Method and program for operating semiconductor device evaluation device by means of character string

Abstract

The purpose of the present disclosure is to provide technology with which it is possible to obtain, from a language model by means of input of a character string into the language model, an instruction to an evaluation device for evaluating a semiconductor device. In a method according to the present disclosure, an instruction to an evaluation device for evaluating a semiconductor device is inputted into a language model, whereby specification data describing operation specifications of the evaluation device is received from the language model. The language model is fine-tuned in advance by means of at least one of domain knowledge pertaining to the semiconductor device or domain knowledge relating to the evaluation device (see fig. 2).

Description

Method and program for operating semiconductor device evaluation equipment using character strings

[0001] The present disclosure relates to a technique for operating a semiconductor evaluation device using a character string.

[0002] In recent years, there has been active development of technology that uses a large language model (LLM) to receive answers to questions written in natural language from an artificial intelligence (AI) in natural language. Non-Patent Document 1 below describes the GPT model developed by OpenAI as one example.

[0003] Long Ouyang et al., “Training language models to follow instructions with human feedback”, 2022, URL: https: / / arxiv.org / abs / 2203.02155

[0004] A general-purpose large-scale language model such as that described in Non-Patent Document 1 learns the frequency of occurrence of character strings by, for example, repeatedly masking part of a character string and estimating the masked part for various document data. This allows knowledge in various fields (sometimes called domain knowledge) to be learned in advance using character strings, and when a question related to that field is given, an answer can be output based on the previously learned knowledge.

[0005] Although such large-scale language models can learn knowledge in various fields, they may not have sufficient knowledge in a specific field. Furthermore, while such large-scale language models are useful in the process of providing a question as input and obtaining an answer to that question, they are not necessarily optimized for applications such as obtaining operating instructions for a specific device. The technology described in Non-Patent Document 1 describes a method for training a large-scale language model, but does not particularly consider specific methods for such applications.

[0006] The present disclosure has been made in consideration of the above-described problems, and aims to provide a technology that enables instructions for an evaluation device that evaluates semiconductor devices to be obtained from a language model by inputting a string of characters into the language model.

[0007] The method according to the present disclosure includes inputting instructions for a semiconductor device evaluation device into a language model, and receiving specification data describing operational specifications of the evaluation device from the language model, the language model having been fine-tuned in advance using at least one of domain knowledge of the semiconductor device and domain knowledge of the evaluation device.

[0008] According to the method of the present disclosure, instructions for an evaluation device that evaluates semiconductor devices can be obtained from a language model by inputting a string of characters into the language model. Other problems, configurations, advantages, and the like of the present disclosure will become apparent from the description of the following embodiments.

[0009] 1 is a block diagram of a semiconductor evaluation system 100 according to a first embodiment. FIG. 2 is a diagram illustrating an example of a text creation system 200 that uses a general-purpose LLM 210 configured based on a large-scale language model. FIG. 3 is a diagram illustrating an example of domain knowledge learned by fine tuning. FIG. 4 is a diagram illustrating the distinction between unsupervised learning and supervised learning. FIG. 5 is a schematic diagram illustrating a process in which a learning module 113 performs unsupervised learning. FIG. 6 is a schematic diagram illustrating another process in which a learning module 113 performs unsupervised learning. FIG. 7 is a schematic diagram illustrating a process in which a learning module 113 performs supervised learning. FIG. 8 is a schematic diagram illustrating variations in combinations of input / output / correct answer / loss function in supervised learning. FIG. 9 is a schematic diagram illustrating a process of reinforcement learning performed by a reinforcement learning module 114. FIG. 10 is a schematic diagram illustrating an example of a re-learning procedure by the reinforcement learning module 114. FIG. 11 is a schematic diagram illustrating another example of a re-learning procedure by the reinforcement learning module 114. FIG. 12 is a schematic diagram illustrating another example of a re-learning procedure by the reinforcement learning module 114. FIG. 13 is a schematic diagram illustrating another example of a re-learning procedure by the reinforcement learning module 114. FIG. 14 is a schematic diagram illustrating an example of an input to an LLM and an output from an LLM.

[0010] 1 is a block diagram of a semiconductor evaluation system 100 according to a first embodiment of the present disclosure. The semiconductor evaluation system 100 is a system that performs evaluation processes such as inspection and measurement on semiconductor devices (samples). The semiconductor evaluation system 100 includes a computer system 101, a semiconductor evaluation tool 102, a design data storage medium 103, and a simulator 104, which are connected to each other by a communication bus 105.

[0011] The computer system 101 has one or more processors 106 (including at least one of a CPU and a GPU), a memory 107, and a user interface 108. The user interface 108 includes a display (not shown) and one or more input devices. The display can display image data output from the semiconductor evaluation tool 102, information output from the processor 106, and the like.

[0012] The semiconductor evaluation tool 102 is, for example, a CD-SEM (Critical Dimension-Scanning Electron Microscope), which is a device that generates an image and a luminance signal waveform based on the detection of secondary electrons and backscattered electrons emitted from a sample when the sample is irradiated with an electron beam. The CD-SEM is a device that measures the dimensions and two-dimensional shape of a pattern formed on a semiconductor device based on the luminance signal waveform. The semiconductor evaluation tool 102 may also be an inspection device that includes a charged particle beam, such as an SEM for defect review, or an optical inspection device.

