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

Abstract

The purpose of the present disclosure is to provide a technique capable of obtaining, from a language model, an instruction for an evaluation apparatus for evaluating a semiconductor device by inputting a character string into the language model. The method of the present disclosure receives specification data describing an operation specification of an evaluation device from a language model by inputting an instruction for the evaluation device of a semiconductor device to the language model. The language model is finely tuned in advance by at least one of domain knowledge of the semiconductor device and domain knowledge related to the evaluation device.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to a technology of evaluating a semiconductor device by a string operation. BACKGROUND

[0002] In recent years, a technology of accepting an answer to a question item described using natural language by an artificial intelligence (AI) through natural language using a large language model (LLM) has been widely developed. As one example thereof, Non Patent Literature 1 describes a GPT model developed by Open AI.

[0003] PRIOR ART DOCUMENT

[0004] NON-PATENT LITERATURE

[0005] Non Patent Literature 1: Long Ouyang et al., “Training language models to follow instructions with human feedback”, 2022, URL: https: / / arxiv.org / abs / 2203.02155 SUMMARY

[0006] PROBLEMS TO BE SOLVED BY THE INVENTION

[0007] A general large language model like Non Patent Literature 1 learns the frequency of occurrence of a string by repeatedly performing, for various document data, for example, masking a part of a string to estimate the masked part. Thereby, knowledge in various fields (sometimes also referred to as domain knowledge) is pre-learned through a string, and when a question related to the field is given, an answer can be output based on the pre-learned knowledge.

[0008] Such a large language model can learn knowledge in various fields, but there are cases where knowledge related to a specific field is insufficient. In addition, such a large language model is useful in a process of giving a question as input to obtain an answer to the question, but in a use such as obtaining an operation instruction for a specific device, it is considered not necessarily optimal. Non Patent Literature 1 describes a training method of a large language model, but does not particularly consider a specific method related to such a use.

[0009] The present disclosure is achieved in view of the above-described problems, and an object thereof is to provide a technology capable of obtaining an instruction for an evaluation device for evaluating a semiconductor device, and the like from a language model by a string input to the language model.

[0010] MEANS FOR SOLVING THE PROBLEMS

[0011] The method of the present disclosure receives specification data describing the action specification of the evaluation device for evaluating the semiconductor device from a language model by inputting an instruction for the evaluation device to the language model. The language model is fine-tuned in advance by at least one of domain knowledge of the semiconductor device or domain knowledge related to the evaluation device.

[0012] As another way for achieving the above object, specification data describing the action specification of a second evaluation device is received from a language model by inputting specification data describing the action specification of a first evaluation device for evaluating the semiconductor device and an instruction for the second evaluation device to the language model. The language model is fine-tuned by comparison information of output information of the first evaluation device and output information of the second evaluation device.

[0013] As another way for achieving the above object, data related to a defect position is received from a language model by inputting information related to a layout of the semiconductor device and instruction information predicting the position of a defect generated on the semiconductor device to the language model. The language model is fine-tuned by comparison information of data related to the defect position output from the language model and data related to the correct position of the defect.

[0014] As still another way for achieving the above object, data related to the cause of an anomaly is received by inputting an investigation instruction of an anomaly of a manufacturing process of the semiconductor device to the language model. The language model is fine-tuned by article data related to the cause and association data of evidence of the cause.

[0015] Effects of Invention

[0016] According to the method of the present disclosure, an instruction for an evaluation device for evaluating a semiconductor device or the like can be obtained from a language model by inputting a character string to the language model. Other problems, structures, advantages, and the like of the present disclosure become clear through the description of the following embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a block diagram of a semiconductor evaluation system 100 of Embodiment 1.

[0018] Figure 2 represents an example of an article production system 200 using a general LLM 210 configured based on a large-scale language model.

[0019] Figure 3 represents an example of domain knowledge learned by fine-tuning.

[0020] Figure 4 represents a distinction between unsupervised learning and supervised learning.

[0021] Figure 5A is a schematic diagram representing a process in which the learning module 113 implements unsupervised learning.

[0022] Figure 5B is a schematic diagram representing another process in which the learning module 113 implements unsupervised learning.

[0023] Figure 6 is a schematic diagram representing a process in which the learning module 113 implements supervised learning.

[0024] Figure 7 represents a change in the combination of input / output / correct answer / loss function in supervised learning.

