JP2022096340A - Processing system, learning inference processing device, control method of learning inference processing device, and control program of learning inference processing device - Google Patents

Abstract

To provide a processing system, a learning inference processing device, a control method, and a control program for realizing identification of an initial value of a basic operation parameter that shortens a setup time.SOLUTION: A processing system (1) according to a first embodiment of the present disclosure includes: a processing device (10) that performs processing; and a learning inference processing device (40) including a first inference unit (41) that inputs control information for controlling the processing device (10) output by a second inference unit (42) when the second inference unit (42) is learned, one of recipe information regarding the specifications of the processing executed by the processing device (10) and device information regarding the activation operation of the processing device (10), and a control information measurement result which is a result obtained by measuring the control information at the end of the activation of the processing device (10), and infers and outputs time information, and a second inference unit (42) that inputs one of the recipe information and the device information and the control information measurement result and outputs the control information to the processing device (10).SELECTED DRAWING: Figure 2

Description

The present disclosure relates to a learning inference processing device for acquiring control information in a processing device that performs various processes, a control method of the learning inference processing device, a control program of the learning inference processing device, and a processing system including these.

In a processing apparatus that executes various processes, it is important to specify appropriate control information in order to appropriately execute the processes. As such a processing device, for example, an ion implantation processing device that executes an ion implantation process into a semiconductor substrate or the like is known (see, for example, Patent Document 1 below).

The ion implantation processing apparatus as shown in Patent Document 1 includes various modules for generating an ion beam such as an ion source and a mass spectrometer (separator). Then, these modules are configured to operate based on various preset basic operation parameters. When operating this ion implanter, it is necessary to first perform a start-up operation. Specifically, when a desired recipe (a set of processing conditions such as ion species, ion beam energy, dose amount, etc.) is read in the control device that controls each module, the control device first tells each module the recipe. Specify the initial value of the basic operation parameter corresponding to. Next, this controller automatically adjusts the initial values of the basic operation parameters so that the ion beam according to the recipe is generated while detecting various information about the generated ion beam, and obtains the desired adjustment value. Identify. By specifying this adjustment value, the start-up operation of the ion implantation processing device is completed. When the start-up operation is completed, the ion implantation processing apparatus is in a processable state, and the ion implantation processing is executed based on the operator's instruction or the like. Specifically, the basic operation parameters referred to here are the voltage value and current value of the power supply connected to each module of the ion injection processing device, and the type of gas supplied to the plasma chamber and the processing chamber in the ion source. The flow rate, the magnetic flux density of the magnetic field generated in the plasma chamber, the distance between the electrodes of the ion beam extraction electrode, the acceleration / deceleration of the ion beam, etc., which refer to the main control parameters of the ion injection processing device. be.

Japanese Unexamined Patent Publication No. 2008-066664

By the way, regarding the automatic adjustment of the initial value of the basic operation parameter by the above-mentioned control device, usually, the device is actually operated by software created based on the knowledge based on the experience of a skilled engineer or by the initial value of the basic operation parameter. Adjustments are made taking into account the results of the operation. Therefore, if the difference between the initial value of the basic operation parameter and the adjustment value obtained after the automatic adjustment is small, the time required for the automatic adjustment is shortened, and the setup time from the loading of the recipe to the completion of the start-up of the device is shortened. can do. Here, the initial values of the above-mentioned basic operation parameters are, in many cases, predetermined values stored in advance corresponding to the recipe or the information of the recipe and the ion implantation processing apparatus. Therefore, the initial value of the specified basic operation parameter is often not close to the adjusted value, and there is room for improvement in specifying the initial value.

In view of the above points, the present disclosure discloses a processing system, a learning inference processing device, a control method of the learning inference processing device, and a learning inference processing device in order to specify initial values of basic operation parameters that shorten the setup time. The purpose is to provide a control program for.

In order to achieve the above object, the processing system according to the first aspect of the present disclosure includes a processing device that performs processing and the processing device that is output by the second inference unit when learning the second inference unit. Either the control information for controlling the processing device, the recipe information regarding the specifications of the processing executed by the processing device, or the device information regarding the start-up operation of the processing device, and the control information were measured at the start and end of the processing device. The control information measurement result, which is the result, is input, and the first inference unit that infers and outputs the time information, the recipe information or the device information, and the control information measurement result are input. , A learning inference processing apparatus having the second inference unit that outputs the control information to the processing apparatus.

Such a processing system simply infers the control information output to the processing apparatus by using the second inference unit learned by using the time information output by the first inference unit, thereby simply performing the recipe information and the conventional one. / Or, as compared with the case of obtaining the control information as the specified value corresponding to the device information, the control information in consideration of the setup time in the processing device can be obtained. Therefore, it becomes possible to provide a processing system in which the setup time in the processing apparatus is shortened.

The processing system according to the second aspect of the present disclosure is the processing system according to the first aspect of the present disclosure. In the processing system according to the first aspect of the present disclosure, the second inference unit included in the learning inference processing device is output by the first inference unit. Learning is repeated by updating the weight value of the neural network included in the second inference unit until the time information becomes less than or equal to the predetermined reference value, and the learned AI when the time information becomes less than or equal to the reference value. Includes modules.

In such a processing system, the learning of the second inference unit is repeated until the time information becomes equal to or less than the reference value, and the learned AI module obtained as a result of the learning of the second inference unit always has the time information. It becomes possible to infer control information that is below the reference value. This makes it possible to provide a processing system in which the setup is completed in an arbitrary setup time by adjusting the reference value to a desired setup time.

The processing system according to the third aspect of the present disclosure is the processing system according to the first or second aspect of the present disclosure, wherein the first inference unit is further a recipe relating to a specification of processing executed by the processing apparatus. Enter either information or device information related to the startup operation of the processing device, whichever is the other.

In such a processing system, the inference accuracy of the AI module is improved by increasing the number of explanatory variables input to the first inference unit.

The processing system according to the fourth aspect of the present disclosure is at least one of recipe information or device information input to the second inference unit in the processing system according to the first to third aspects of the present disclosure. When the recipe information or device information that actually controls the processing device is the same, the actual control information that is the control information used when the processing device actually operates and the second inference unit The squared error with the output control information is further input to the second inference unit, and the second inference unit learns so that the value of the difference information defined by the squared error and the time information becomes small. repeat.