[0013] The design data storage medium 103 stores design data that describes the physical circuit structure of a semiconductor device in, for example, a GDS II or OASIS file format.

[0014] The simulator 104 generates an observation image of the sample through simulation, or estimates a signal profile indicating the intensity distribution of secondary electrons and backscattered electrons. Specifically, the simulator 104 uses random numbers to reproduce the generation process of secondary electrons generated by a beam incident on the sample and backscattered electrons emitted from the sample, and performs repeated calculations to calculate the emission angle, energy, etc. of the secondary electrons and backscattered electrons. Monte Carlo simulation, etc., is used for the calculation.

[0015] The simulator 104 has a built-in processor (not shown) and a storage medium for storing data and programs required for the simulation. The simulator 104 is configured to read out necessary information (e.g., the pattern shape included in the semiconductor device, the material composition of the semiconductor device) from the design data storage medium 103, and to estimate an electron microscope image (when the semiconductor evaluation tool 102 is an electron microscope) based on previously stored structural information of the semiconductor evaluation tool 102, such as the position of a detector in the semiconductor evaluation tool 102 (the position relative to the beam irradiation location), and control conditions of the measurement device, including the beam irradiation conditions, which will be described later.

[0016] The simulation performed by the simulator 104 may derive images, brightness information, profiles, etc. using not only trajectory calculation methods such as Monte Carlo simulation, but also models that define the relationships between layout data (including three-dimensional data such as film thickness), coefficients that vary depending on the material, device conditions of the electron microscope, brightness information, etc. Furthermore, image data, etc. may be estimated by inputting known information into a learning device that has undergone learning using data sets of device conditions of the electron microscope, material information, 3D layout information, and brightness information (image data, profiles, etc.) as training data.

[0017] The memory 107 is configured with volatile and / or nonvolatile memory. The memory 107 has a non-transitory computer-readable storage medium. The memory 107 stores the following data: (a) an operating system 109 that executes various measurement or inspection-related processing steps (described later); (b) a learning module 110 that generates a learning model based on inputs such as a large-scale language model (described later), text data, and information output from the semiconductor evaluation tool 102; (c) a text creation module 111 that creates text data and the like using the model learned by the learning module 110; and (d) a measurement database 112 that stores measurement results and the like output from the semiconductor evaluation tool 102.

[0018] The learning module 110 has two modules: (a) a supervised / unsupervised learning module 113; and (b) a reinforcement learning module 114. The writing module 111 has a writing model 115. These sub-modules and models are described in more detail below.

[0019] In addition to the above, the memory 107 (e.g., a non-transitory computer-readable storage medium) may store, for example, the following: (a) an operating program executed by one or more processors included in the semiconductor evaluation tool 102; and (b) a recipe generation module that generates the operating program (recipe) based on parameters input from the user interface 108.

[0020] The recipe generation module is configured to execute recipe generation and correction processing of an already generated recipe, for example, based on the output of a text creation module 111 (described later). The recipe generation module is configured to generate a recipe based on, for example, design data, a simulation image generated by the simulator 104 based on input of the design data, parameters input from the user interface 108, etc.

[0021] FIG. 2 is a diagram showing an example of a writing system 200 that uses a general-purpose LLM 210 configured based on a large language model (LLM). FIG. 2 illustrates a writing system 200 that is primarily used for measuring or inspecting semiconductor devices. The writing system 200 generates an LLM suitable for measuring or inspecting semiconductor devices based on the general-purpose LLM 210. The LLM for measuring or inspecting semiconductor devices is based on the existence of a general-purpose LLM 210 that has been trained in advance using a large amount of language data.

[0022] The document creation system 200 includes: (a) an equipment supplier-side computer system 220 that performs fine tuning of the general-purpose LLM 210 mainly based on domain knowledge of the evaluation equipment; and (b) an equipment user-side computer system 230 that performs further fine tuning of the evaluation equipment LLM 222 provided by the equipment supplier based on domain knowledge of the semiconductor device (e.g., domain knowledge about manufacturing conditions in the manufacturing process of semiconductor wafers).

[0023] The computer system 220 includes a learning module 221 for the evaluation device LLM. The learning module 221 performs fine tuning on the generic LLM 210 mainly using domain knowledge of the evaluation device (semiconductor evaluation tool 102 in FIG. 1 ). As a result, the generic LLM 210 further learns the domain knowledge of the evaluation device and is reconfigured as an evaluation device LLM 222.

[0024] The computer system 230 includes a semiconductor LLM learning module 231. The learning module 231 performs further fine tuning of the evaluation device LLM 222 using domain knowledge of semiconductor devices. As a result, the evaluation device LLM 222 further learns domain knowledge of semiconductor devices and is reconfigured as the semiconductor LLM 232.