[0025] Figure 8 is a schematic diagram representing a process in which the reinforcement learning module 114 implements reinforcement learning.

[0026] Figure 9 is a schematic diagram representing an example of the relearning step of the reinforcement learning module 114.

[0027] Figure 10 is a schematic diagram representing another example of the relearning step of the reinforcement learning module 114.

[0028] Figure 11 is a schematic diagram representing another example of the relearning step of the reinforcement learning module 114.

[0029] Figure 12 is a schematic diagram representing another example of the relearning step of the reinforcement learning module 114.

[0030] Figure 13 represents an example of input to the LLM and output from the LLM.

[0031] Figure 14 is a schematic diagram representing a process in which supervised learning is implemented based on a dataset containing article data and SEM images, etc.

[0032] Figure 15 represents a step of generating a process for reducing mechanical errors between a plurality of measurement devices.

[0033] Figure 16 represents a step of adjusting device conditions of other measurement devices so as to reduce mechanical errors between a reference device of the measurement device and the other measurement devices.

[0034] Figure 17 is a schematic diagram representing a step of adjusting device conditions of other measurement devices so as to reduce mechanical errors between a reference device of the measurement device and the other measurement devices (including manual operation).

[0035] Figure 18is a diagram showing an example of a GUI screen provided with a display column of an article input to an LLM and a display column of an adjusted content of an LLM output.

[0036] Figure 19 is a diagram showing an example of a GUI screen provided with a display column of an article input to an LLM and a display column of an adjusted content (parameters and articles) of an LLM output.

[0037] Figure 20 represents an outline of an LLM configured to output a recipe for causing an apparatus to act and a measurement instruction.

[0038] Figure 21 is a diagram showing a process of predicting a defect position using an LLM.

[0039] Figure 22 represents a process of predicting a defect position using an LLM and generating a recipe based on the predicted defect position to an input to the LLM.

[0040] Figure 23 represents a step of using an LLM to derive a cause of an abnormal image and evidence.

[0041] Figure 24 represents a step of fine-tuning a cause determination model of an abnormal image.

[0042] Figure 25 represents a step of fine-tuning a cause analysis model of an abnormal image.

[0043] Figure 26 represents a step of reinforcement learning of a cause determination model and a cause analysis model of an abnormal image. DETAILED DESCRIPTION

[0044] <Embodiment 1: System Structure>

[0045] Figure 1 is a block diagram of a semiconductor evaluation system 100 of Embodiment 1 of the present disclosure. The semiconductor evaluation system 100 is a system that performs an evaluation process such as inspection, measurement, and the like on a semiconductor device (a sample). The semiconductor evaluation system 100 is provided with 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 through a communication bus 105.

[0046] The computer system 101 includes one or more processors 106 (including at least one of a CPU and a GPU), a memory 107, and a user interface 108.

[0047] The user interface 108 includes a display, which is not shown, one or more input devices. The display is capable of displaying image data output from the semiconductor evaluation tool 102, information output from the processor 106, and the like.

[0048] The semiconductor evaluation tool 102 is, for example, a CD-SEM (Critical Dimension-Scanning Electron Microscope) that generates an image, a brightness signal waveform based on detection of secondary electrons, backscattered electrons emitted from a sample by irradiating the sample with an electron beam. The CD-SEM measures the size, two-dimensional shape of a pattern formed on a semiconductor device based on the brightness signal waveform. In addition to this, the semiconductor evaluation tool 102 can also be an inspection device including a charged particle beam such as a SEM for defect review, an optical inspection device.

[0049] The design data storage medium 103 stores design data describing the physical circuit structure of a semiconductor device, for example, in the form of a GDS II or OASIS file.

[0050] The simulator 104 generates an observation image of a sample by simulation, or estimates a signal profile representing the intensity distribution of secondary electrons, backscattered electrons. Specifically, the generation process of secondary electrons generated by a beam incident to a sample, backscattered electrons emitted from the sample is reproduced using random numbers, and a calculation is repeatedly performed, whereby the emission angle, energy, and the like of the secondary electrons, backscattered electrons are calculated. Monte Carlo simulation or the like is used in the calculation.

