缺陷的量化表示和至少第二物理缺陷的量化表示组合成与第一物理缺陷和第二物理缺陷相关联的第一逻辑错误的量化表示；以及至少基于第一逻辑错误的量化表示，生成与一个部分的失效有关的量化表示。在这种实施例中，以从底向上的方式来构建与失效有关的量化表示，使得所构建的量化表示能够反映待分析元件的局部失效。
在第二方面的一些实施例中，该方法还包括：在将多个量化表示应用于根因分析模型之前，基于与多个诊断报告分别指示的潜在根本原因有关的面积，来调整多个量化表示。以此方式，可以减小或甚至消除不同的根因所对应的元件面积的影响，这有助于进一步提高根因确定的准确性。
在第二方面的一些实施例中，一组输入样本中的至少一个输入样本所包括的根本原因是 根据根因分析模型确定并且经由测试验证的。以此方式，可以利用新设计或新工艺中的真实样本来更新根因分析模型。这使得经更新的根因分析模型更好地适应新设计或新工艺。
在第二方面的一些实施例中，待分析元件包括待分析晶圆，并且多个部分包括待分析晶圆的多个裸片。半导体工艺的发展迅速，新设计和新工艺不断出现。将可持续学习的根因分析模型应用于晶圆，能够有效节约资源和成本，提高效率。
在本申请公开的第三方面，提供了一种电子设备，包括：至少一个处理器；至少一个存储器，至少一个存储器被耦合到至少一个处理器并且存储用于由至少一个处理器执行的指令，指令当由至少一个处理器执行时，使得设备实现第二方面中的任意一种实现方式中的方法。
在本申请公开的第四方面，提供了一种计算机可读存储介质，其上存储有一条或多条计算机指令，其中一条或多条计算机指令被处理器执行实现第二方面中的任意一种实现方式中的方法。
在本申请公开的第五方面，提供一种计算机程序产品，当计算机程序产品在计算机上运行时，使得计算机执行第二方面中的任意一种实现方式中的方法的部分或全部步骤的指令。
可以理解地，上述提供的第三方面的电子设备、第四方面的计算机存储介质或者第五方面的计算机程序产品均用于执行第二方面所提供的方法。因此，关于第二方面的解释或者说明同样适用于第三方面、第四方面和第五方面。此外，第三方面、第四方面和第五方面所能达到的有益效果可参考对应方法中的有益效果，此处不再赘述。
应当理解，发明内容部分中所描述的内容并非旨在限定本申请的实施例的关键或重要特征，亦非用于限制本申请的范围。本申请的其它特征将通过以下的描述变得容易理解。
附图说明
结合附图并参考以下详细说明，本申请的上述和其他特征、优点及方面将变得更加明显。在附图中，相同或相似的附图标记表示相同或相似的元素，其中：
图1示出了本申请的多个实施例能够在其中实现的示例环境的示意图；
图2示出了根据本申请的一些实施例的从诊断报告构建量化表示的示意图；
图3示出了根据本申请的一些实施例的根因分析模型的架构的示意图；
图4示出了根据本申请的一些实施例的更新根因分析模型的示意图；
图5示出了根据本申请的一些实施例的扩展特征库的示意图；
图6示出了根据本申请的一些实施例的确定失效原因的过程的流程图；
图7示出了根据本申请的一些实施例的用于确定失效原因的装置的示意性框图；以及
图8示出了能够实施本申请的多个实施例的计算设备的示意性框图。
具体实施方式
下面将参照附图更详细地描述本申请的实施例。虽然附图中显示了本申请的某些实施例，然而应当理解的是，本申请可以通过各种形式来实现，而且不应该被解释为限于这里阐述的实施例，相反提供这些实施例是为了更加透彻和完整地理解本申请。应当理解的是，本申请的附图及实施例仅用于示例性作用，并非用于限制本申请的保护范围。
在本申请的实施例的描述中，术语“包括”及其类似用语应当理解为开放性包含，即“包括但不限于”。术语“基于”应当理解为“至少部分地基于”。术语“一个实施例”或“该实施例”应当理解为“至少一个实施例”。术语“第一”、“第二”等等可以指代不同的或相同的对象。术语“和 /或”表示由其关联的两项的至少一项。例如“A和/或B”表示A、B、或者A和B。下文还可能包括其他明确的和隐含的定义。
如本文所使用的，“神经网络”能够处理输入并且提供相应输出，其通常包括输入层和输出层以及在输入层与输出层之间的一个或多个隐藏层。在深度学习应用中使用的神经网络通常包括许多隐藏层，从而延长网络的深度。神经网络的各个层按顺序相连，从而前一层的输出被提供作为后一层的输入，其中输入层接收神经网络的输入，而输出层的输出作为神经网络的最终输出。在本文中，术语“神经网络”、“网络”、“神经网络模型”和“模型”可替换地使用。
如上文所简要提及的，需要使用根因分析技术来确定问题的根本原因。广义上，根因分析是指标识问题或故障的根本原因的过程。具体到半导体领域，根因分析可以用于发现导致芯片失效的根本原因，从而有效保证芯片在量产过程中的高产能。
在半导体领域中，为了定位新设计或新工艺中的问题，通常首先进行芯片测试。芯片测试的主要流程为：首先根据自动化测试设备(ATE)产生的测试向量对芯片或芯片的一部分进行测试，根据逻辑电路输出结果从错误词典中找到产生的错误类型，从而获得ATE失效数据；结合网表图案布局，对ATE失效数据中失效的引脚进行布局感知诊断，并结合先验知识，生成诊断报告。诊断报告主要包括关键面积、疑似区域等。根因分析主要针对生成的诊断报告进行分析，得到导致芯片失效的根因。具体地，利用根因分析，可以从诊断报告中推断出各种根因发生的概率分布，从而定位主要根因，以解决芯片失效问题。
在常规的RCA方案中，使用贝叶斯模型计算各个根因的概率分布。基于真实物理关系分析，可在根因、物理缺陷和逻辑错误之间建立相应的概率联系。在该RCA方案中，可以从失效产生的诊断报告中提取相应数据并计算相应的转移概率。使用基于非监督学习的期望最大(EM)算法进行求解，从而得到最能够解释一系列诊断报告的根因分布。EM算法的求解过程包括迭代进行的期望步骤和最大化步骤。在迭代的期望步骤中，使用当前迭代中第k个根因c k的概率分布P(c k)′来估计第n个诊断报告r n的后验概率P(c k|r n)′。在迭代的最大化步骤中，通过最大化N个诊断报告的后验概率P(c k|r n)′之和来更新根因的概率分布P(c k)′。
这种常规方法存在很大不足，因为不同工艺对根因分析准确性影响较大。特别是随着技术的不断演进，工艺尺寸越来越小，中段工艺层(即较底层的金属层)的根因更加难以区分。每次工艺更新需要依赖各个相关方的合作交流进行提升，耗费较大的人力和时间成本。
上文以芯片失效中的根因分析为例，描述了常规RCA方案的问题。在其他一些系统性体系的失效分析中，也存在类似问题。例如，航天器是由各种复杂且精细的部件构成的体系。在航天器的失效分析中，也需要使用RCA确定失效的根因。