In such a processing system, by learning the second inference unit using the actual control information, the control information output by the second inference unit becomes a value close to the actual control information in consideration of the actual control information. .. As a result, it is possible to prevent the control information output by the second inference unit from becoming an impractical value although the time information is calculated to be less than or equal to the reference value, and an error occurs when setting up the processing device. It can be suppressed.

The processing system according to the fifth aspect of the present disclosure is the processing system according to the fourth aspect of the present disclosure, wherein the difference information uses the time information and the squared error, and the difference information = α1 × time. It is defined by the formula of information + α2 × squared error (α1 and α2 are arbitrary coefficients).

Such a processing system can shorten the setup time in the second inference unit by adopting two pieces of information, time information and squared error of actual control information and control information, as difference information used for learning. Moreover, control information as a practical value can be inferred.

In the processing system according to the sixth aspect of the present disclosure, the relationship between the α1 and the α2 satisfies 0.01 ≦ α1 / α2 ≦ 0.1 in the processing system according to the fifth aspect of the present disclosure. ..

In such a processing system, since the difference information is a value that is relatively greatly affected by the squared error, it is possible to almost eliminate the situation where an impractical value is inferred in the second inference unit.

The learning inference processing device according to the seventh aspect of the present disclosure controls the processing device, and controls the processing device output by the second inference unit when the second inference unit is trained. This is the result of measuring one of the control information for the purpose, the recipe information regarding the specifications of the processing executed by the processing device, or the device information regarding the start-up operation of the processing device, and the control information at the start and end of the processing device. A first inference unit that inputs a control information measurement result and infers and outputs time information, one of the recipe information or the device information, and the control information measurement result are input and the first. It includes the second inference unit that outputs the control information to the processing device based on the time information output by the inference unit of the above.

Such a learning inference processing device infers control information to be output to the processing device using the second inference unit learned using the time information output by the first inference unit, and outputs this to the processing device. By doing so, the setup time in the processing device from which the control information is output can be shortened.

The control method of the learning inference processing device according to the eighth aspect of the present disclosure is the learning in the control method of the learning inference processing device including the first inference unit and the second inference unit while controlling the processing device. The control information for controlling the processing device output by the second inference unit, the recipe information regarding the specifications of the processing executed by the processing device, or the device information regarding the start-up operation of the processing device. By learning using either one and the time information inferred and output by the first inference unit, which inputs the control information measurement result which is the result of measuring the control information at the start and end of the processing apparatus. The second inference unit is generated, and the second inference unit inputs either one of the recipe information or the device information and the control information measurement result, and transfers the control information to the processing device. It is to output.

In such a control method of the learning inference processing device, the control information output to the processing device is inferred by using the second inference unit learned by using the time information output by the first inference unit, and this is processed. By outputting to the device, the setup time in the processing device from which the control information is output can be shortened.

The control program of the learning inference processing device according to the ninth aspect of the present disclosure controls the processing device, and is the learning in the control program of the learning inference processing device including the first inference unit and the second inference unit. The control information for controlling the processing device output by the second inference unit to the inference processing device, the recipe information regarding the specifications of the processing executed by the processing device, or the device information regarding the start-up operation of the processing device. By learning using either one and the time information inferred and output by the first inference unit, which inputs the control information measurement result which is the result of measuring the control information at the start and end of the processing apparatus. The second inference unit is generated, the second inference unit is made to input either one of the recipe information or the device information and the control information measurement result, and the control information is transmitted to the processing device. It is to be output.

The control program of such a learning inference processing device infers the control information output to the processing device using the second inference unit learned by using the time information output by the first inference unit, and processes it. By outputting to the device, the setup time in the processing device from which the control information is output can be shortened.

By providing the above-mentioned configuration, it is possible to specify the initial value of the basic operation parameter so that the setup time becomes shorter than before. It becomes possible to provide a control program for the device.

It is a schematic explanatory drawing which shows an example of the ion implantation processing system which concerns on 1st Embodiment of this disclosure. It is a functional block diagram which shows an example of the information processing apparatus which includes the learning inference processing apparatus which concerns on 1st Embodiment of this disclosure. It is a schematic explanatory diagram which shows an example of the learning process in the learning inference processing apparatus which concerns on 1st Embodiment of this disclosure. It is a flowchart which shows an example of the control method of the learning inference processing apparatus which concerns on 1st Embodiment of this disclosure. It is a flowchart which shows an example of the setup process of the ion implantation processing apparatus in the ion implantation processing system which concerns on 1st Embodiment of this disclosure. It is a schematic explanatory diagram which shows an example of the learning process in the learning inference processing apparatus which concerns on the 2nd Embodiment of this disclosure. It is a flowchart which shows an example of the control method of the learning inference processing apparatus which concerns on 2nd Embodiment of this disclosure. It is a schematic explanatory diagram which shows an example of the learning process in the learning inference processing apparatus which concerns on 3rd Embodiment of this disclosure.

Hereinafter, embodiments for carrying out the present disclosure will be described with reference to the drawings. In the following, the scope necessary for the explanation to achieve the object of the present disclosure will be schematically shown, and the scope necessary for the explanation of the relevant part of the present disclosure will be mainly explained. It shall be based on known technology.

<First Embodiment>
FIG. 1 is a schematic explanatory view showing an example of an ion implantation processing system 1 as a processing system according to the first embodiment of the present disclosure. As shown in FIG. 1, the ion implantation processing system 1 according to the present embodiment mainly includes an ion implantation processing apparatus 10 and an information processing apparatus 20. In the present embodiment, the ion implantation processing system 1 is exemplified and described as the processing system, but it can also be applied to other processing systems that require prior setup when executing the processing.

The ion implantation processing apparatus 10 is an apparatus that implants ions into the substrate W as a processing object by irradiating the substrate W made of a semiconductor wafer or the like with a desired ion beam IB. The ion implantation processing apparatus 10 mainly includes an ion source 11, a mass spectrometer (separator) 12, an analysis (separation) slit 13, an acceleration tube 14, an energy separator 15, a scanner 16, and the like. Those including the beam parallelizer 17 and the substrate driving device 18 can be adopted.