[0051] A processor, which is not shown, and a storage medium storing data and programs required for simulation are built into the simulator 104. The simulator 104 is configured to read out required information (example: pattern shape included in a semiconductor device, material composition of a semiconductor device) from the design data storage medium 103, and estimate an electron microscope image (in the case where the semiconductor evaluation tool 102 is an electron microscope) based on configuration information of the semiconductor evaluation tool 102, such as the position of a detector (position relative to the position of a beam irradiation site) in the semiconductor evaluation tool 102 described later, control conditions of a measurement device including the irradiation conditions of a beam, and the like, which are stored in advance.

[0052] The simulation implemented by the simulator 104 not only uses a trajectory calculation method such as a Monte Carlo simulation, but also can use a model defined in association with layout data (including three-dimensional data such as film thickness), coefficients that vary depending on materials, device conditions of an electron microscope, brightness information, and the like to derive an image, brightness information, a profile, and the like. In addition, it is also possible to estimate image data and the like by inputting known information to a learner that has implemented learning using a data set of device conditions of an electron microscope, material information, 3D layout information, brightness information (image data, a profile, and the like) as training data.

[0053] The memory 107 is constituted by a volatile and / or non-volatile 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 performs processing procedures related to various measurements or inspections described later; (b) a learning module 110 that generates a learning model based on input of a large-scale language model described later, article data, and information output from the semiconductor evaluation tool 102, and the like; (c) an article production module 111 that produces article data and the like using a model obtained by implementing learning 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.

[0054] The learning module 110 has the following 2 modules: (a) a supervised / unsupervised learning module 113; and (b) a reinforcement learning module 114. The article production module 111 has an article production model 115. Details of these sub-modules and models will be described later.

[0055] In addition to the above, the memory 107 (for example, a non-transitory computer-readable storage medium) can also store, for example: (a) an action program executed by one or more processors included in the semiconductor evaluation tool 102; and (b) a process generation module that generates the action program (process) based on parameters and the like input from the user interface 108.

[0056] The process generation module is configured to perform process generation, correction processing of a generated process, and the like, for example, based on output of the article production module 111 described later. The process generation module is configured to generate a process based on design data, a simulation image generated by the simulator 104 based on input of the design data, parameters input from the user interface 108, and the like, for example.

[0057] Figure 2 An example of an article production system 200 using a general-purpose LLM 210 configured based on a large language model (LLM) is shown. Figure 2An article creation system 200 for measuring or inspecting a semiconductor device is exemplified. The article creation system 200 generates an LLM suitable for measurement or inspection of a semiconductor device based on a general-purpose LLM 210. The LLM for measurement or inspection of a semiconductor device is premised on the general-purpose LLM 210 that has learned in advance through a large amount of language data.

[0058] The article creation system 200 includes: (a) a device vendor-side computer system 220 that mainly performs fine-tuning of the general-purpose LLM 210 by evaluating domain knowledge of a device; and (b) a device user-side computer system 230 that performs further fine-tuning of the evaluation device LLM 222 provided by the device vendor by domain knowledge of a semiconductor device (e.g., domain knowledge related to manufacturing conditions in a manufacturing process of a semiconductor wafer).

[0059] The device vendor-side computer system 220 includes an evaluation device LLM learning module 221. The evaluation device LLM learning module 221 performs fine-tuning of the general-purpose LLM 210 mainly by evaluating domain knowledge of a device (e.g., domain knowledge related to manufacturing conditions in a manufacturing process of a semiconductor wafer) of the semiconductor evaluation tool 102. Figure 1 By this, the general-purpose LLM 210 further learns the domain knowledge of the evaluation device, and is restructured into the evaluation device LLM 222.

[0060] The device user-side computer system 230 includes a semiconductor LLM learning module 231. The semiconductor LLM learning module 231 performs further fine-tuning of the evaluation device LLM 222 by domain knowledge of a semiconductor device. By this, the evaluation device LLM 222 further learns the domain knowledge of the semiconductor device, and is restructured into the semiconductor LLM 232.

[0061] The device user-side computer system 230 further includes a reinforcement learning module 233. The reinforcement learning module 233 performs reinforcement learning described later on the semiconductor LLM 232, thereby performing relearning on the semiconductor LLM 232 so that the semiconductor LLM 232 can output an evaluation process in which the evaluation device can achieve a desired target value. By this, the semiconductor LLM 232 is restructured into an article creation model.

[0062] The device user-side computer system 230 further includes an article creation module 234 that includes an article creation model obtained by performing the reinforcement learning described above. The article creation module 234 uses the article creation model 115 to output an output string as a response to an input string.