为了至少部分地解决上述问题以及其他潜在问题，本申请的各种实施例提供了一种用于确定失效原因的方案。总体而言，根据在此描述的各种实施例，首先获得根因分析模型。该根因分析模型被配置为基于分别与待分析元件的多个部分的失效有关的多个量化表示，来标识待分析元件失效的根本原因。如果接收到针对根因分析模型的一组输入样本，获取与根因分析模型的历史训练有关的一组特征样本。每个样本包括分别参考元件失效的根本原因、和与参考元件的多个参考部分的失效有关的多个参考量化表示。换言之，输入样本和特征样本是带标记样本。接下来，基于输入样本和特征样本所包括的多个参考量化表示和参考元件失效的根本原因，再次训练根因分析模型以更新根因分析模型的网络参数。
在本申请的实施例中，利用根因分析模型可以快速确定被分析元件失效的根因。利用新 工艺或新设计的新增样本和与历史训练有关的特征样本来更新根因分析模型，可以实现根因分析模型的持续学习，使根因分析模型针对不同设计和工艺进行适应性调整。这可以提升根因分析模型对新工艺和新设计的识别准确性，同时保证对历史全任务的稳定表现，从而提供更高准确性的根因确定结果。此外，根因分析模型的更新可以自动完成，从而节约人力、物料和时间成本并提升效率。
以下参考附图来描述本申请的各种示例实施例。
示例环境
图1示出了本申请的多个实施例能够在其中实现的示例环境100的示意图。总体上，示例环境100包括根因分析引擎105和关于待分析元件101的多个诊断报告120-1、120-2、……、120-N，其也统称为多个诊断报告120或单独称为诊断报告120，其中N为大于等于1的整数。应当理解，图1所示的环境100仅是示例性的，而无意限制本申请的范围。
如图1所示，待分析元件101包括N个部分102-1、102-2、……、102-N，其也统称为多个部分102或单独称为部分102。在图1的示例中，待分析元件101是晶圆(wafer)，而多个部分102是该晶圆所包括的多个裸片(die)。在另一示例中，待分析元件101可以是航天器，而多个部分可以是该航天器所包括的多个部件。
每个诊断报告120对应于待分析元件101的一个部分，并且指示与该部分失效有关的诊断信息，例如失效的潜在根因、物理缺陷和逻辑错误等。作为示例，诊断报告120-2指示与部分102-2失效有关的诊断信息，例如部分102-2失效的潜在根因、物理缺陷和逻辑错误等。诊断报告120可以是以任何合适的方式生成的，本申请的实施例在此方面不受限制。
根因分析引擎105包括推理模块110和持续学习模块160(也可以简称为“学习模块”)。推理模块110用于根据关于待分析元件101的多个诊断报告120，来标识待分析元件101失效的根因。如图1所示，推理模块110接收多个诊断报告120，并且将诊断报告中关于失效的诊断信息转换为量化表示。具体地，推理模块110从多个诊断报告120，生成分别与多个部分102的失效有关的N个量化表示130-1、130-2、……、130-N，其也统称为多个量化表示130或单独称为量化表示130。每个量化表示130从一个诊断报告120得出。从诊断报告120得出的这种量化表示130可以视为待分析元件101失效的局部原始特征。在一些实施例中，推理模块110还可以包括调整操作140。调整操作140利用权重对每个量化表示130进行调整，以消除不同的根因所对应于的器件面积的影响。
量化表示130或经调整的量化表示130被应用到根因分析模型150。根因分析模型150基于接收到的量化表示130，生成分析结果155，分析结果155标识待分析元件101失效的根因。例如，分析结果155可以包括针对不同根因的概率。又如，分析结果可以包括最可能的预定数目个(例如，一个)根因。在本申请的实施例应用于芯片失效分析的情况下，不同根因可以包括但不限于，金属层断路(诸如M1断路、M2断路等)，金属层短路(诸如M1短路、M2短路等)，过孔(VIA)短路(例如VIA1短路、VIA2短路等)，较大的器件出现问题(在本文中称为“器件问题”)，例如与周围部分发生短路等。
持续学习模块160用于训练根因分析模型150，以更新根因分析模型150的网络参数。具体地，持续学习模块160接收一组输入样本180。输入样本180可以是或可以包括尚未用于训练根因分析模型150的样本。例如，输入样本180可以是新工艺或新设计中产生的新增样本。输入样本180是带标记样本，包括：参考元件失效的根因，以及分别与参考元件的多 个参考部分的失效有关的多个参考量化表示。
在一些实施例中，输入样本180可以是应用根因分析模型150被分析的样本。具体地，输入样本180所包括的标记(即，参考元件失效的根因)是根据根因分析模型150确定，并经过测试验证的，如虚线箭头所示。例如，在芯片失效的应用场景中，推理模块110基于失效晶圆的多个裸片的诊断报告，标识该晶圆失效的主要根因。然后，可以在实验上针对该主要根因进行物理验证分析(PFA)。如果由根因分析模型150标识的主要根因通过了PFA验证，则与该晶圆有关的数据可以作为输入样本180由持续学习模块160使用。
备选地或附加地，在一些实施例中，输入样本180可以是通过注错仿真生成的样本。例如，在发展了根因分析模型150的历史训练中未学习过的新工艺或新设计后，可以通过注错仿真生成针对新工艺或新设计的新增样本，以作为用于更新根因分析模型150的输入样本180。以此方式，可以在新工艺或新设计出现后，快速生成新增样本，使经更新的根因分析模型150迅速适用于新工艺或新设计。
持续学习模块160包括特征库190。特征库190用于再现(例如，存储或生成)与根因分析模型150的历史训练有关的样本，其也称为“特征样本”。来自特征库190的特征样本与输入样本180一起组成全样本集合。持续学习模块160进而利用持续学习算法170，基于全样本集合来再次训练根因分析模型150。附加地，输入样本180可以用于扩展特征库190，以使经扩展的特征库190在后续更新中再现输入样本180。
在一些实施例中，特征库190可以包括存储装置，用于存储历史训练任务中用于训练根因分析模型150的历史样本。例如，在历史样本的数量较小的情况下，可以在训练任务结束后，将历史样本存储在特征库190中。在这种实施例中，在接收到输入样本180后，可以从特征库190中直接取回历史样本作为特征样本，并与输入样本一起组成全样本集合。如果输入样本180的数量较少，可以直接将输入样本180存储在特征库190中，以扩展特征库190。