The ion source 11 generates an ion beam that is the source of the ribbon-shaped ion beam IB that irradiates the substrate W. Further, the mass spectrometer 12 analyzes and separates ions of a predetermined type from the ion beam IB generated from the ion source 11 based on their mass numbers and valences. Further, the analysis slit 13 selects and derives a desired ion from a plurality of types of ions separated by the mass spectrometer 12. Furthermore, the accelerating tube 14 accelerates (or decelerates) the ion beam IB derived from the analysis slit 13 so as to have a predetermined energy.

The energy separator 15 separates and sorts and derives ions of desired energy from the ion beam IB accelerated by the acceleration tube 14. Further, the scanner 16 reciprocates the ion beam IB derived from the energy separator 15 in one direction (X direction in FIG. 1) electrically or magnetically. Further, the beam parallelizer 17 electrically or magnetically bends back the ion beam IB derived from the scanner 16 and performs parallel scanning of the ion beam IB in cooperation with the scanner 16 to the substrate W. The purpose is to irradiate the ion beam IB in a ribbon shape. Furthermore, the substrate driving device 18 is a device for sequentially holding the substrate W as a processing target and positioning and holding it at a predetermined position and angle. A well-known structure can be adopted for the holding structure of the substrate W and the operating structure for operating the substrate W in the substrate driving device 18, and detailed description thereof will be omitted here. Further, it is possible that the substrate W and the substrate driving device 18 are housed in a processing chamber (not shown) and the atmospheric gas or the like is controlled.

The information processing device 20 is a device that performs various controls of the ion implantation processing device 10, and can be realized by, for example, a computer. The information processing device 20 may be capable of realizing the function of the learning inference processing device 40 according to the present embodiment. The information processing device 20 includes, for example, a processor 21 represented by a CPU (Central Processing Unit) as a central processing unit that executes various operations in a computer, and data necessary for the operations. A communication interface for wirelessly (or wired) communication between the volatile and / or non-volatile memory 22 and various components of the ion injection processor 10 (ion source 11, mass analyzer 12, board drive device 18, etc.). Those containing at least 23 can be adopted. The information processing device 20 according to the present embodiment can be provided with these configurations to enable various functions and operations of the control device 30 and the learning inference processing device 40, which will be described later.

FIG. 2 is a functional block diagram showing an example of an information processing device 20 including a learning inference processing device 40 according to the first embodiment of the present disclosure. As shown in FIG. 2, the information processing apparatus 20 according to the present embodiment outputs a control device 30 that performs various controls of the ion injection processing device 10 and predetermined control information used in the control device 30. It is possible to realize a learning inference processing device 40 that learns an AI (Artificial Information) module as an artificial intelligence module and provides an inference result by the learned AI module to a control device. In the present embodiment, an embodiment in which the learning inference processing device 40 is included in the information processing device 20 will be exemplified, but the present disclosure is not limited to this, and for example, the learning inference processing device 40 may be referred to as a control device 30. Data communication is possible with the information processing device 20 including the information processing device 20, and it can also be realized by another computer or the like different from the information processing device 20.

As shown in FIG. 2, the control device 30 includes a parameter acquisition unit 31, a basic operation parameter prediction unit 32, a basic operation parameter automatic adjustment unit 33, a data storage unit 34, an input / output interface 35, and a configuration thereof. It may include an internal bus 36 connecting the elements. Among these configurations, the parameter acquisition unit 31 acquires various parameters (information) used in the basic operation parameter prediction unit 32. Further, the basic operation parameter prediction unit 32 is based on the various parameters acquired by the parameter acquisition unit 31, and is preferable as the initial value of the basic operation parameter. It is also called "operation parameter"). The basic operation parameter prediction unit 32 functions as a predictor that outputs the inference result of the learned AI module learned by the learning inference processing device 40 described later, and the learned AI module used in the predictor. The details of this will be described later. Further, the types and numbers of parameters adopted as the basic operation parameters in the present embodiment are appropriately adjusted according to the configuration of the ion implantation processing apparatus 10 and the like.

The basic operation parameter automatic adjustment unit 33 inputs the predicted basic operation parameter specified by the basic operation parameter prediction unit 32 as the initial value of the basic operation parameter, so that a desired ion beam can be generated in the corresponding ion injection processing device 10. The initial value of the input basic operation parameter is automatically adjusted so that it is generated, and the adjusted value is specified. The basic operation parameter automatic adjustment unit 33 is communicably connected to each of the above-mentioned components of the ion implantation processing device 10, and the actual ion implantation processing operation of the ion implantation processing device 10 is based on the above adjustment value. Will be executed. Further, the automatic adjustment by the basic operation parameter automatic adjustment unit 33 can be performed by software for specifying the adjustment value of the basic operation parameter set for each ion implantation processing device 10. The software may be stored in the basic operation parameter automatic adjustment unit 33, and can be used as knowledge based on the experience of a skilled engineer or accumulated information of measured values when the ion implantation processing device 10 is actually operated. It may be generated based on. Further, the data storage unit 34 is a database capable of storing various information used in the control device 30, and can store at least the device parameters described later and the previous startup result. Further, the input / output interface 35 is an interface capable of inputting various information including recipe parameters described later and an operation instruction to the ion injection processing device 10, and is a well-known interface including a keyboard, a pointing device, a touch panel, a monitor, and the like. It is composed of a user interface (UI: User Interface).

In the present embodiment, the parameter acquisition unit 31 acquires three parameters. Specifically, the recipe parameters (recipe information) related to the specifications of the ion implantation process executed by the ion implantation process device 10, the device parameters (device information) related to the start-up operation of the ion implantation process device 10, and the ion implantation process device 10 The previous start-up result (control information measurement result) as a result of measuring the basic operation parameters in the ion implantation processing apparatus 10 at the end of the start-up of the Of the above three parameters, the recipe parameter can be obtained from the input / output interface 35, and the device parameter and the previous startup result can be obtained from the data storage unit 34. Here, it is particularly noteworthy that the parameters to be acquired include the previous start-up result. This is based on knowing that the start-up operation in the ion implantation processing apparatus 10 is affected by the operation content and the end state of the immediately preceding start-up.