[0063] In Figure 2In the illustrated system configuration, the device vendor side computer system 220 performs fine tuning by the domain knowledge of the evaluation device such as an electron microscope, and the device user side computer system 230 performs fine tuning by the domain knowledge of the semiconductor device. However, the computer system that performs fine tuning is not limited to this. For example, the device vendor side computer system 220 can perform fine tuning by the domain knowledge related to the semiconductor device, and the device user side computer system 230 can perform fine tuning by the domain knowledge of the evaluation device. Also, with respect to reinforcement learning, it can be performed in any computer system. Furthermore, the device vendor side computer system 220 and the device user side computer system 230 can be integrally configured.

[0064] The device vendor side computer system 220 and the device user side computer system 230 can be respectively configured by the computer system 101 explained in the Figure 1 In the embodiment, the computer system 101 explained in the Figure 1 is an example in which these modules are concentrated in a single computer system 101.

[0065] <Embodiment 1: Fine tuning>

[0066] The general-purpose LLM 210 is configured in advance so that by pre-learning various domain knowledge, an answer string reflecting the domain knowledge can be output for an input question string. The domain knowledge that can be learned is various. For example, in a case where a specific book is learned, an abstract of the book can be answered. Or, in a case where a specific programming language is learned, if a desired function is described in natural language and input to the general-purpose LLM 210, a source code in which the function is installed can be output.

[0067] The general-purpose LLM 210 has such versatility, but with respect to knowledge of a field that is not pre-learned, a useful answer cannot be output, and only a simple general degree of answer can be made. Therefore, in order for the general-purpose LLM 210 to learn knowledge of a specific field, adjustment called fine tuning needs to be performed. The purpose of the present disclosure is to input an action required for a semiconductor device evaluation device (the semiconductor evaluation tool 102 in the Figure 1 ) to the LLM, and to obtain an action specification (evaluation process) that can achieve the requirement as an answer from the LLM. In other words, the purpose is to operate the semiconductor device evaluation device by natural language or the like. Therefore, in the embodiment 1, first, the general-purpose LLM 210 is caused to learn the domain knowledge of the evaluation device and the domain knowledge of the semiconductor device by fine tuning.

[0068] Figure 3 Examples of domain knowledge representing learning by fine tuning. Figure 3 The "evaluation device" row of the "domain knowledge" column in Table 1 is an example of domain knowledge related to an evaluation device learned by the evaluation device LLM with the learning module 221. Figure 3 The other rows of the "domain knowledge" column in Table 1 are examples of domain knowledge related to a semiconductor device learned by the semiconductor LLM with the learning module 231.

[0069] As for domain knowledge related to an evaluation device, fine tuning can be performed by causing the general-purpose LLM 210 to learn document data such as a user's manual of the evaluation device, a specification sheet of the functions of the evaluation device, and the like. By this, knowledge about the operation method (use method) of the evaluation device can be learned. Also, as for the specification of image processing performed by the evaluation device or the result of performing image processing, learning as domain knowledge of the evaluation device is possible. As for the specification of image processing, learning is possible from a specification sheet, a textbook, or the like that describes the contents of the image processing. As for the result of image processing, learning is possible by learning the change in the observed image before and after actually performing image processing.

[0070] A recipe is data that describes an action specification that specifies an evaluation process performed by an evaluation device. Depending on the user of the evaluation device, there are parameters that are prohibited from being used within the recipe or are subject to restrictions within the recipe. Such user-specific restrictions (parameter items subject to restrictions and the contents of the restrictions) are referred to as local rules. Local rules are rules specific to a user, and thus it is more appropriate to learn them as domain knowledge specific to a semiconductor device being evaluated by the user than to learn them as general domain knowledge of the evaluation device. Therefore, local rules are learned as domain knowledge of a semiconductor device.

[0071] Since a recipe describes instructions for an evaluation device as is, it looks like a list of parameters from the user's perspective, and the readability is not necessarily high. A measurement instruction is a document that describes the manufacturing steps of a recipe in natural language, and it is sometimes easier for a user to read a measurement instruction than a recipe. Therefore, a measurement instruction is also learned as domain knowledge of a semiconductor device.

[0072] The specification of image processing performed by an evaluation device can also be learned as domain knowledge of a semiconductor device. The learning method is the same as when learning domain knowledge of an evaluation device.