备选地或附加地，在一些实施例中，特征库190可以包括样本生成器。样本生成器被配置为再现在历史训练任务中用于训练根因分析模型的历史样本的特征。在这种实施例中，在接收到输入样本180后，可以根据样本生成器，生成能够反映历史样本的特征的特征样本，并与输入样本一起组成全样本集合。可以基于输入样本180重新训练样本生成器，以扩展特征库190。下文将参考图4和图5来详细描述这种实施例。
应当理解，图1所示的环境100仅是示例性的，而无意限制本申请的范围。可以以任何合适的方式对根因分析引擎105的各个功能进行划分。例如，从诊断报告120得到量化表示130的功能可以独立于推理模块110而实现。也即，在一些实施例中，推理模块110可以接收量化表示130，并将量化表示130或经调整的量化表示130应用于根因分析模型150。在一些实施例中，推理模块110和持续学习模块160可以实现在相同的计算设备或计算系统上。在另一些实施例中，推理模块110和持续学习模块160可以实现不同但彼此通信的计算设备或计算系统上。
推理模块的示例实现
如图1所示，可以从诊断报告120构建待分析元件101的一部分102失效的量化表示130。现在参考图2，图2示出了根据本申请的一些实施例的从诊断报告120构建量化表示130的示意图。诊断报告120可以是图1所示的多个诊断报告120-1、120-2、……、120-N中的任一个。可以按如图2所示的递归树结构200来组织诊断报告120中与失效有关的信息。
递归树包括自底向上的多个层级，例如根因层级、物理缺陷层级和逻辑缺陷层级。根因层级包括由诊断报告120指示的引起对应的部分102失效的潜在的多个根因210-1、210-2、210-3，也统称为多个根因210或单独称为根因210。与根因210有关的信息可以包括根因的名称和关键面积。关键面积是指对应的根因所涉及的器件的面积。
根因层级之上的物理缺陷层级包括多个子疑似区域220-1、220-2、220-3，其也统称为多个子疑似区域220或单独称为疑似区域220。每个子疑似区域220可以是一个物理缺陷，并与根因层级的至少一个根因相关联。例如，子疑似区域220-1与根因210-1、210-2、210-3等相关联。与子疑似区域220有关的信息可以包括缺陷的类型和由诊断报告120给出的得分。该得分可以表示缺陷的发生概率。
在物理缺陷层级之上是逻辑错误层级，其又可以划分为两个子层级。较低的子层级包括多个疑似区域230-1、230-2、230-3，其统称为多个疑似区域230或单独称为疑似区域230。每个疑似区域230与多个子疑似区域相关联，例如可以包括多个子疑似区域。在图2的示例中，疑似区域230-1与子疑似区域220-1、220-2、220-3等相关联。
较高的子层级包括多个症状240-1、240-2、240-3，其也统称为多个症状240或单独地称为症状240。症状240可以是指在测试中直接表现出的错误。每个症状240与多个疑似区域相关联。例如，症状240-1与疑似区域230-1、230-2、230-3相关联。
推理模块110可以基于递归树结构200，通过自底向上递归的方式来构建量化表示130。具体地，针对每个根因210，推理模块110可以从诊断报告120中提取与该根因有关的信息，例如根因的名称(诸如，M1断路、VIA2短路等)和关键面积。然后，将所提取的信息量化，以生成该根因的量化表示(例如，向量)。该根因的量化表示和其他根因的量化表示被组合成上一层级的相关联的子疑似区域220的量化表示。在图2的示例中，根因210-1、210-2、210-3等的量化表示被组合成子疑似区域220-1的量化表示。
针对每个子疑似区域220，推理模块110可以从诊断报告120中提取与相应子疑似区域220有关的信息，例如类型和得分。基于所提取的信息，该子疑似区域的量化表示和其他子疑似区域的量化表示被组合成上一层级的相关联的疑似区域230的量化表示。在图2的示例中，子疑似区域220-1、220-2、220-3的量化表示被组合成疑似区域230-1的量化表示。
类似地，每个疑似区域230的量化表示和其他的疑似区域的量化表示被组合成上一层级的相关联的症状240的量化表示。在图2的示例中，疑似区域230-1、230-2、230-3的量化表示被组合成症状240-1的量化表示。最终，多个症状240-1、240-2、240-3的量化表示被组合成与诊断报告120相对应的量化表示130。量化表示130可以被实现为一维向量，但不限于此。每个量化表示130对应于待分析元件101的一个部分102，并且是待分析元件101失效的局部原始特征。
在本申请的实施例中，可以通过任何合适的方式将下一层级的多个量化表示组合成上一层级的量化表示。在一个示例中，可以对下一层级的多个量化表示进行平均或取最大值，以生成上一层级的量化表示。例如，可以通过对根因210-1、210-2、210-3的量化表示进行平均，来生成子疑似区域220-1的量化表示。在另一示例中，可以将下一层级的多个量化表示作为不同的通道进行联接(concatenate)，将联接在一起的多个量化表示作为上一层级的量化表示。例如，作为不同通道而联接的根因210-1、210-2、210-3的量化表示可以用作子疑似区域220-1的量化表示。
返回参考图1。在一些实施例中，推理模块110还可以包括调整操作140。调整操作140 可以根据下式对每个量化表示130进行调整：
Figure PCTCN2021101866-appb-000001
其中ReportFeat i表示从第i个诊断报告提取的量化表示，w i表示针对第i个诊断报告的调整权重，i的取值为1至N。
权重w i用于减小或甚至消除不同的根因所对应的器件面积的影响。因此，权重w i可以是基于诊断报告120所指示的不同根因的关键面积(如上文参考图2所描述的)而确定的。在一个示例中，可以从关于待分析元件101的多个诊断报告120提取的RCA常数作为权重w i。在这种情况下，权重w i对于所有诊断报告是相同的。可以理解的是，权重w i具有与ReportFeat i相同的维度，式(1)的计算是逐元素的。在这种实施例中，通过减小与不同根因有关的面积的影响，可以使根因分析模型150最终标识的根因更为准确。
在获得量化表示130之后，根因分析模型150基于量化表示130，标识待分析元件101失效的根因。参考图3，图3示出了根据本申请的一些实施例的根因分析模型150的架构的示意图。在图3的示例中，根因分析模型150总体上包括特征提取器310和分类器320。