Regarding the above three parameters, as the recipe parameters, numerical values for obtaining a desired ion implantation result, specifically, the ion type of the ion beam IB to be irradiated, the number of charges, the energy set value, the current amount set value, and the like can be mentioned. .. Similarly, as device parameters, information about the target ion implantation processing device 10, specifically, processing mode (for example, cold start, with ion type switching), presence / absence of plasma flat gun (PFG: Plasma Flood Gun), argon. Whether or not the inside of the chamber is cleaned with (Ar) gas or boron trifluoride (BF3) gas can be mentioned. In addition, as the result of the previous start-up, various measured values in the ion injection processing apparatus 10, specifically, the current value of the ion beam IB, the mass number, the valence, the energy, the gas flow rate used in the ion source 11, the arc electrode and the filament. Measured values such as voltage and current values applied to, electrode positions, etc. can be mentioned.

The learning inference processing device 40 is a device capable of obtaining a learned AI module that can be used in the basic operation parameter prediction unit 32 by executing learning of the AI module. As shown in FIG. 2, the learning inference processing device 40 can mainly include a first inference unit 41, a second inference unit 42, and a learning data set storage unit 43. In the present embodiment, in order to simplify the explanation, an embodiment in which the learning data set storage unit 43 is provided in the learning inference processing device 40 is described, but the present disclosure is limited to this. However, for example, the learning data set storage unit 43 can be realized by an arbitrary external memory other than the information processing device 20, a cloud server, or the like.

The learning data set storage unit 43 is a plurality of learning data sets (specifically, a first learning data set TD1, a third learning data set TD3, or a third learning data set) input to the second inference unit 42. The learning data set TD4) of 4 is stored. The first learning data set TD1 used in the present embodiment can include the recipe parameter, the device parameter, and the previous start-up result at any time in the past.

FIG. 3 is a schematic explanatory diagram showing an example of a learning process in the learning inference processing device 40 according to the first embodiment of the present disclosure. It is particularly noteworthy that the learning inference processing device 40 shown here employs an optimization algorithm in machine learning. As the optimization algorithm, in detail, it is particularly preferable to adopt the Adam algorithm (Adam Optimizer) among various gradient descent methods. Hereinafter, the learning method in the learning inference processing device 40 will be briefly described together with the functions of each part.

The second inference unit 42 can be realized by a trained AI module including a neural network model. As illustrated in FIG. 3, the neural network in the neural network model included in the second inference unit 42 includes an arbitrary number of neurons constituting an input layer and an arbitrary number of neurons constituting an intermediate layer (also referred to as a hidden layer). It has a number of neurons and any number of neurons that make up the output layer, and multiple nodes are stretched between the neurons in one layer and the neurons in the other layers adjacent to the one layer. It may have a configuration. Further, each node is associated with a weight, and the neurons constituting the intermediate layer are associated with a bias. In the following, these weights and biases are collectively referred to as a weight value. Adjusting this weight value by backpropagation (error backpropagation method) in the learning process described later is the learning of the neural network model in the present embodiment. Then, the neural network model whose weight value is adjusted by learning is stored in the second inference unit 42 as a learned AI module. In addition, although FIG. 3 exemplifies the one provided with only one intermediate layer for the sake of simplification of the figure, the number of intermediate layers may be two or more.

The data included in the first learning data set TD1 in the learning data set storage unit 43 is associated with the input layer of the neural network model of the second inference unit 42 as input data (explanatory variable). .. Specifically, in the present embodiment, the recipe parameter, the device parameter, and the previous start-up result are input to the second inference unit 42 as input data. Here, although not shown, preprocessing for adjusting the number of values specified by the above-mentioned three parameters is mainly performed in order to adjust the degree of influence on the output result of the above-mentioned three parameters. It is preferable to carry out. When the input data is input to the input layer of the neural network model of the second inference unit 42, the predicted basic operation parameter is output to the output layer as an output result (objective variable). The value of the predicted basic operation parameter output here corresponds to the value that can be used as the initial value of the basic operation parameter as it is (for example, the voltage value or the current value when generating plasma in the ion source 11). It may be, or an arbitrary feature quantity that can obtain a value that can be used as an initial value of the basic operation parameter is shown by executing an arbitrary post-processing (not shown) on the value output to the output layer. It may be a value.

The first inference unit 41 can be realized by an AI module including a neural network model like the second inference unit 42. However, the neural network model included in the first inference unit 41 is a trained AI module after learning has been performed in advance, and the weight values in the trained AI module are shown in FIG. 4 below. It should be noted in particular that the learning process according to the present embodiment described by using is not adjusted. The second learning data set TD2 is input as input data to the input layer of the neural network model included in the first inference unit 41. The second learning data set TD2 has the predicted basic operation parameter output by the second inference unit 42 and the input data input to the second inference unit 42 when the predicted basic operation parameter is output. That is, the recipe parameters included in the first learning data set TD1, the device parameters, and the previous start-up result are included. When the input data composed of the second learning data set TD2 is input to the input layer of the neural network model in the first inference unit 41, the output layer of the neural network model in the first inference unit 41 is input. Is output by inferring the predicted setup time (time information) as an output result. This predicted setup time predicts the setup time required to complete the start-up of the ion injection processing device 10 when the predicted basic operation parameter is input to the basic operation parameter automatic adjustment unit 33 as the initial value of the basic operation parameter. It is the value that was set.

The trained AI module included in the neural network model included in the first inference unit 41 described above can be acquired, for example, by performing well-known supervised learning. Specifically, first, it corresponds to the recipe parameter at an arbitrary time point, the device parameter, the previous start-up result, and the initial value of the basic operation parameter specified based on these three parameters (for example, by manual operation by the engineer). A plurality of learning data sets composed of a predicted basic operation parameter and a setup time as an actually measured value when the initial value of the basic operation parameter is input to the basic operation parameter automatic adjustment unit 33 are prepared. Next, when the input data is input to the input layer, the pre-learning AI module for outputting the value corresponding to the setup time to the output layer is prepared. Then, the prepared pre-learning neural network with the recipe parameters, device parameters, previous startup results, and predicted basic operation parameters as input data and the setup time as teacher data in the prepared multiple learning data sets. By sequentially inputting input data to the model and training the output data using the teacher data, a desired trained AI module can be acquired. In the present embodiment, the first inference unit 41 includes a neural network model, but the first inference unit 41 is not limited to such an embodiment. That is, for example, a well-known simulator can be used as an alternative as long as the predicted setup time can be output based on the second learning data set TD2.