[0025] The computer system 230 further includes a reinforcement learning module 233. The reinforcement learning module 233 performs reinforcement learning, which will be described later, on the semiconductor LLM 232, thereby performing re-learning on the semiconductor LLM 232 so that the semiconductor LLM 232 can output an evaluation recipe that enables the evaluation device to achieve a desired target value. As a result, the semiconductor LLM 232 is reconfigured as a sentence creation model.

[0026] The computer system 230 further includes a sentence creation module 234 including the sentence creation model that has been subjected to the reinforcement learning. The sentence creation module 234 uses the sentence creation model 115 to output an output string as a response to an input string.

[0027] In the system configuration illustrated in FIG. 2 , the equipment supplier's computer system 220 performs fine tuning using domain knowledge of an evaluation device such as an electron microscope, and the equipment user's computer system 230 performs fine tuning using domain knowledge of semiconductor devices. However, the computer systems that perform fine tuning are not limited to this. For example, the equipment supplier's computer system 220 may perform fine tuning using domain knowledge related to semiconductor devices, or the equipment user's computer system 230 may perform fine tuning using domain knowledge of the evaluation device. Similarly, reinforcement learning may be performed in either computer system. Furthermore, the equipment supplier's computer system 220 and the equipment user's computer system 230 may be configured as an integrated unit.

[0028] The computer systems 220 and 230 can each be configured by the computer system 101 described in Fig. 1. That is, (a) the learning module 221 and the learning module 231 can each be configured by the supervised / unsupervised learning module 113, (b) the reinforcement learning module 233 can be configured by the reinforcement learning module 114, and (c) the writing module 234 can be configured by the writing module 111 and the writing model 115. Fig. 1 shows an example in which these modules are integrated into a single computer system 101.

[0029] <First Embodiment: Fine Tuning> The general-purpose LLM 210 is configured in advance to learn various domain knowledge so that it can output an answer string that reflects that domain knowledge in response to an input question string. A wide range of domain knowledge can be learned. For example, if a specific book has been studied, a summary of that book can be provided as an answer. Alternatively, if a specific programming language has been studied, a desired function can be described in natural language and input to the general-purpose LLM 210, and source code that implements that function can be output.

[0030] Although the general-purpose LLM 210 has such versatility, it cannot output useful answers for knowledge in fields that it has not previously learned, and can only provide answers that are merely general. Therefore, in order to have the general-purpose LLM 210 learn knowledge in a specific field, tuning known as fine tuning must be performed. The present disclosure aims to input desired operations for a semiconductor device evaluation system (semiconductor evaluation tool 102 in FIG. 1 ) into the LLM and obtain an operational specification (evaluation recipe) that can realize the desired operation as a response from the LLM. In other words, the objective is to operate the semiconductor device evaluation system using natural language, etc. Therefore, in embodiment 1, the general-purpose LLM 210 is first fine-tuned to learn domain knowledge of the evaluation system and domain knowledge of semiconductor devices.

[0031] 3 shows an example of domain knowledge learned by fine-tuning. The "Evaluation Device" row in the "Domain Knowledge" column in Fig. 3 is an example of domain knowledge related to the evaluation device learned by the learning module 221. The other rows in the "Domain Knowledge" column in Fig. 3 are examples of domain knowledge related to semiconductor devices learned by the learning module 231.

[0032] Fine tuning of domain knowledge related to the evaluation device can be performed by having the general-purpose LLM 210 learn document data such as the instruction manual for the evaluation device and the functional specifications of the evaluation device. This allows knowledge related to the operation (usage) of the evaluation device to be learned. Furthermore, the specifications of the image processing performed by the evaluation device or the results of the image processing can also be learned as domain knowledge of the evaluation device. The specifications of the image processing can be learned from document data such as specifications or textbooks that describe the contents of the image processing. The results of the image processing can be learned by studying the changes in the observed image before and after actually performing the image processing.

[0033] A recipe is data describing the operational specifications that define the evaluation process performed by the evaluation equipment. Depending on the user of the evaluation equipment, there may be parameters that are prohibited from being used in the recipe or that impose restrictions on the recipe. Such user-specific restrictions (parameter items that impose restrictions and the content of the restrictions) are called local rules. Because local rules are user-specific, it is more appropriate to learn them as domain knowledge specific to the semiconductor device that the user is evaluating, rather than learning them as general domain knowledge of the evaluation equipment. Therefore, we decided to learn local rules as domain knowledge of the semiconductor device.

[0034] Since recipes describe the instructions for the evaluation equipment, they appear to users as a list of parameters and are not necessarily easy to read. Measurement instructions are documents that describe the steps for creating a recipe in natural language, and in some cases, measurement instructions are easier for users to read. Therefore, we decided to learn measurement instructions as part of the domain knowledge of semiconductor devices.

[0035] The specifications of the image processing performed by the evaluation device can also be learned as domain knowledge of semiconductor devices. The learning method is the same as when learning as domain knowledge of the evaluation device.

[0036] A user who manufactures semiconductor devices may have a simulator that simulates the manufacturing process of semiconductor devices. This simulator uses design data that describes the three-dimensional shape, materials, and manufacturing process information (e.g., manufacturing conditions of semiconductor manufacturing equipment) of the sample to be measured (semiconductor device). In addition, the simulator may also use design data that describes the two-dimensional shape of the shape pattern formed on the semiconductor device. This design data can be learned as domain knowledge of semiconductor devices.