特征提取器310基于N个量化表示130-1、130-2、……、130-N，生成对应的N个局部失效特征301-1、301-2、……、301-N，其也统称为多个局部失效特征301或单独称为局部失效特征301。由于每个量化表示130对应于待分析元件101的一个部分102，因此每个局部失效特征301表征相应的部分102的失效。在芯片失效分析的场景中，每个局部失效特征301对应于晶圆的一个裸片，并且表征该裸片的失效。在这种情况下，局部失效特征301可以视为裸片特征嵌入。
接下来，多个局部失效特征301被组合成针对待分析元件101的全局失效特征302。换言之，多个局部失效特征301被聚合成全局失效特征302。在芯片失效分析的场景中，全局失效特征302可以视为晶圆特征嵌入。在本申请的实施例中，可以通过任何合适的方式来组合多个局部失效特征301。例如，可以通过对多个局部失效特征301进行平均，来生成全局失效特征302。又如，可以将多个局部失效特征301联接在一起，作为全局失效特征302。
全局失效特征302被应用于分类器320。分类器320基于全局失效特征302确定不同的根因引起待分析元件101失效的概率。换言之，分类器320基于全局失效特征302生成根因概率分布。
以芯片失效分析为例，不同的根因或分类器320所划分的类别可以包括金属层断路、金属层短路、过孔短路、器件问题等。作为示例，金属层断路可以包括M0断路、M1断路、M2断路、M3断路、M4断路、M5断路、M6断路、M7断路、M8断路、M9断路等十个不同的根因。金属层短路可以包括M0短路、M1短路、M2短路、M3短路、M4短路、M5短路、M6短路、M7短路、M8短路、M9短路等十个不同的根因。过孔短路可以包括VIA0短路、VIA1短路、VIA2短路、VIA3短路、VIA4短路、VIA5短路、VIA6短路、VIA7短路、VIA8短路、VIA9短路等十个不同的根因。应当理解，以上所列举的类别及其数目以及类别下的具体根因及其数目仅是示例性的，而无意限制本申请的范围。在本申请的各种实施例中，可以针对分类器320设置任何合适类型和数目的根因来生成根因概率分布。
与金属层断路、金属层短路、过孔短路不同，器件问题是难以枚举的类型。为此，在本申请的一些实施例中，针对分类器320引入计数型器件(Count Cell)类别。计数型器件类别的意义在于，某个器件在诊断报告120中出现的次数越多，该器件的短路是真正根因的概率越大。在输出各个计数型器件类别(例如，计数型器件0至计数型器件9)之后，计数型器 件推断器330可以对每个计数型器件类别进行映射和召回。例如，计数型器件0对应于通过递归树结构200得到的值最大的器件，计数型器件1对应于通过递归树结构200得到的值第二大的器件，以此类推。然后，可以从诊断报告120中提取值最大的器件的器件名称、值第二大的器件的器件名称，以此类推。将计数型器件名称及对应的概率作为分析结果155的一部分输出。
从以上描述中可以看出，根因分析模型150是一种端到端的架构。因此，在应用阶段，利用根因分析模型150可以快速确定待分析元件失效的根因。
应当理解，以上参考图3描述的根因分析模型150的架构仅是示例性的，而无意限制本申请的范围。也可以利用其他架构来实现根因分析模型150，例如特征提取器和分类器的作业可以整体实现，即由同一个网络来实现。
在训练阶段，可以通过注错仿真的方式生成带标记的训练样本，以对根因分析模型150进行训练。训练样本对应于参考元件(例如，注错仿真中的晶圆)，并且包括与参考元件的多个部分(例如，多个裸片)的失效有关的量化表示以及参考元件失效的根因(例如，注错仿真所注入的根因)。
在训练阶段，在一些实施例中，还可以实现迁移能力增强操作340，其用于消除训练中涉及的不同工艺或设计之间的差异。以芯片失效为例，如果训练中既涉及a纳米工艺也涉及b纳米工艺(其中，a不等于b)，考虑到这两种不同工艺中相同类型的器件的面积不同，因此为了提高根因分析模型的准确性，需要进行迁移能力增强。例如，迁移能力增强操作340可以对训练中涉及的每个参考元件(例如，训练样本对应的晶圆)的特征做归一化处理，从而消除不同工艺或设计之间的差异。
在本申请的实施例中，特征提取器310和分类器320可以基于任何合适的神经网络结构来实现。在一些实施例中，可以基于深度神经网络(DNN)来构建特征提取器310和分类器320。在另一些实施例中，可以基于其他类型的神经网络来构建特征提取器310和分类器320，例如卷积神经网络、循环神经网络、变换器(Transformer)等。应当理解，本申请的范围在此方面不受限制。
持续学习模块的示例实现
返回参考图1，如上文参考图1所描述的，持续学习模块160用于更新经训练的根因分析模型150。具体地，尚未用于训练根因分析模型150的输入样本180可以与来自特征库190的特征样本组成全样本集合。持续学习模块160进而可以利用全样本集合来再次训练根因分析模型150。输入样本180和来自特征库190的特征样本是带标记样本，包括参考元件失效的根因、以及分别与参考元件的多个参考部分的失效有关的多个参考量化表示。相应地，利用全样本集合对根因分析模型150的训练是监督学习过程。例如，持续学习模块160可以将多个参考量化表示应用于根因分析模型150，以确定由根因分析模型150标识的、参考元件失效的潜在原因。通过最小化参考元件失效的根本原因(即，真值)与所标识的潜在原因(即，预测值)之间的差异，可以更新根因分析模型150的网络参数。例如，可以基于该差异来计算损失函数，并通过最小化损失函数来更新根因分析模型150的网络参数。
鉴于直接存储历史训练中使用的历史样本占用大量的存储空间，在一些实施例中，特征库190可以包括样本生成器。现在参考图4，图4示出了根据本申请的一些实施例的利用样本生成器410来更新根因分析模型150的示意图。
样本生成器410被配置为再现在历史训练任务中用于训练根因分析模型的历史样本的特征。换言之，由样本生成器410生成的针对每次历史训练任务的特征样本与该次历史训练任务中的历史样本具有相同的分布特性。为此，需要用历史训练任务中的历史样本来训练样本生成器410。