The time comparison unit 50 is for comparing whether or not the predicted setup time output by the first inference unit 41 is equal to or less than a preset reference value. Although this reference value can be set as appropriate, the average setup time when the ion implantation processing apparatus 10 is started up is adopted as the reference value in the present embodiment. In the learning inference processing device 40 according to the present embodiment, when the time comparison unit 50 determines that the predicted setup time is larger than the reference value, the second inference is set with the predicted setup time as the loss value (loss function). Input to the unit 42. Then, in the second inference unit 42, the acquired loss value becomes equal to or less than the average setup time, that is, the first inference unit 41 uses the value of the predicted basic operation parameter output by the second inference unit 42. Learning (back propagation) is executed so that the output predicted setup time is equal to or less than the average setup time. If the time comparison unit 50 determines that the predicted setup time is less than or equal to the reference value, the learning process using the other first learning data set TD1 is performed, or a series of learning processes are completed. It becomes. In the present embodiment, the average setup time is adopted as the reference value, but it is naturally possible to set a time shorter than the average setup time as the reference value. In that case, the trained AI module obtained as a result of the learning will be able to output the predicted basic operation parameters that can complete the start-up operation in a shorter setup time.

FIG. 4 is a flowchart showing an example of a control method of the learning inference processing device 40 according to the first embodiment of the present disclosure. Hereinafter, with reference to FIG. 4, a control method of the learning inference processing device 40 having the above-described configuration, that is, a machine learning process executed by using a computer such as an information processing device 20 will be described. The control method shown below will be described on the premise of the structure of the learning inference processing device 40 described above, but the structure of the learning inference processing device on which the control method is implemented does not affect the control method. Can be changed as appropriate in.

As shown in FIG. 4, the learning inference processing device 40 first prepares a plurality of first learning data sets TD1 to be used for learning in order to execute a series of learning processes (step S11), and a second. The AI module before learning is prepared in the inference unit 42 of (step S12). If the number of the first training data set TD1 group prepared in step S11 is too small, the prediction accuracy of the trained AI module obtained as a result of learning may be insufficient or overtraining may occur. Therefore, the number of the first learning data sets TD1 to be prepared should be appropriately set in consideration of the degree of learning, the learning rate, and the like. Further, in the neural network model of the pre-learning AI module prepared in step S12, the weight value associated with the node and the neuron may be set to the initial value.

When the preparation of the training data set group and the pre-learning AI module is completed, next, one first training data set TD1 to be used for learning is selected from the plurality of prepared first training data set TD1 groups. (Step S13). Then, using the recipe parameters, the device parameters, and the previous start-up result in the selected first learning data set TD1 as input data, the neurons of the input layer of the neural network model in the second inference unit 42 are used. Is input to (step S14). When the input data is input to the input layer in step S14, the predicted basic operation parameter is output to the output layer of the second inference unit 42 as an output result. Then, the output predicted basic operation parameter and the three parameters in the first learning data set TD1 selected in step S13 are set as the second learning data set TD2, and the second learning data. The set TD2 is input as input data to the input layer of the neural network model in the first inference unit 41 (step S15). When the input data is input to the input layer of the first inference unit 41 in step S15, the predicted setup time is output as an output result to the output layer of the first inference unit 41. Therefore, the output predicted setup time is then sent to the time comparison unit 50 for comparison with the average setup time as a reference value (step S16).

As a result of step S16, when the predicted setup time is larger than the reference value, that is, longer than the average setup time (No in step S17), the predicted setup time is set as the loss value, and the loss value is set to be equal to or less than the average setup time. In addition, back propagation (error back propagation) is executed in the second inference unit 42 to adjust (update) the weight value in the neural network model of the AI module before learning (step S18). When the weight value is adjusted in step S18, the process returns to step S14 and the steps shown in steps S14 to S17 are repeated.

As a result of step S16, when the predicted setup time is equal to or less than the reference value, that is, the same as the average setup time or shorter than the average setup time (Yes in step S17), the predicted basic operation output immediately before by the second inference unit 42 is performed. It is determined that the value of the parameter is an appropriate value, and the training process using the corresponding first training data set TD1 is completed. Then, it is determined whether or not a series of learning is completed (step S19), and for example, when the first learning data set TD1 that requires learning prepared in step S11 remains (No in step S19). To, return to step S13, select one unlearned first learning data set TD1, and execute the steps shown in steps S14 to S18. When the first learning data set TD1 requiring learning does not remain, that is, when learning is completed for all of the plurality of first learning data sets TD1 prepared in advance in step S11 (Yes in step S19). , The AI module after the weight value is adjusted in the second inference unit 42 is regarded as a learned AI module (step S20), and a series of control methods are completed.

By executing the control method of the learning inference processing device 40 described above, it is possible to output the predicted basic operation parameters by inputting the recipe parameters, the device parameters, and the three data of the previous start-up result. You can get a finished AI module.

The control method of the learning inference processing device 40 described above is in the form of a control program including computer-readable instructions executed by one or a plurality of processors in a computer such as the information processing device 20 or other hardware, or the control method thereof. It can be provided in the form of a computer-readable storage medium containing a control program.

FIG. 5 is a flowchart showing an example of the setup process of the ion implantation processing apparatus 10 in the ion implantation processing system 1 according to the first embodiment of the present disclosure. For example, when the ion implantation processing apparatus 10 is started, the setup process shown in FIG. 5 is started. Specifically, when the ion implantation processing device 10 is started, first, the parameter information necessary for the parameter acquisition unit 31 to drive the ion implantation processing device 10, the recipe parameter, the device parameter, and the previous stand in the present embodiment. Acquire three of the raised results (step S31). At this time, the recipe parameters may be acquired from the input / output interface 35, and the device parameters and the previous startup result may be acquired from the data storage unit 34. Next, the parameter information acquired in step S31 is sent to the basic operation parameter prediction unit 32 via the internal bus 36, and the basic operation parameter prediction unit 32 uses the received parameter information as the second second in the learning inference processing device 40. It is input to the input layer of the learned AI module stored in the inference unit 42 (step S32).