[0073] A user who manufactures a semiconductor device sometimes has a simulator that simulates a manufacturing process of the semiconductor device. The simulator uses design data that describes a three-dimensional shape, a material, and manufacturing process information (example: manufacturing conditions of a semiconductor manufacturing apparatus) for a measurement target sample (semiconductor device). Also, design data that describes a two-dimensional shape of a shape pattern formed on the semiconductor device is sometimes used. These design data can be learned as domain knowledge of the semiconductor device.

[0074] In a case where the evaluation device performs evaluation by irradiating an electron beam (or other arbitrary charged particle beam) to a sample (semiconductor device), data that describes an interaction between the electron beam and a material of the sample can be learned as domain knowledge of the semiconductor device.

[0075] Figure 4 A distinction between unsupervised learning and supervised learning is indicated. A user manual or a function specification of the evaluation device can be learned by unsupervised learning. A specification of image processing can be learned by unsupervised learning. A result of image processing can be learned by supervised learning. A measurement instruction can be learned using either one of unsupervised learning and supervised learning. A local rule can be learned by unsupervised learning. Design data can be learned by supervised learning. An interaction between an electron beam and a sample can be learned by supervised learning.

[0076] Figure 5A is a schematic diagram indicating a process in which the learning module 113 performs unsupervised learning. The unsupervised learning can be performed by, for example, Masked LM. Masked LM is a method of learning a frequency of occurrence of a string in a document by repeatedly masking a part of the document and estimating the masked part, thereby learning knowledge described in the document. Learning is performed by feeding back a difference (loss) between an estimation result of the masked part and a correct string to the LLM. As an example of a loss function, there are, for example, cross-entropy error, but not limited thereto.

[0077] Figure 5B is a schematic diagram indicating another process in which the learning module 113 performs unsupervised learning. In addition to Masked LM, unsupervised learning can be performed by NSP (Next Sentence Prediction). Other arbitrary methods can also be used. In NSP, it is repeatedly predicted whether two articles are connected in terms of meaning. Thereby, a correct connection between articles can be learned. The loss function is the same as Masked LM.

[0078] Figure 6is a schematic diagram showing a process in which the learning module 113 implements supervised learning. Here, an example of design data described in natural language strings for a request to the evaluation device and the LLM input is shown. In this example, the LLM output describes a measurement instruction sheet for a manufacturing step of an evaluation process for implementing the requested step. As the correct data, a measurement instruction sheet that can actually satisfy the request is provided. By feeding back the loss between the measurement instruction sheet of the LLM output and the correct measurement instruction sheet to the LLM, supervised learning can be implemented. Figure 3

[0079] Figure 7 shows a change in the combination of input / output / correct answer / loss function in supervised learning. In the case of design data as input, a measurement instruction sheet as output and correct answer, the loss function is, for example, cross-entropy error. In the case of a measurement instruction sheet as input, a process as output and correct answer, the loss function is, for example, cross-entropy error. In the case of learning the interaction between an electron beam and a sample, the input is the shooting conditions (voltage of each part, current of each part, etc.) and three-dimensional design data when the observation image of the semiconductor device is shot by the evaluation device, or the shooting conditions and information related to the material of the semiconductor device, the output is the observation image, and the loss function is, for example, mean square error. In the case of learning the result of image processing, the input is a process (in particular, a filter condition in image processing implemented by the evaluation device), the output is an observation image, and the loss function is, for example, mean square error.

[0080] The learning module 113 (or the evaluation device LLM learning module 221 and the semiconductor LLM learning module 231) implements unsupervised learning / supervised learning according to the above steps, whereby the general-purpose LLM 210 can learn the domain knowledge of the evaluation device and the domain knowledge of the semiconductor device by fine-tuning. Thus, the semiconductor LLM 232 can output answers that reflect these domain knowledges when given an input sentence related to the evaluation device or the semiconductor device.

[0081] <Implementation Example 1>

[0082] The computer system 101 of the embodiment 1 (or the device supplier side computer system 220 and the device user side computer system 230) makes the general-purpose LLM 210 learn the domain knowledge of the evaluation device and the domain knowledge of the semiconductor device by fine-tuning. Thus, the semiconductor LLM 232 can output answers that reflect these domain knowledges when given an input sentence related to the evaluation device or the semiconductor device. In the fine-tuning, a measurement instruction sheet or a process is learned, so the semiconductor LLM 232 can output them as output (refer to Figure 7 ​That is, if a requirement for the evaluation device is input into the semiconductor LLM232 via a string, a response corresponding to the requirement can be obtained from the semiconductor LLM232 as a measurement instruction or process. Thus, instructions for the evaluation device can be given via the LLM through a string. In other words, the evaluation device can be operated via a string.