[0037] When an evaluation device performs evaluation by irradiating a sample (semiconductor device) with an electron beam (or any other charged particle beam), data describing the interaction between the electron beam and the sample material can be learned as domain knowledge of the semiconductor device.

[0038] Figure 4 shows the division between unsupervised learning and supervised learning. The instruction manual or functional specifications of the evaluation device can be learned by unsupervised learning. The image processing specifications can be learned by unsupervised learning. The image processing results can be learned by supervised learning. The measurement instructions can be learned using either unsupervised or supervised learning. The local rules can be learned by unsupervised learning. The design data can be learned by supervised learning. The interaction between the electron beam and the sample can be learned by supervised learning.

[0039] FIG. 5A is a schematic diagram showing the process by which the learning module 113 performs unsupervised learning. Unsupervised learning can be performed, for example, by Masked LM. Masked LM is a technique that learns the frequency of occurrence of character strings in a document by repeatedly masking a portion of the document and estimating the masked portion, thereby learning the knowledge described in the document. Learning progresses by feeding back the difference (loss) between the estimation result of the masked portion and the correct character string to the LLM. An example of a loss function is, for example, cross-entropy error, but is not limited to this.

[0040] FIG. 5B is a schematic diagram showing another process in which the learning module 113 performs unsupervised learning. Unsupervised learning can be performed by NSP (Next Sentence Prediction) in addition to Masked LM. Any other method may be used. NSP repeatedly predicts whether two sentences are semantically connected. This makes it possible to learn the correct connection between sentences. The loss function is the same as that of Masked LM.

[0041] FIG. 6 is a schematic diagram showing the process by which the learning module 113 performs supervised learning. Here, an example is shown in which a request for an evaluation device written in a natural language string and the design data described in FIG. 3 are input to the LLM. In this example, the LLM outputs a measurement instruction document describing the procedure for creating an evaluation recipe that implements the requested procedure. A measurement instruction document that can actually satisfy the request is provided as the correct answer data. Supervised learning can be performed by feeding back to the LLM the loss between the measurement instruction document output by the LLM and the correct measurement instruction document.

[0042] FIG. 7 shows variations in the combination of input / output / correct answer / loss function in supervised learning. When design data is used as input and measurement instructions are used as output and correct answer, the loss function is, for example, cross-entropy error. When measurement instructions are used as input and a recipe is used as output and correct answer, the loss function is, for example, cross-entropy error. When learning about the interaction between an electron beam and a sample, the input is the imaging conditions (voltages at each part, currents at each part, etc.) and 3D design data when the evaluation device captures an observed image of the semiconductor device, or the imaging conditions and information about the semiconductor device material, the output is the observed image, and the loss function is, for example, mean squared error. When learning about the results of image processing, the input is the recipe (particularly the filtering conditions for the image processing performed by the evaluation device), the output is the observed image, and the loss function is, for example, mean squared error.

[0043] The learning module 113 (or the learning modules 221 and 231) performs unsupervised learning / supervised learning according to the above procedure, thereby allowing the general-purpose LLM 210 to learn domain knowledge of the evaluation device and domain knowledge of the semiconductor device through fine tuning. As a result, when an input sentence related to the evaluation device or the semiconductor device is given, the semiconductor LLM 232 can output a response that reflects this domain knowledge.

[0044] Summary of First Embodiment The computer system 101 (or the computer systems 220 and 230) according to the first embodiment fine-tunes the general-purpose LLM 210 to learn domain knowledge of the evaluation equipment and domain knowledge of the semiconductor device. As a result, when an input statement related to the evaluation equipment or the semiconductor device is given, the semiconductor LLM 232 can output a response reflecting this domain knowledge. Since the semiconductor LLM 232 learns measurement instructions and recipes in the fine tuning process, the semiconductor LLM 232 can output these as output (see FIG. 7 ). That is, when a request for the evaluation equipment is input to the semiconductor LLM 232 using a character string, a response corresponding to the request can be obtained from the semiconductor LLM 232 using a measurement instruction or a recipe. This allows instructions to the evaluation equipment to be given using a character string via the LLM. In other words, the evaluation equipment can be operated using a character string.

[0045] <Embodiment 2> In the first embodiment, an example configuration was described in which instructions to the evaluation device can be given by character strings via the LLM. The first embodiment can at least achieve the objective of operating the evaluation device by character strings. However, depending on the parameter items and their value configurations described in the recipe for the evaluation device, it is possible that the operation of the evaluation device may not necessarily be desirable (e.g., the desired target value may not be achieved). Therefore, in the second embodiment of the present disclosure, based on the configuration of the first embodiment, an example configuration in which the LLM can output a more desirable recipe is described. The configuration of the computer system 101 (or the computer systems 220 and 230) is the same as that of the first embodiment.