在下文中，用t来表示与输入样本180相对应的当前训练任务，并且用1～t-1来表示历史训练任务。相应地，当前训练任务的输入样本180可以表示为D t，而与历史训练任务有关的特征样本可以表示为D 1～t-1。例如，样本生成器410可以针对每次历史训练任务，利用随机数z来生成特征样本，从而获得与历史训练任务1～t-1有关的特征样本D 1～t-1。可以理解的是，特征样本D 1～t-1并非是原始的诊断报告，而是带标记数据，包括多个量化表示和对应的失效根因。特征样本D 1～t-1进而与当前任务的样本D t组成全样本集合401，即D 1～t。持续学习算法170进而基于全样本集合401来再次训练根因分析模型150，以更新模型的网络参数。
可以利用任何合适的神经网络来实现样本生成器410。在一些实施例中，可以利用条件生成式对抗网络(GAN)来实现样本生成器410。备选地，在一些实施例中，也可以基于其他神经网络来实现样本生成器410，例如可以利用变分自编码器(VAE)或条件VAE等来实现样本生成器410。
返回参考图1。当前的特征库190仅能再现1～t-1次历史训练任务的样本特征。为了应对第t+1次训练任务，可以基于当前任务t的输入样本D t来扩展特征库190。例如，可以利用全样本集合D 1～t来再次训练样本生成器410以更新样本生成器410的网络参数，使得经更新的样本生成器410能够再现1～t训练任务中的样本的特征。
现在参考图5，图5示出了根据本申请的一些实施例的扩展特征库的示意图。在图5的示例中，样本生成器410是利用条件GAN来实现的。相应地，样本生成器410的训练需要借助于样本判别器520。具体来说，针对第1至t次训练任务中的每次训练任务，样本生成器410生成该次训练任务的样本，并输入到样本判别器520。全样本集合401中的该次训练任务的样本也被输入到样本判别器520作为真值。
基于此，样本判别器520判断来自样本生成器410的样本是真样本还是假样本，并得到生成式对抗网络损失函数L GAN和交叉熵损失函数L CE。损失函数L GAN和L CE用于训练样本生成器410和样本判别器520两者，直到样本判别器520无法判断来自样本生成器410的样本是真样本还是假样本。以此方式，样本生成器410被更新，更新后的样本生成器410能够再现第1至t次训练任务中的样本的特征。也即，特征库190被扩展。
在这种实施例中，样本生成器410是基于记忆回放(Memory Replay)GAN来实现的，并且可以实现生成式特征回放。以此方式，可以在节省存储空间的情况下实现根因分析模型的持续学习，而不遗忘历史任务。
示例过程、装置和设备
图6示出了根据本申请的一些实施例的确定失效原因的过程600的流程图。过程600例如可以由图1中的根因分析引擎105来实施。为了方便描述，以下参考图1-5来描述过程600。
在框610，根因分析引擎105接收针对根因分析模型150的一组输入样本。一组输入样本例如可以是尚未用于训练根因分析模型150的样本。根因分析模型150被配置为基于分别与待分析元件101的多个部分的失效有关的多个量化表示130，来标识待分析元件101失效的根本原因。
在一些实施例中，一组输入样本中的至少一个输入样本所包括的根本原因是根据根因分析模型150确定并且经由测试验证的。在一些实施例中，待分析元件101包括待分析晶圆，并且多个部分包括待分析晶圆的多个裸片。
在框620，根因分析引擎105获取与根因分析模型150的历史训练有关的一组特征样本。一组输入样本和一组特征样本中的每个样本包括参考元件失效的根本原因、和分别与参考元件的多个参考部分的失效有关的多个参考量化表示。
在一些实施例中，根因分析引擎105根据样本生成器，生成一组特征样本。样本生成器410被配置为再现用于训练根因分析模型150的历史样本的特征。该组特征样本反映历史样本的特征。
在一些实施例中，根因分析引擎105基于一组输入样本，更新样本生成器410的网络参数。以此方式，经更新的样本生成器410能够再现一组输入样本的特征。
在框630，根因分析引擎105基于参考元件失效的根本原因和多个参考量化表示，训练根因分析模型150。换言之，根因分析模型150基于一组输入样本和一组特征样本被再次训练。
在一些实施例中，为了训练所述根因分析模型，根因分析引擎105通过将多个参考量化表示应用于根因分析模型150，确定由根因分析模型150标识的、参考元件失效的潜在原因；以及通过最小化参考元件失效的根本原因与潜在原因之间的差异，更新根因分析模型150的网络参数。
在一些实施例中，过程600还包括：根因分析引擎105接收关于待分析元件101的多个诊断报告120，每个诊断报告对应于待分析元件101的多个部分中的一个部分，并且指示该部分失效的潜在根本原因、物理缺陷和逻辑错误；根因分析引擎105从多个诊断报告120，生成分别与多个部分的失效有关的多个量化表示130；以及根因分析引擎105通过将多个量化表示130应用于根因分析模型150，标识待分析元件101失效的根本原因。
在一些实施例中，为了标识待分析元件101失效的根本原因，根因分析引擎105通过将多个量化表示130应用于根因分析模型150中的特征提取器310，从多个量化表示130生成多个局部失效特征301，每个局部失效特征与多个部分中的一个部分相对应；根因分析引擎105将多个局部失效特征301组合为针对待分析元件101的全局失效特征302；以及根因分析引擎105通过将全局失效特征302应用于根因分析模型150中的分类器320，确定不同的根本原因引起待分析元件失效的概率。
在一些实施例中，为了生成多个量化表示130，针对多个诊断报告120中的给定诊断报告，根因分析引擎105将给定诊断报告中指示的多个潜在根本原因的量化表示组合成与多个潜在根本原因相关联的第一物理缺陷的量化表示；根因分析引擎105将第一物理缺陷的量化表示和至少第二物理缺陷的量化表示组合成与第一物理缺陷和第二物理缺陷相关联的第一逻辑错误的量化表示；以及根因分析引擎105至少基于第一逻辑错误的量化表示，生成与一个部分的失效有关的量化表示。
在一些实施例中，过程600还包括：在将多个量化表示130应用于根因分析模型150之前，根因分析引擎105基于与多个诊断报告分别指示的潜在根本原因有关的面积，来调整多个量化表示130。