When the parameter information is input to the learned AI module in the second inference unit 42 in step S32, the predicted basic operation parameter is output to the output layer of the learned AI module. The basic operation parameter prediction unit 32 sends the output predicted basic operation parameter to the basic operation parameter automatic adjustment unit 33, and the basic operation parameter automatic adjustment unit 33 uses the received predicted basic operation parameter as the initial value of the basic operation parameter. , Is input to the software for specifying the adjustment value of the basic operation parameter stored in the basic operation parameter automatic adjustment unit 33 (step S33). The software in which the initial value of the basic operation parameter is input automatically adjusts the optimum basic operation parameter for operating the ion injection processing device 10 based on the initial value (step S34), and the basic operation obtained as a result is obtained. Complete a series of setup processes by identifying parameter adjustments.

Since the ion implantation processing system 1 according to the present embodiment can realize the above-mentioned setup process, the ion implantation processing system 1 has basic operation parameters such that the time required for setting up the ion implantation processing apparatus 10 is shortened. The initial value can be obtained and the setup time can be shortened.

<Second embodiment>
In the learning process of the second inference unit 42 according to the first embodiment described above, the learning process is performed so that the predicted setup time is smaller than the reference value. However, in rare cases, if the value of the predicted basic operation parameter output by the learned AI module obtained through such a learning method is input to the first inference unit 41, a predicted setup time smaller than the reference value can be obtained. However, it may not be practical to use as the initial value of the basic operation parameter. Specifically, for example, as a basic operating parameter, an event occurs in which the beam current value supplied to the plasma chamber in the ion source 11 is inferred as a value so small that plasma cannot be continuously generated in the plasma chamber. obtain. Therefore, as the second embodiment, the following includes a learning inference processing device having a learned AI module capable of inferring the value of a practical predictive basic operation parameter, a control method of the learning inference processing device, a control program, and the like. The processing system including the above will be described. The basic structure of the ion implantation processing system and the learning inference processing apparatus according to the present embodiment is the same as the structure described with reference to FIGS. 1 and 2 in the first embodiment, and is related to the learning process. Only the point to do is different. Therefore, regarding the basic structure of the ion implantation processing system and the learning inference processing apparatus according to the present embodiment, the description thereof is omitted by referring to the one described in the first embodiment, and the first embodiment is used. The explanation will focus on the different points.

FIG. 6 is a schematic explanatory diagram showing an example of a learning process in the learning inference processing device 40A according to the second embodiment of the present disclosure. As shown in FIG. 6, the learning inference processing device 40A according to the present embodiment includes a learning data set (third learning data set TD3) stored in the learning data set storage unit 43 and the learning data set TD3. The second inference unit 42A learned using the training data set is mainly different from that of the first embodiment. On the other hand, among those provided in the learning inference processing device 40A, the first inference unit 41 and the time comparison unit 50 can adopt the same ones as described in the first embodiment.

The third learning data set TD3 used in the learning inference processing apparatus 40A according to the present embodiment is an ion injection processing system 1 based on the three parameters in addition to the recipe parameter, the apparatus parameter and the previous start-up result. Includes the basic operation parameters (actual control information; hereinafter, this value is referred to as "actual basic operation parameters") actually used when operating. The actual basic operation parameters used in the present embodiment are set in advance according to, for example, the above three parameters, and the initial values of the basic operation parameters input to the automatic adjustment unit 33 (that is, the ion implantation processing device 10 is actually operated. It may be the actual value of the basic operation parameter used at the time of making it. Alternatively, it may be the average value of the basic operation parameters at the time of the previous operation or the basic operation parameters at the time of several operations in the past. As for this actual basic operation parameter, any specific method can be adopted as long as it is a value specified based on a specific recipe parameter, device parameter, and the previous start-up result.

In relation to the above, since the recipe parameter, the device parameter, and the previous start-up result in the third learning data set TD3 in the present embodiment are associated with the actual basic operation parameter, these three parameters are set. The parameters are the same as those actually used for controlling the ion implantation processing device 10 in the control device 30.

As shown in FIG. 6, the second inference unit 42A of the learning inference processing device 40A according to the present embodiment is realized by the AI module provided with the neural network model, and the input data and the output data are , It is the same as the second inference unit 42 of the learning inference processing device 40 according to the first embodiment described above. On the other hand, the second inference unit 42A uses the parameter loss (difference information) when learning (back-propagating) the neural network model, and the learning inference processing device 40 according to the first embodiment. It is different from the second reasoning unit 42 of. The learning method using parameter loss will be described later.

In addition to the above-described configuration, the learning inference processing device 40A according to the present embodiment further includes a parameter loss calculation unit 60. The parameter loss calculation unit 60 is output by the actual basic operation parameter in the third learning data set TD3, the predicted basic operation parameter output by the second inference unit 42A, and the first inference unit 41. The parameter loss (loss function) used for learning the AI module of the second inference unit 42A is calculated based on the predicted setup time. Since the above-mentioned parameter loss is used for learning in the learning process of the learning inference processing device 40A according to the present embodiment, is the time comparison unit 50 according to the present embodiment larger than the reference value? All you have to do is compare whether or not. Alternatively, the difference between the predicted setup time and the reference value may be calculated.

The value of the parameter loss specified by the parameter loss calculation unit 60 described above is preferably specified by using at least the squared error between the actual basic operation parameter and the predicted basic operation parameter and the predicted setup time. Further, the value of the parameter loss can be calculated by various methods, but it is preferably defined by using the following formula (1).

Difference information = α1 x time information + α2 x squared error (1)

At this time, α1 and α2 can be set to arbitrary coefficients, but as an example, if the relationship between α1 and α2 satisfies 0.01 ≦ α1 / α2 ≦ 0.1, the value of the square error is learned. It is greatly reflected in, and is more preferable.

The parameter loss defined by the above equation (1) is greatly affected by the square error. Therefore, the predicted basic operation parameters inferred by the learned AI module learned using this parameter loss fully consider the values of the basic operation parameters actually used for the control of the ion implantation processing apparatus 10. ..

FIG. 7 is a flowchart showing an example of a control method of the learning inference processing device 40A according to the second embodiment of the present disclosure. Hereinafter, the control method of the learning inference processing device 40A according to the present embodiment will be described with reference to FIGS. 6 and 7.