[0046] 8 is a schematic diagram showing the process of reinforcement learning performed by the reinforcement learning module 114. The reinforcement learning module 114 (or the reinforcement learning module 233) retrains the LLM (semiconductor LLM 232 in embodiment 1) so as to output a recipe that can achieve the target value by feeding back to the LLM the results of evaluation of the semiconductor device by the evaluation device using the recipe (specification data in FIG. 8) output by the LLM, along with the target value. This is considered to enable the LLM to output a desirable recipe that can achieve the target value. The same can be considered when the LLM outputs a measurement instruction sheet.

[0047] The procedure for re-learning by the reinforcement learning module 114 is roughly as follows: Re-learning of the LLM is performed using at least one of the output data output by the evaluation device (or text data generated based on this; specific examples will be described later) as learning data. A string of characters describing instructions to the evaluation device in natural language sentences is input to the re-learned LLM. Data specifying the operation of the evaluation device (called specification data; for example, a recipe) is obtained as output from the LLM. The specification data is used to operate the evaluation device, and output data from the evaluation device is obtained. Re-learning is performed using the output data as learning data. Re-learning can be performed by repeating the above steps.

[0048] When relearning is performed, the target values ​​for each evaluation item are input to the LLM along with the evaluation results of the semiconductor device obtained by the evaluation device. Furthermore, the specification data used to obtain the evaluation results is also input. This allows the difference between the target values ​​and the evaluation results to be fed back, and also allows the specification data that caused the evaluation results to be learned. Therefore, relearning can be performed so that the LLM outputs specification data that brings the evaluation results closer to the target values.

[0049] The items evaluated by the evaluation device (i.e., items for which the evaluation results and target values ​​are targeted) can be, for example, the following: (a) throughput of the evaluation device; (b) values ​​measured by the evaluation device (e.g., pattern size); (c) feature quantities of observed images of semiconductor devices acquired by the evaluation device (e.g., noise amount, various histograms); and (d) probability of defect occurrence at each location on a semiconductor device. The evaluation device outputs an evaluation value for each of these evaluation items. These are fed back to the LLM together with the target values ​​for each evaluation item.

[0050] Examples of specification data that the LLM can output include: (a) data (i.e., recipe) that specifies the process that the evaluation device performs when evaluating a semiconductor device; (b) predicted defect positions in a semiconductor device (prediction of whether or not a defect exists at the measurement target position); (c) predicted causes of anomalies when an observed image of a semiconductor device acquired by the evaluation device is abnormal and measures to improve the abnormality; (d) an image quality improvement model that converts the observed image into an image of higher image quality; and (e) a pattern enhancement model that enhances specific shape patterns contained in the observed image. The LLM performs learning and re-learning in advance so as to output the above specification data.

[0051] When the LLM outputs the cause of the anomaly and its remedy, it may output, for example, a recipe modified to improve the anomaly and the cause of the anomaly along with the recipe before the improvement. This allows the user to understand which part of the recipe before the improvement had a problem. The image quality improvement model and the pattern enhancement model may be configured using data such as a mathematical model that implements these functions, or may be configured using any other data and / or program, such as artificial intelligence, that implements these functions.

[0052] FIG. 9 is a schematic diagram showing an example of a re-learning procedure by the reinforcement learning module 114. In this example, the LLM is pre-configured to receive as input a sentence describing instructions for the evaluation device and design data for a semiconductor device, and to output an operating recipe for the evaluation device as specification data. Since the design data may not necessarily be written in a format suitable for the LLM, a conversion module may be used to convert the design data into a format that is easy for the LLM to interpret. The input sentence may also be further converted appropriately (e.g., converting a natural language string into a numerical vector) before being input to the LLM.

[0053] The evaluation results of semiconductor devices output by the evaluation device may be fed back to the LLM as they are, or the evaluation results may be further processed to convert the output from the evaluation device into an evaluation value, which is then fed back to the LLM. One example of evaluation processing is for a user to manually evaluate the evaluation results output by the evaluation device and then generate an evaluation value by quantifying the results using rankings or the like. Any other evaluation processing may also be used.

[0054] Examples of items that may be considered (i.e., quantified) when further evaluating the evaluation results output by the evaluation device include the following: (a) throughput of the evaluation device (the number of semiconductor devices evaluated per unit time); (b) evaluation results (semiconductor device measurement success / failure); (c) the amount of noise in the observation image of the semiconductor device acquired by the evaluation device, various histograms, and measurement values ​​(e.g., pattern size, amount of deviation between upper and lower layers); (d) evaluation score indicating whether the measurement was successful or not, such as the image correlation value between an AP image (a unique circuit pattern image near the measurement position) and a captured image; (e) stability of the evaluation (e.g., matching success rate after long-term operation, variance of measurement values); (f) coverage of measurement positions for hotspots; (g) uniqueness of the observation image (AP image), etc.

[0055] Examples of algorithms that can be used to learn pairs of evaluation results (evaluation values) / target values / specification data include, but are not limited to, Temporal Difference Learning (TD learning) and Proximal Policy Optimization (PPO).