图7示出了根据本申请的一些实施例的用于确定失效原因的装置700的示意性框图。装置700可以用于实现图1中所示的根因分析引擎105。如图7所示，装置700包括推理模块 701，例如图1中所示的推理模块110。推理模块701被配置为基于分别与待分析元件的多个部分的失效有关的多个量化表示，根据根因分析模型，来标识所述待分析元件失效的根本原因。装置700还包括学习模块702，例如图1中所示的持续学习模块160。学习模块702包括输入样本接收单元710，其被配置为接收针对根因分析模型的一组输入样本。学习模块702还包括特征样本获取单元720，其被配置为获取与根因分析模型的历史训练有关的一组特征样本，一组输入样本和一组特征样本中的每个样本包括参考元件失效的根本原因、和分别与参考元件的多个参考部分的失效有关的多个参考量化表示。学习模块702进一步包括模型更新单元730，其被配置为基于参考元件失效的根本原因和多个参考量化表示，训练根因分析模型。
在一些实施例中，特征样本获取单元720进一步被配置为：根据样本生成器，生成一组特征样本。样本生成器被配置为再现用于训练根因分析模型的历史样本的特征，并且一组特征样本反映历史样本的特征。
在一些实施例中，学习模块702还包括：生成器更新单元，被配置为基于一组输入样本，更新样本生成器的网络参数。经更新的样本生成器能够再现一组输入样本的特征。
在一些实施例中，模型更新单元730进一步被配置为：通过将多个参考量化表示应用于根因分析模型，确定由根因分析模型标识的、参考元件失效的潜在原因；以及通过最小化参考元件失效的根本原因与潜在原因之间的差异，更新根因分析模型的网络参数。
在一些实施例中，推理模块701包括：诊断报告接收单元，其被配置为接收关于待分析元件的多个诊断报告，每个诊断报告对应于待分析元件的多个部分中的一个部分，并且指示一个部分失效的潜在根本原因、物理缺陷和逻辑错误；量化表示生成单元，其被配置为从多个诊断报告，生成分别与多个部分的失效有关的多个量化表示；以及根本原因标识单元，其被配置为通过将多个量化表示应用于根因分析模型，标识待分析元件失效的根本原因。
在一些实施例中，根本原因标识单元进一步被配置为：通过将多个量化表示应用于根因分析模型中的特征提取器，从多个量化表示生成多个局部失效特征，每个局部失效特征与多个部分中的一个部分相对应；将多个局部失效特征组合为针对待分析元件的全局失效特征；以及通过将全局失效特征应用于根因分析模型中的分类器，确定不同的根本原因引起待分析元件失效的概率。
在一些实施例中，量化表示生成单元进一步被配置为：针对多个诊断报告中的给定诊断报告，将给定诊断报告中指示的多个潜在根本原因的量化表示组合成与多个潜在根本原因相关联的第一物理缺陷的量化表示；将第一物理缺陷的量化表示和至少第二物理缺陷的量化表示组合成与第一物理缺陷和第二物理缺陷相关联的第一逻辑错误的量化表示；以及至少基于第一逻辑错误的量化表示，生成与一个部分的失效有关的量化表示。
在一些实施例中，推理模块701还包括：量化表示调整单元，其被配置为在将多个量化表示应用于根因分析模型之前，基于与多个诊断报告分别指示的潜在根本原因有关的面积，来调整多个量化表示。
在一些实施例中，一组输入样本中的至少一个输入样本所包括的根本原因是根据根因分析模型确定并且经由测试验证的。
在一些实施例中，待分析元件包括待分析晶圆，并且多个部分包括待分析晶圆的多个裸片。
图8示出了能够实施本申请的多个实施例的计算设备800的示意性框图。设备800可以 用于实现根因分析引擎105或其至少一部分，例如推理模块110或持续学习模块160。如图所示，设备800包括计算单元801，其可以根据存储在随机存取存储器(RAM)和/或只读存储器(ROM)802的计算机程序指令或者从存储单元807加载到RAM和/或ROM 802中的计算机程序指令，来执行各种适当的动作和处理。在RAM和/或ROM 802中，还可存储设备800操作所需的各种程序和数据。计算单元801和RAM和/或ROM 802通过总线803彼此相连。输入/输出(I/O)接口804也连接至总线803。
设备800中的多个部件连接至I/O接口804，包括：输入单元805，例如键盘、鼠标等；输出单元806，例如各种类型的显示器、扬声器等；存储单元807，例如磁盘、光盘等；以及通信单元808，例如网卡、调制解调器、无线通信收发机等。通信单元808允许设备800通过诸如因特网的计算机网络和/或各种电信网络与其他设备交换信息/数据。
计算单元801可以是各种具有处理和计算能力的通用和/或专用处理组件。计算单元801的一些示例包括但不限于中央处理单元(CPU)、图形处理单元(GPU)、各种专用的人工智能(AI)计算芯片、各种运行机器学习模型算法的计算单元、数字信号处理器(DSP)、以及任何适当的处理器、控制器、微控制器等。计算单元801执行上文所描述的各个方法和处理，例如过程600。例如，在一些实施例中，过程600可被实现为计算机软件程序，其被有形地包含于机器可读介质，例如存储单元807。在一些实施例中，计算机程序的部分或者全部可以经由RAM和/或ROM和/或通信单元808而被载入和/或安装到设备800上。当计算机程序加载到RAM和/或ROM并由计算单元801执行时，可以执行上文描述的过程600的一个或多个步骤。备选地，在其他实施例中，计算单元801可以通过其他任何适当的方式(例如，借助于固件)而被配置为执行过程600。
用于实施本申请的方法的程序代码可以采用一个或多个编程语言的任何组合来编写。这些程序代码可以提供给通用计算机、专用计算机或其他可编程数据处理装置的处理器或控制器，使得程序代码当由处理器或控制器执行时使流程图和/或框图中所规定的功能/操作被实施。程序代码可以完全在机器上执行、部分地在机器上执行，作为独立软件包部分地在机器上执行且部分地在远程机器上执行或完全在远程机器或服务器上执行。
在本申请的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦除可编程只读存储器(EPROM或快闪存储器)、光纤、便捷式紧凑盘只读存储器(CD-ROM)、光学储存设备、磁储存设备、或上述内容的任何合适组合。
此外，虽然采用特定次序描绘了各操作，但是这应当理解为要求这样操作以所示出的特定次序或以顺序次序执行，或者要求所有图示的操作应被执行以取得期望的结果。在一定环境下，多任务和并行处理可能是有利的。同样地，虽然在上面论述中包含了若干具体实现细节，但是这些不应当被解释为对本申请的范围的限制。在单独的实施例的上下文中描述的某些特征还可以组合地实现在单个实现中。相反地，在单个实现的上下文中描述的各种特征也可以单独地或以任何合适的子组合的方式实现在多个实现中。
尽管已经采用特定于结构特征和/或方法逻辑动作的语言描述了本主题，但是应当理解所 附权利要求书中所限定的主题未必局限于上面描述的特定特征或动作。相反，上面所描述的特定特征和动作仅仅是实现权利要求书的示例形式。