[0083] <Implementation Method 2>

[0084] In Embodiment 1, a structural example was described in which instructions for the evaluation device could be assigned via a string through an LLM. According to Embodiment 1, the purpose of operating the evaluation device via a string can be largely achieved. However, depending on the composition of the parameters and their values ​​described in the process of the evaluation device, it is considered that the operation of the evaluation device may not be as desired (e.g., the expected target value cannot be achieved). Therefore, in Embodiment 2 of this disclosure, a structural example was described in which the LLM can output a more desirable process based on the structure of Embodiment 1. The structure of the computer system 101 (or the device supplier-side computer system 220 and the device user-side computer system 230) is the same as in Embodiment 1.

[0085] Figure 8 This is a schematic diagram illustrating the reinforcement learning process implemented by reinforcement learning module 114. Reinforcement learning module 114 (or reinforcement learning module 233) uses the process data output by the evaluation device via an LLM (semiconductor LLM 232 in Embodiment 1). Figure 8 The evaluation results of the semiconductor equipment (based on the specifications data) are fed back to the LLM along with the target values, allowing the LLM to relearn and output a process that achieves the target values. Therefore, it is believed that the LLM can output an ideal process that achieves the target values. The same consideration can be made when the LLM outputs a measurement instruction.

[0086] The relearning steps of reinforcement learning module 114 are roughly as follows: At least one of the output data from the evaluation device (or text data generated based on it, specific examples described later) is used as learning data to perform LLM relearning. For the relearned LLM, a string describing the instructions for the evaluation device is input in natural language. As output from the LLM, data specifying the actions of the evaluation device (called specification data, for example, process data) is obtained. The evaluation device is activated using the specification data, resulting in output data from the evaluation device. The output data is used as learning data to perform relearning. By repeating the above process, relearning can be performed.

[0087] In implementing the relearning, the target value of each evaluation item is input to the LLM together with the evaluation result of the semiconductor device by the evaluation device. Also, the specification data used when the evaluation result is obtained is input together. By this, it is possible to feedback the difference between the target value and the evaluation result, and to learn the specification data that is the cause of the evaluation result together. Therefore, it is possible to perform the relearning so that the LLM outputs the specification data that makes the evaluation result close to the target value.

[0088] As the item evaluated by the evaluation device (i.e., the item that becomes the object of the evaluation result and its target value), for example, consider the following items: (a) the throughput of the evaluation device; (b) the value measured by the evaluation device (example: pattern size); (c) the characteristic amount of the observation image of the semiconductor device obtained by the evaluation device (example: noise amount, various histograms); (d) the defect generation probability of each part of the semiconductor device. The evaluation device outputs the evaluation value with respect to each of these evaluation items. They are fed back to the LLM together with the target value of each evaluation item.

[0089] As the specification data output by the LLM, for example, consider the following data: (a) data that specifies the processing (i.e., the process) performed when the evaluation device evaluates the semiconductor device; (b) the prediction result of the defect position in the semiconductor device (prediction result of whether there is a defect at the measurement object position); (c) the prediction of the abnormality cause when the observation image of the semiconductor device obtained by the evaluation device is abnormal and its improvement countermeasure; (d) a quality improvement model that transforms the observation image into an image of higher quality; (e) a pattern emphasis model for emphasizing a specific shape pattern contained in the observation image. The LLM performs learning in advance and relearning so that the above specification data is output.

[0090] In the case where the LLM outputs the abnormality cause and its improvement countermeasure, for example, the process changed in order to improve the abnormality and the cause of the abnormality of the process before the improvement can be output together. By this, the user can understand which part of the process before the improvement is problematic. The quality improvement model and the pattern emphasis model can be constituted by data such as mathematical models installed with these functions, or other arbitrary data and / or programs such as artificial intelligence installed with these functions.

[0091] Figure 9is a schematic diagram showing an example of the relearning step of the reinforcement learning module 114. In this example, the LLM is constructed in advance so that it receives, as input, an article describing an instruction to the evaluation device and design data of the semiconductor device, and outputs, as specification data, a recipe of the operation of the evaluation device. Since there is a case where the design data is not described in a form suitable for the LLM to receive, the design data can be made to pass through a transformation module that transforms it into a form that the LLM can easily interpret. As for the input sentence, it can also be input to the LLM after a more appropriate transformation (example: transformation of a natural language string into a numerical vector) is performed.