[0056] FIG. 10 is a schematic diagram showing another example of a relearning procedure by the reinforcement learning module 114. In the relearning procedure described in FIG. 9, when the target value and evaluation result are fed back to the LLM, they may be converted in advance into a format or value suitable for learning by the LLM. FIG. 10 shows one such example. Evaluation results may be fed back to the LLM without conversion and converted and then fed back to the LLM. FIG. 10 shows an example of such a combination. Alternatively, evaluation results from an evaluation device may be used in combination with evaluation values ​​obtained by further evaluating those evaluation results. Because the evaluation values ​​are quantified, they can be fed back to the LLM as is. However, because the evaluation results themselves have various formats and values, some conversion is desirable. The configuration of FIG. 10 is useful in such cases.

[0057] The target values ​​may be automatically extracted from an input statement input to the LLM. In the example shown in FIG. 9, these values ​​correspond to throughput and measurement accuracy. The extraction process may be performed by the user, or by inputting the input statement to an LLM (which may be the same as semiconductor LLM 232 or a separately configured LLM) that has been pre-configured to extract the target values ​​from the input statement.

[0058] FIG. 11 is a schematic diagram showing another example of a relearning procedure by the reinforcement learning module 114. In the relearning procedure described with reference to FIGS. 9 and 10 , the configuration of FIG. 11 is useful when verifying how the evaluation results of the evaluation device change when a recipe is changed. Specifically, the recipe output by the LLM and the evaluation results (and / or evaluation values) obtained when the evaluation device is operated using the recipe are fed back to the LLM. This allows learning of recipe / evaluation results (and / or evaluation values) / target values ​​as a pair. In other words, it is possible to learn a recipe that produced a certain evaluation result by associating it with that evaluation result. This allows learning of the relationship between the evaluation results and the recipe, thereby enabling the LLM to predict the evaluation results that will be obtained when the recipe is changed.

[0059] FIG. 12 is a schematic diagram showing another example of a relearning procedure by the reinforcement learning module 114. In the relearning procedure described in FIGS. 9 to 11, the LLM may output measurement instructions as an intermediate output. The measurement instructions are further input to the LLM to output a recipe. The LLM that outputs the measurement instructions (large-scale language model 1 in FIG. 12) and the LLM that outputs the recipe (large-scale language model 2 in FIG. 12) may be configured as separate LLMs, or they may be configured as an integrated whole. For convenience of description, these are represented as separate LLMs in FIG. 12.

[0060] Since a recipe is an instruction itself to an evaluation device, it may be written in a format that is not necessarily suitable for a user to understand. In contrast, a measurement instruction is a document that describes the procedure for creating a recipe using natural language strings, making it easy for a user to understand. The configuration shown in Figure 12 is useful in that it allows a user to understand what instructions were given to derive a recipe in the form of a measurement instruction.

[0061] As explained in Figures 9 to 12, the output data output by the evaluation device may be fed back to the LLM as is, or data generated by converting the output data (e.g., text data that is easy for the LLM to learn) may be fed back to the LLM.

[0062] <Summary of Second Embodiment> The computer system 101 (or the computer systems 220 and 230) according to the second embodiment applies reinforcement learning to the semiconductor LLM 232, thereby enabling the semiconductor LLM 232 (or the sentence creation model 115 after reinforcement learning) to output a recipe that enables the evaluation device to achieve a target value when evaluating a semiconductor device. This allows the evaluation device to be operated by a character string, and also allows recipes and other specification data that enable the desired target value to be obtained by a character string instruction.

[0063] For example, if an abnormality is found in an observation image acquired by an evaluation device, it is necessary to revise the measurement instruction manual or recipe. However, creating a measurement instruction manual or recipe can be very costly and time-consuming for the user. According to the second embodiment, even in such a case, if the LLM is instructed to create a recipe or measurement instruction manual using a character string, a recipe or measurement instruction manual that achieves the desired target value can be obtained. This reduces the burden on the user.

[0064] Third Embodiment In the first and second embodiments, an example has been described in which text or design data is input to an LLM and a recipe is obtained as output. The input to the LLM and the output from the LLM are not limited to these, and the LLM can be trained to achieve other inputs / outputs. In a third embodiment of the present disclosure, examples of these inputs / outputs will be described. The configuration of the computer system 101 (or the computer systems 220 and 230) is the same as in the first embodiment.

[0065] 13 shows an example of input to and output from an LLM. The LLM here may be any of the general-purpose LLM 210, the evaluation device LLM 222, the semiconductor LLM 232, and the writing model 115.

[0066] When instructing the LLM to improve a program (such as source code, executable code, or machine learning model, the same applies below), a directive (sentence) requesting the LLM to do so is input, along with the program before the improvement. The LLM is pre-trained to output an improved program that incorporates the instructed improvements to the input program. The program output by the LLM can be used in an evaluation device. For example, the image quality improvement model and pattern enhancement model described in embodiment 1 can be obtained as output from the LLM.

[0067] The LLM can also be trained in advance to generate a program. In this case, a sentence specifying a desired function is input to the LLM, and the LLM outputs a program that implements that function. For example, the LLM can output the image quality improvement model or pattern enhancement model described in the first embodiment.

[0068] The LLM can be trained in advance to output the most suitable recipe for evaluating an input image (an observation image of a shape pattern formed on a semiconductor device). In this case, when the image is input to the LLM, a sentence instructing the LLM to output the most suitable recipe for evaluating the shape pattern of the image is also input. The LLM outputs the optimal recipe according to the instruction.