As shown in FIG. 7, the learning inference processing device 40A first prepares a plurality of third learning data sets TD3 to be used for learning in order to execute a series of learning processes (step S41), and a second. The AI module before learning is prepared in the inference unit 42A of the above (step S42). When the preparation of the training data set group and the pre-learning AI module is completed, next, one third training data set TD1 to be used for training is selected from the plurality of prepared third training data sets TD3 group. (Step S43). Then, in the selected third learning data set TD3, three parameters excluding the actual basic operation parameter, that is, the recipe parameter, the device parameter, and the previous start-up result are used as input data, and the second inference is made. Inputs are made to neurons in the input layer of the neural network model in part 42A (step S44). When the input data is input to the input layer in step S44, the predicted basic operation parameter is output to the output layer of the second inference unit 42A as an output result. Then, the output predicted basic operation parameter, the recipe parameter in the first third learning data set TD3 selected in step S43, the device parameter, and the previous start-up result are used as the second learning data. The set TD2 is used, and the second learning data set TD2 is input to the input layer of the neural network model in the first inference unit 41 as input data (step S45). When the input data is input to the input layer of the first inference unit 41 in step S45, the predicted setup time is output as an output result to the output layer of the first inference unit 41. Therefore, the output predicted setup time is then sent to the time comparison unit 50 for comparison with the average setup time as a reference value (step S46).

As a result of step S46, when the predicted setup time is larger than the reference value, that is, longer than the average setup time (No in step S47), the parameter loss used for learning of the second inference unit 42A in the parameter loss calculation unit 60 Calculate the value (step S48). Then, back propagation is executed in the second inference unit 42A so that the calculated parameter loss value becomes small, and the weight value in the neural network model of the AI module before learning is adjusted (step S49). .. When the weight value is adjusted in step S49, the process returns to step S44 and the steps shown in steps S44 to S47 are repeated.

As a result of step S47, when the predicted setup time is equal to or less than the reference value, that is, the same as the average setup time or shorter than the average setup time (Yes in step S47), the predicted basic operation output by the second inference unit 42A immediately before is performed. It is determined that the value of the parameter is an appropriate value, and the training process using the corresponding third training data set TD3 is completed. Then, it is determined whether or not a series of learning is completed (step S50), and if the predetermined end condition is not satisfied, for example, the third learning data set TD3 that requires learning prepared in step S41 If it remains (No in step S50), the process returns to step S43, the unlearned third learning data set TD3 is selected, and the steps shown in steps S44 to S49 are executed. Then, when the third learning data set TD3 that needs to be learned does not remain, that is, the learning is completed for all of the plurality of third learning data sets TD3 prepared in advance in step S41 (Yes in step S50). In the second inference unit 42A, the AI module after the weight value is adjusted is set as the learned AI module (step S51), and a series of control methods are completed. The end condition defined in step S50 is not limited to the above example, and substitutes various other methods capable of determining whether or not the inference accuracy of the AI module being trained has reached a desired level. Can be implemented as a target.

By executing the control method of the learning inference processing device 40A described above, by inputting three data of the recipe parameter, the device parameter, and the previous start-up result as the second inference unit 42A, the predicted basic operation parameter. You can get a trained AI module that can output. Further, in the above-mentioned control method, since the actual basic operation parameter is used in the learning, the predicted basic operation parameter output by the learned AI module of the second inference unit 42A is surely a practical value. Will be output to. As a result, the initial value of the basic operation parameter input to the basic operation parameter automatic adjustment unit 33 is always a practical value, so that an impractical value is not input as the initial value, and the initial value is not input at the time of setup. The error rate can be reduced.

The control method of the learning inference processing device 40A described above is in the form of a control program for causing a computer or the like to execute the control method, or stores the control program, as in the first embodiment. It can be provided in the form of a computer-readable storage medium. Further, the ion implantation treatment system according to the present embodiment can execute the same setup process as that described with reference to FIG. 5 in the first embodiment, and shorten the time required for the setup. It can be said that it is a possible system.

<Third embodiment>
As described above, in the learning inference processing devices 40 and 40A according to the first and second embodiments, the recipe parameters and the devices are used as input data (explanatory variables) input to the second inference units 42 and 42A. Three parameters, the parameter and the result of the previous startup, are adopted. However, if at least one of these three parameters, the recipe parameter and the device parameter, is used as input data, it is possible to generate an AI module capable of obtaining a desired inference result. Therefore, the learning inference processing device 40B according to the third embodiment will be described below as a modified example of the first embodiment. The learning inference processing device 40B according to the present embodiment is the same as the learning inference processing device 40 according to the first embodiment except that the structure of the data used for learning is different. Will explain only the part different from the learning inference processing device 40 according to the first embodiment, and omit the description about the same part as the learning inference processing device 40 according to the first embodiment. Similarly, the configuration of the ion implantation processing system including the learning inference processing device 40B, the control method and the control program of the learning inference processing device 40B are also the same as those described in the first embodiment, and thus the description thereof is omitted. do.

FIG. 8 is a schematic explanatory diagram showing an example of a learning process in the learning inference processing device 40B according to the third embodiment of the present disclosure. As shown in FIG. 8, the learning inference processing device 40B according to the present embodiment does not include the device parameters in the input data input to the second inference unit 42B and the first inference unit 41B. That is, the fourth learning data set TD4 used in the learning inference processing device 40B according to the present embodiment is composed of only the recipe parameter and the previous start-up result.