Claims (23)

  1. 一种确定失效原因的装置，其特征在于，包括：
    推理模块，被配置为基于分别与待分析元件的多个部分的失效有关的多个量化表示，根据根因分析模型，来标识所述待分析元件失效的根本原因；以及
    学习模块，与所述推理模块耦合，并且包括：
    输入样本接收单元，被配置为接收针对所述根因分析模型的一组输入样本；
    特征样本获取单元，被配置为获取与所述根因分析模型的历史训练有关的一组特征样本，所述一组输入样本和所述一组特征样本中的每个样本包括参考元件失效的根本原因、和分别与所述参考元件的多个参考部分的失效有关的多个参考量化表示；以及
    模型更新单元，被配置为基于所述参考元件失效的根本原因和所述多个参考量化表示，训练所述根因分析模型。
  2. 根据权利要求1所述的装置，其特征在于，所述特征样本获取单元进一步被配置为：
    根据样本生成器，生成所述一组特征样本，所述样本生成器被配置为再现用于训练所述根因分析模型的历史样本的特征，并且所述一组特征样本反映所述历史样本的特征。
  3. 根据权利要求2所述的装置，其特征在于，所述学习模块还包括：
    生成器更新单元，被配置为基于所述一组输入样本，更新所述样本生成器的网络参数。
  4. 根据权利要求1至3任一所述的装置，其特征在于，所述模型更新单元进一步被配置为：
    通过将所述多个参考量化表示应用于所述根因分析模型，确定由所述根因分析模型标识的、所述参考元件失效的潜在原因；以及
    通过最小化所述参考元件失效的根本原因与所述潜在原因之间的差异，更新所述根因分析模型的网络参数。
  5. 根据权利要求1至4任一所述的装置，其特征在于，所述推理模块包括：
    诊断报告接收单元，被配置为接收关于所述待分析元件的多个诊断报告，每个诊断报告对应于所述待分析元件的所述多个部分中的一个部分，并且指示所述一个部分失效的潜在根本原因、物理缺陷和逻辑错误；
    量化表示生成单元，被配置为从所述多个诊断报告，生成分别与所述多个部分的失效有关的所述多个量化表示；以及
    根本原因标识单元，被配置为通过将所述多个量化表示应用于所述根因分析模型，标识所述待分析元件失效的根本原因。
  6. 根据权利要求5所述的装置，其特征在于，所述根本原因标识单元进一步被配置为：
    通过将所述多个量化表示应用于所述根因分析模型中的特征提取器，从所述多个量化表示生成多个局部失效特征，每个局部失效特征与所述多个部分中的一个部分相对应；
    将所述多个局部失效特征组合为针对所述待分析元件的全局失效特征；以及
    通过将所述全局失效特征应用于所述根因分析模型中的分类器，确定不同的根本原因引起所述待分析元件失效的概率。
  7. 根据权利要求5所述的装置，其特征在于，所述量化表示生成单元进一步被配置为：
    针对所述多个诊断报告中的给定诊断报告，
    将所述给定诊断报告中指示的多个潜在根本原因的量化表示组合成与所述多个潜 在根本原因相关联的第一物理缺陷的量化表示；
    将所述第一物理缺陷的量化表示和至少第二物理缺陷的量化表示组合成与所述第一物理缺陷和所述第二物理缺陷相关联的第一逻辑错误的量化表示；以及
    至少基于所述第一逻辑错误的量化表示，生成与所述一个部分的失效有关的量化表示。
  8. 根据权利要求5所述的装置，其特征在于，所述推理模块还包括：
    量化表示调整单元，被配置为在将所述多个量化表示应用于所述根因分析模型之前，基于与所述多个诊断报告分别指示的潜在根本原因有关的面积，来调整所述多个量化表示。
  9. 根据权利要求1至8任一所述的装置，其特征在于，所述一组输入样本中的至少一个输入样本所包括的根本原因是根据所述根因分析模型确定并且经由测试验证的。
  10. 根据权利要求1至9任一所述的装置，其特征在于，所述待分析元件包括待分析晶圆，并且所述多个部分包括所述待分析晶圆的多个裸片。
  11. 一种确定失效原因的方法，其特征在于，包括：
    接收针对根因分析模型的一组输入样本，所述根因分析模型被配置为基于分别与待分析元件的多个部分的失效有关的多个量化表示，来标识所述待分析元件失效的根本原因；
    获取与所述根因分析模型的历史训练有关的一组特征样本，所述一组输入样本和所述一组特征样本中的每个样本包括参考元件失效的根本原因、和分别与所述参考元件的多个参考部分的失效有关的多个参考量化表示；以及
    基于所述参考元件失效的根本原因和所述多个参考量化表示，训练所述根因分析模型。
  12. 根据权利要求11所述的方法，其特征在于，获取所述一组特征样本包括：
    根据样本生成器，生成所述一组特征样本，所述样本生成器被配置为再现用于训练所述根因分析模型的历史样本的特征，并且所述一组特征样本反映所述历史样本的特征。
  13. 根据权利要求12所述的方法，其特征在于，还包括：
    基于所述一组输入样本，更新所述样本生成器的网络参数。
  14. 根据权利要求11至13任一项所述的方法，其特征在于，训练所述根因分析模型包括：
    通过将所述多个参考量化表示应用于所述根因分析模型，确定由所述根因分析模型标识的、所述参考元件失效的潜在原因；以及
    通过最小化所述参考元件失效的根本原因与所述潜在原因之间的差异，更新所述根因分析模型的网络参数。
  15. 根据权利要求11至14任一所述的方法，其特征在于，还包括：
    接收关于所述待分析元件的多个诊断报告，每个诊断报告对应于所述待分析元件的所述多个部分中的一个部分，并且指示所述一个部分失效的潜在根本原因、物理缺陷和逻辑错误；
    从所述多个诊断报告，生成分别与所述多个部分的失效有关的所述多个量化表示；以及
    通过将所述多个量化表示应用于所述根因分析模型，标识所述待分析元件失效的根本原因。
  16. 根据权利要求15所述的方法，其特征在于，标识所述待分析元件失效的根本原因包括：
    通过将所述多个量化表示应用于所述根因分析模型中的特征提取器，从所述多个量化表示生成多个局部失效特征，每个局部失效特征与所述多个部分中的一个部分相对应；
    将所述多个局部失效特征组合为针对所述待分析元件的全局失效特征；以及
    通过将所述全局失效特征应用于所述根因分析模型中的分类器，确定不同的根本原因引起所述待分析元件失效的概率。
  17. 根据权利要求15所述的方法，其特征在于，生成所述多个量化表示包括：
    针对所述多个诊断报告中的给定诊断报告，
    将所述给定诊断报告中指示的多个潜在根本原因的量化表示组合成与所述多个潜在根本原因相关联的第一物理缺陷的量化表示；
    将所述第一物理缺陷的量化表示和至少第二物理缺陷的量化表示组合成与所述第一物理缺陷和所述第二物理缺陷相关联的第一逻辑错误的量化表示；以及
    至少基于所述第一逻辑错误的量化表示，生成与所述一个部分的失效有关的量化表示。
  18. 根据权利要求15所述的方法，其特征在于，还包括：
    在将所述多个量化表示应用于所述根因分析模型之前，基于与所述多个诊断报告分别指示的潜在根本原因有关的面积，来调整所述多个量化表示。
  19. 根据权利要求11至18任一所述的方法，其特征在于，所述一组输入样本中的至少一个输入样本所包括的根本原因是根据所述根因分析模型确定并且经由测试验证的。
  20. 根据权利要求11至19任一所述的方法，其特征在于，所述待分析元件包括待分析晶圆，并且所述多个部分包括所述待分析晶圆的多个裸片。
  21. 一种电子设备，其特征在于，包括：
    至少一个处理器；
    至少一个存储器，所述至少一个存储器被耦合到所述至少一个处理器并且存储用于由所述至少一个处理器执行的指令，所述指令当由所述至少一个处理器执行时，使所述电子设备执行根据权利要求11-20中任一项所述的方法。
  22. 一种计算机可读存储介质，其特征在于，其上存储有计算机程序，所述程序被处理器执行时实现根据权利要求11-20中任一项所述的方法。
  23. 一种计算机程序产品，其特征在于，包括计算机可执行指令，其中所述计算机可执行指令在被处理器执行时实现根据权利要求11-20中任一项所述的方法。
PCT/CN2021/101866 2021-06-23 2021-06-23 一种确定失效原因的方法及装置 WO2022266890A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202180099751.1A CN117561502A (zh) 2021-06-23 2021-06-23 一种确定失效原因的方法及装置
PCT/CN2021/101866 WO2022266890A1 (zh) 2021-06-23 2021-06-23 一种确定失效原因的方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/101866 WO2022266890A1 (zh) 2021-06-23 2021-06-23 一种确定失效原因的方法及装置