[0069] As described in the second embodiment, the LLM can be configured to input design data and instructions for a semiconductor device and output a recipe suitable for evaluating the semiconductor device of the design data. In this case, the LLM can output a measurement instruction as an intermediate output in addition to directly outputting the recipe.

[0070] As described in the second embodiment, the LLM can be configured to output a remedy when an abnormality is detected in an observed image. In this case, an observed image containing an abnormality is input to the LLM, and an instruction to output a recipe capable of correcting the abnormality is also input. The LLM is trained in advance to output a corrected recipe capable of correcting the observed image of the portion (circuit pattern) where the abnormality occurred. Furthermore, the cause of the abnormality and the corrections in the recipe may also be output. Although the cause of the abnormality and the corrections in the recipe are not themselves involved in the automatic operation of the evaluation device, this information can be used when explaining the significance and cause of the recipe correction to the user.

[0071] <Regarding Modifications of the Present Disclosure> The present disclosure is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present disclosure, and it is not necessary to include all of the configurations described. Furthermore, a part of one embodiment can be replaced with a configuration of another embodiment. Furthermore, a configuration of another embodiment can be added to a configuration of one embodiment. Furthermore, a part of the configuration of each embodiment can be added to, deleted from, or substituted for a part of the configuration of another embodiment.

[0072] In the above embodiments, learning module 113, reinforcement learning module 114, and sentence creation module 111 (or learning modules 221, 231, reinforcement learning module 233, and sentence creation module 234) can be configured by hardware such as a circuit device that implements these functions, or by software that implements these functions being executed by a computing device such as a CPU (Central Processing Unit).

[0073] In the above embodiments, all or some of the learning module 113, reinforcement learning module 114, and sentence creation module 111 (or learning modules 221, 231, reinforcement learning module 233, and sentence creation module 234) may be configured as an integrated unit, or these modules may be configured separately and then configured to operate in conjunction with each other. In either case, these modules implement a method of operating the evaluation device with character strings via the LLM.

[0074] 101: Computer system 102: Semiconductor evaluation tool 111: Text creation module 113: Supervised / unsupervised learning module 114: Reinforcement learning module

Claims

1. A method executed by a computer, comprising a step of receiving specification data describing operational specifications of an evaluation device from a pre-trained language model by inputting instructions for the evaluation device to evaluate a semiconductor device, the language model having been fine-tuned in advance using at least one of domain knowledge of the semiconductor device or domain knowledge of the evaluation device.

2. The method according to claim 1, wherein the step of receiving the specification data comprises the steps of: inputting an input string describing instructions for the evaluation device to the language model; and receiving a response to the input string from the language model, wherein the language model has been fine-tuned in advance so that, by learning data describing technical knowledge about the semiconductor device and data describing technical knowledge about the evaluation device, the language model receives a request regarding the semiconductor device or the evaluation device as the input string and outputs an answer to the request as the response, and wherein, in the step of inputting the input string, a string requesting the language model to output data describing instructions for the evaluation device is input as the input string, and, in the step of receiving the response, the language model receives the response based on a result of the fine-tuning, thereby receiving from the language model the response reflecting technical knowledge regarding each of the semiconductor device and the evaluation device as the instruction.

3. The method according to claim 1, further comprising a step of performing the fine-tuning on the language model, wherein in the step of performing the fine-tuning, the language model is made to learn at least one of the following technical knowledge regarding the evaluation device by unsupervised learning: an instruction manual for the evaluation device; and a functional specification of the evaluation device.

4. The method according to claim 1, further comprising a step of performing the fine-tuning on the language model, wherein in the step of performing the fine-tuning, data describing the specifications of the image processing performed by the evaluation device is trained in the language model by unsupervised learning, or data describing the results of the image processing performed by the evaluation device is trained in the language model by supervised learning, as technical knowledge related to the evaluation device.

5. The method according to claim 1, further comprising a step of performing the fine-tuning on the language model, wherein in the step of performing the fine-tuning, the language model is made to learn, by unsupervised learning, at least one of the following technical knowledge relating to the semiconductor device: data describing parameter items that impose constraints in a recipe that specifies the operation of the evaluation apparatus and the contents of those constraints; data describing the specifications of the image processing performed by the evaluation apparatus; and measurement instructions describing the procedure for creating an evaluation recipe that defines the processing to be performed by the evaluation apparatus when evaluating the semiconductor device.

6. The method according to claim 1, further comprising a step of performing the fine tuning on the language model, wherein in the step of performing the fine tuning, the language model is made to learn, by supervised learning, at least one of the following technical knowledge regarding the semiconductor device: a measurement instruction document describing a procedure for creating an evaluation recipe defining a process to be performed by the evaluation equipment when evaluating the semiconductor device; data describing at least one of the shape, material, or manufacturing conditions of the semiconductor device; data describing a layout of a pattern shape formed on the semiconductor device; and data describing an interaction between a material of the semiconductor device and a charged particle beam irradiated to the semiconductor device by the evaluation equipment when evaluating the semiconductor device.