The learning inference processing device 40B has a first inference unit 41B composed of an AI module including a neural network model, and a second inference unit composed of an AI module including a neural network model like the first inference unit 41B. A unit 42B and a time comparison unit 50 are provided. Of these, the time comparison unit 50 is the same as that described in the first and second embodiments, but the first inference unit 41B and the second inference unit 42B are inputs used for the learning. The data differ from those described in the first and second embodiments. Specifically, the input data input to the second inference unit 42B is the parameters in the fourth learning data set TD4, that is, the recipe parameter and the previous start-up result. When the input data composed of the above two parameters is input to the input layer of the neural network model of the second inference unit 42B, the predicted basic operation parameter is output to the output layer as an output result. The first inference unit 41B includes a trained AI module as shown in the first and second embodiments. The fifth learning data set TD5 is input as input data to the input layer of the neural network model of the trained AI module included in the first reasoning unit 41B. In the fifth learning data set TD5, the predicted basic operation parameter output by the second inference unit 42B and the input data input to the second inference unit 42B when the predicted basic operation parameter is output. That is, the recipe parameters included in the fourth learning data set TD4 and the previous start-up result are included. When the input data composed of the fifth learning data set TD5 is input to the input layer of the neural network model in the first inference unit 41B, the predicted setup time is output as an inference result to the output layer as an output result. Will be done. Then, the predicted setup time output from the first inference unit 41B is compared with the reference value in the time comparison unit 50, and if it is larger than the reference value, the difference is input to the second inference unit 42B as a loss value. Then, the learning of the second inference unit 42B is executed.

By providing the above-described configuration, the learning inference processing device 40B according to the present embodiment learns a learned AI module capable of inferring a predicted basic operation parameter having a short setup time based on only two parameters. You will be able to do it.

In the third embodiment described above, as the fourth learning data set TD4, the one including the recipe parameter and the previous start-up result is illustrated, but instead of this, the device parameter and the previous start-up result are illustrated. It can also include two of. Further, the fourth learning data set TD4 may include the actual basic operation parameters described in the second embodiment in addition to the above-mentioned two parameters. The learning process and the setup process using such a learning data set are common to the methods described in the first and second embodiments described above.

The learning process in each of the above-described embodiments has been described only when so-called batch learning is performed for the purpose of simplifying the explanation, but the present disclosure is not limited to this, and online learning can also be adopted. .. Further, in each embodiment, a learning AI module used by the basic operation parameter prediction unit 32 is stored in the second inference units 42, 42A, and 42B. It is also possible to separate only the completed AI module, store it in the control device 30 or another storage device, and use it.

The present disclosure is not limited to the above-described embodiment, and various modifications can be made without departing from the gist of the present disclosure. And all of them are included in the technical idea of the present disclosure.

1 Ion implantation processing system (processing system)
10 Ion implantation processing device (processing device)
11 Ion source 12 Mass analyzer 13 Analytical slit 14 Accelerator tube 15 Energy separator 16 Scanner 17 Beam parallelizer 18 Board drive device 20 Information processing device 21 Processor 22 Memory 23 Communication interface 30 Control device 31 Parameter acquisition unit 32 Basic operation Parameter prediction unit 33 Basic operation Parameter automatic adjustment unit 34 Data storage unit 35 I / O interface 36 Internal bus 40 Learning inference processing device 41, 41B First inference unit 42, 42A, 42B Second inference unit 43 Learning data set storage Part 50 Time comparison part 60 Parameter loss calculation part IB Ion beam TD1 to TD5 First to fifth learning data set W board

Claims (9)

The processing device that performs the processing and
Control information for controlling the processing device output by the second inference unit when learning the second inference unit, recipe information regarding specifications of processing executed by the processing device, or startup of the processing device. A first inference unit that inputs one of the device information related to the operation and the control information measurement result that is the result of measuring the control information at the start and end of the processing device, infers the time information, and outputs the time information.
A learning inference processing device having the second inference unit for inputting the recipe information or the device information and the control information measurement result and outputting the control information to the processing device. ,
Processing system. In the processing system
The second inference unit included in the learning inference processing device is
Learning is repeated by updating the weight value of the neural network included in the second inference unit until the time information output by the first inference unit becomes equal to or less than a predetermined reference value, and the time information is equal to or less than the reference value. Equipped with a trained AI module when
The processing system according to claim 1. In the processing system
The first inference unit further
Input either recipe information regarding the specifications of the processing executed by the processing device or device information regarding the startup operation of the processing device, whichever is the other.
The processing system according to claim 1 or 2. When at least one of the recipe information or the device information input by the processing device to the second inference unit is the same as the recipe information or the device information that actually controls the processing device.
The square error between the actual control information, which is the control information used when the processing device actually operates, and the control information output by the second inference unit is further input to the second inference unit.
The second inference unit repeats learning so that the value of the difference information defined by the square error and the time information becomes small.
The processing system according to any one of claims 1 to 3. The difference information uses the time information and the squared error.
Difference information = α1 x time information + α2 x squared error (α1 and α2 are arbitrary coefficients)
Defined by the formula of
The processing system according to claim 4. The relationship between the α1 and the α2 satisfies 0.01 ≦ α1 / α2 ≦ 0.1.
The processing system according to claim 5. In the learning inference processing device that controls the processing device,
The learning inference processing device is
Control information for controlling the processing device output by the second inference unit when learning the second inference unit, recipe information regarding specifications of processing executed by the processing device, or startup of the processing device. A first inference unit that inputs one of the device information related to the operation and the control information measurement result that is the result of measuring the control information at the start and end of the processing device, infers the time information, and outputs the time information.
The first, which inputs the recipe information or the device information and the control information measurement result, and outputs the control information to the processing device based on the time information output by the first inference unit. Equipped with 2 inference units,
Learning inference processing device. In a control method of a learning inference processing device including a first inference unit and a second inference unit while controlling the processing device,
The learning inference processing device
Either the control information for controlling the processing device output by the second inference unit, the recipe information regarding the specifications of the processing executed by the processing device, or the device information regarding the start-up operation of the processing device. The second inference is made by learning using the time information inferred and output by the first inference unit, which inputs the control information measurement result which is the result of measuring the control information at the start and end of the processing device. Generate a part,
The second reasoning unit is
Enter either one of the recipe information or the device information and the control information measurement result.
Output the control information to the processing device.
How to control the learning inference processing device. In a control program of a learning inference processing device that controls a processing device and includes a first inference unit and a second inference unit.
In the learning inference processing device,
Either the control information for controlling the processing device output by the second inference unit, the recipe information regarding the specifications of the processing executed by the processing device, or the device information regarding the start-up operation of the processing device. The second inference is made by learning using the time information inferred and output by the first inference unit, which inputs the control information measurement result which is the result of measuring the control information at the start and end of the processing device. Generate a part,
In the second inference unit,
Either one of the recipe information or the device information and the control information measurement result are input.
The control information is output to the processing device.
A control program for a learning inference processor.

Images