Publications (1)

Publication Number Publication Date
WO2022266890A1 true WO2022266890A1 (zh) 2022-12-29

Family

ID=84545038

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/101866 WO2022266890A1 (zh) 2021-06-23 2021-06-23 一种确定失效原因的方法及装置

Country Status (2)

Country Link
CN (1) CN117561502A (zh)
WO (1) WO2022266890A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115798559A (zh) * 2023-02-10 2023-03-14 长鑫存储技术有限公司 失效单元预测方法、装置、设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109684181A (zh) * 2018-11-20 2019-04-26 华为技术有限公司 告警根因分析方法、装置、设备及存储介质
CN111177416A (zh) * 2020-04-13 2020-05-19 傲林科技有限公司 事件根因分析模型构建方法、事件根因分析方法及装置
CN111178537A (zh) * 2019-12-09 2020-05-19 华为技术有限公司 一种特征提取模型训练方法及设备
US10873456B1 (en) * 2019-05-07 2020-12-22 LedgerDomain, LLC Neural network classifiers for block chain data structures
CN112395170A (zh) * 2020-12-07 2021-02-23 平安普惠企业管理有限公司 智能故障分析方法、装置、设备及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109684181A (zh) * 2018-11-20 2019-04-26 华为技术有限公司 告警根因分析方法、装置、设备及存储介质
US10873456B1 (en) * 2019-05-07 2020-12-22 LedgerDomain, LLC Neural network classifiers for block chain data structures
CN111178537A (zh) * 2019-12-09 2020-05-19 华为技术有限公司 一种特征提取模型训练方法及设备
CN111177416A (zh) * 2020-04-13 2020-05-19 傲林科技有限公司 事件根因分析模型构建方法、事件根因分析方法及装置
CN112395170A (zh) * 2020-12-07 2021-02-23 平安普惠企业管理有限公司 智能故障分析方法、装置、设备及存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115798559A (zh) * 2023-02-10 2023-03-14 长鑫存储技术有限公司 失效单元预测方法、装置、设备及存储介质
CN115798559B (zh) * 2023-02-10 2023-07-04 长鑫存储技术有限公司 失效单元预测方法、装置、设备及存储介质

Also Published As

Publication number Publication date
CN117561502A (zh) 2024-02-13

Similar Documents

Publication Publication Date Title
Yang et al. TLEL: A two-layer ensemble learning approach for just-in-time defect prediction
Ioannides et al. Coverage-directed test generation automated by machine learning--a review
US11604917B2 (en) Static voltage drop (SIR) violation prediction systems and methods
CN107346286B (zh) 一种基于核主成分分析和极限学习机的软件缺陷预测方法
US11093683B2 (en) Test pattern generation systems and methods
CN113657661A (zh) 一种企业碳排放预测方法、装置、计算机设备和存储介质
US9703658B2 (en) Identifying failure mechanisms based on a population of scan diagnostic reports
TW201837702A (zh) 用於評估稀有失效事件的電路良率分析方法及系統
Zhou et al. Imbalanced data processing model for software defect prediction
Zhang et al. A fault diagnosis method of power transformer based on cost sensitive one-dimensional convolution neural network
WO2022266890A1 (zh) 一种确定失效原因的方法及装置
Zhao et al. Towards faithful and consistent explanations for graph neural networks
Chaudhuri et al. Fault-criticality assessment for AI accelerators using graph convolutional networks
Zhang et al. A software defect prediction approach based on bigan anomaly detection
CN114781532A (zh) 机器学习模型的评估方法、装置、计算机设备及介质
Huang et al. Towards smarter diagnosis: A learning-based diagnostic outcome previewer
CN114169460A (zh) 样本筛选方法、装置、计算机设备和存储介质
US20240037313A1 (en) Statistical graph circuit component probability model for an integrated circuit design
CN115098389B (zh) 一种基于依赖模型的rest接口测试用例生成方法
US11467851B1 (en) Machine learning (ML)-based static verification for derived hardware-design elements
Fredriksson et al. Machine learning algorithms for labeling: Where and how they are used?
Wu et al. Circuit-GNN: A Graph Neural Network for Transistor-level Modeling of Analog Circuit Hierarchies
Jabbari et al. Obtaining accurate probabilistic causal inference by post-processing calibration
Zhu et al. Cross-Project Defect Prediction Method based on Feature Distribution Alignment and Neighborhood Instance Selection
US12019971B2 (en) Static voltage drop (SIR) violation prediction systems and methods

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21946390

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 202180099751.1

Country of ref document: CN

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21946390

Country of ref document: EP

Kind code of ref document: A1