7. The method according to claim 6, wherein the supervised learning is performed by at least one of the following: design data of the semiconductor device is used as an input to the language model, and the measurement instructions are used as an output from the language model; the measurement instructions are used as an input to the language model, and the evaluation recipe is used as an output from the language model; imaging conditions when the evaluation device captures an observation image of the semiconductor device and three-dimensional design data of the semiconductor device, or information on the imaging conditions and a material of the semiconductor device, are used as inputs to the language model, and the observation image of the semiconductor device captured by the evaluation device is used as an output from the language model; and filtering conditions for image processing of the observation image of the semiconductor device acquired by the evaluation device is used as an input to the language model, and the observation image is used as an output from the language model.

8. The method according to claim 1, wherein the language model has been previously trained to receive instructions for the evaluation device as a character string and output specification data describing operational specifications of the evaluation device, and the method further comprises the steps of: re-training the language model using at least one of output data output by the evaluation device or text data generated based on the output data as training data; inputting, to the re-trained language model, a character string in a natural language that instructs the evaluation device to evaluate the semiconductor device; receiving the specification data as output from the language model; operating the evaluation device using the specification data received from the language model to evaluate the semiconductor device; and supplying, to the re-training, the output data or the text data obtained by the step of operating the evaluation device.

9. The method according to claim 8, wherein in the step of performing re-learning, the following are input to the language model as the learning data: a target value of the evaluation result obtained by the evaluation device evaluating the semiconductor device, the specification data, and the evaluation result, and the re-learning is performed so that the language model outputs the specification data that brings the evaluation result closer to the target value.

10. The method according to claim 9, wherein in the step of performing re-learning, at least one of the following is input as the target value or evaluation result: a target value for the throughput of the evaluation device or its evaluation result; a target value for a measurement value obtained by the evaluation device measuring the semiconductor device or its evaluation result; a target value for a feature quantity of an observation image of the semiconductor device acquired by the evaluation device or its evaluation result; or a target value for a defect occurrence probability in the semiconductor device or its evaluation result.

11. The method according to claim 9, wherein in the step of performing the relearning, at least one of the following is input as the specification data: an evaluation recipe that defines the processing to be performed by the evaluation device when evaluating the semiconductor device; a prediction result of the defect position in the semiconductor device; a cause of an abnormality and a remedy for the abnormality in the case where the observation image of the semiconductor device acquired by the evaluation device is abnormal; an image quality improvement model that converts the observation image of the semiconductor device output by the evaluation device into an observation image of higher image quality; and a pattern enhancement model that emphasizes a specific shape pattern contained in the observation image of the semiconductor device output by the evaluation device.

12. The method of claim 9, wherein the language model is configured to output, as the specification data, an evaluation recipe that defines the processing to be performed by the evaluation device when evaluating the semiconductor device, and in the step of performing re-learning, the evaluation recipe output by the language model is input as the specification data, and the target value and the evaluation result used as input when the language model outputs the evaluation recipe are input, thereby causing the language model to learn how the evaluation result changes when the evaluation recipe is changed.

13. The method of claim 9, wherein the language model is configured to output, as an intermediate output, measurement instructions describing procedures for creating an evaluation recipe that defines the processing to be performed by the evaluation equipment when evaluating the semiconductor device, and in the step of performing re-learning, the measurement instructions output by the language model as the intermediate output are input again to the language model, thereby performing the re-learning so that the language model outputs the specification data.

14. The method of claim 2, wherein the language model is configured to learn the frequency of occurrence of character strings in each of the data describing technical knowledge about the semiconductor device and the data describing technical knowledge about the evaluation device, and thereby, when the input character string related to either the technical knowledge about the semiconductor device or the technical knowledge about the evaluation device is input, to reflect at least either the technical knowledge about the semiconductor device or the technical knowledge about the evaluation device in the response.

15. The method according to claim 1, wherein the language model is configured to perform at least one of the following: receive a natural language string and a program as input, and output a result of improving the program in accordance with conditions specified by the natural language string; receive a natural language string as input, and output a program that performs processing specified by the natural language string; receive a natural language string and an image of the semiconductor device as input, and output parameters of the evaluation device suitable for the evaluation device to evaluate the shape pattern of the semiconductor device represented by the image while satisfying the conditions specified by the natural language string; receive a natural language string and design data of the semiconductor device as input, and output parameters of the evaluation device suitable for the evaluation device to evaluate the semiconductor device represented by the design data while satisfying the conditions specified by the natural language string, or measurement instructions describing procedures for creating these parameters; receive a natural language string and an image of an abnormal shape pattern of the semiconductor device as input, and output parameters of the evaluation device suitable for the evaluation device to correct the abnormal shape pattern represented by the image while satisfying the conditions specified by the natural language string.

16. A program defining a method executed by a computer, the program being configured to cause the computer to execute a step of receiving specification data describing operational specifications of an evaluation device from a pre-trained language model by inputting instructions for the evaluation device that evaluates a semiconductor device into the language model, the language model having been fine-tuned in advance using at least one of domain knowledge of the semiconductor device and domain knowledge of the evaluation device.

Images