[0092] The evaluation result of the semiconductor device output by the evaluation device can be fed back to the LLM directly, or it can be fed back to the LLM after the evaluation result is further subjected to evaluation processing to transform the output from the evaluation device into an evaluation value. As an example of the evaluation processing, it is considered that the user evaluates the evaluation result output by the evaluation device by manual processing, and numerically values the result by ranking or the like, thereby generating an evaluation value. Any other evaluation processing can also be used.

[0093] As items to be investigated (i.e., numerically valued) when further evaluating the evaluation result output from the evaluation device, for example, the following items are considered: (a) throughput of the evaluation device (number of evaluations of the semiconductor device per unit time); (b) evaluation result (success / failure of measurement of the semiconductor device); (c) noise amount in the observation image of the semiconductor device acquired by the evaluation device, various histograms, measured values (example: pattern size, offset between upper and lower layers); (d) evaluation score indicating whether the measurement was successful, for example, image correlation value of an AP image (image of a unique circuit pattern in the vicinity of the measurement position) and a captured image; (e) stability of evaluation (example: matching success rate at a long time operation, variance of measured values); (f) coverage of the measurement position with respect to hot spots; (g) specificity of the observation image (AP image), and the like.

[0094] As an algorithm for learning the pair of the evaluation result (evaluation value) / target value / specification data, for example, temporal difference learning (TD learning), PPO (Proximal Policy Optimization), or the like can be used, but is not limited thereto.

[0095] Figure 10 is a schematic diagram showing another example of the relearning step of the reinforcement learning module 114. In the relearning step described in Figure 9 In the relearning step described in, when the target value and the evaluation result are fed back to the LLM, they can also be transformed in advance into a form or value suitable for learning by the LLM. Figure 10is shown. As for the evaluation result, the evaluation result fed back to the LLM without transformation and the evaluation result fed back to the LLM after transformation can be combined. In Figure 10 an example of combination is shown. Alternatively, the evaluation result from the evaluation device and the evaluation value obtained by further evaluating the evaluation result can be combined. The evaluation value is numerically expressed, and thus can be directly fed back to the LLM, but the evaluation result itself has various formats and values, and thus it is preferable to implement some kind of transformation. Figure 10 the structure of

[0096] The target value can be automatically extracted from the input sentence input to the LLM, for example. In Figure 9 the example shown, the throughput and the measurement accuracy correspond to the target value. The extraction process can be implemented by the user himself / herself, or can be implemented by inputting the input sentence to an LLM (which can be the same LLM as the semiconductor LLM 232, or can be another LLM) preliminarily configured to extract the target value from the input sentence.

[0097] Figure 11 is a schematic diagram showing another example of the relearning step of the reinforcement learning module 114. In Figure 9~Figure 10 the relearning step explained in Figure 11 the structure of

[0098] Figure 12 is a schematic diagram showing still another example of the relearning step of the reinforcement learning module 114. In Figure 9~Figure 11 the relearning step explained in Figure 12 , the LLM can output a measurement instruction as an intermediate output. By further inputting the measurement instruction to the LLM, a process is output. The LLM (the large-scale language model 1 in Figure 12 ) that outputs the measurement instruction and the LLM (the large-scale language model 2 in Figure 12 ) that outputs the process can be configured as separate LLMs, or can be integrally configured. In , for convenience of description, they are expressed as separate LLMs.

[0099] Since the process is directed at the evaluation device itself, it is sometimes described in a form that is not necessarily suitable for the user to understand its content. In contrast, the measurement instruction is a document that describes the manufacturing steps of the process by a natural language string, and thus the user can easily understand its content. Figure 12 The illustrated structure is useful in that the user can understand in the form of the measurement instruction what kind of instruction was made in order to derive the process.

[0100] As Figure 9~Figure 12 As explained in the above, the output data output by the evaluation device can be fed back to the LLM, or data (example: text data that the LLM easily learns) generated by performing a transformation on the output data can be fed back to the LLM.

[0101] <Embodiment 2: Summary>

[0102] The computer system 101 (or the device vendor side computer system 220 and the device user side computer system 230) of Embodiment 2 causes the semiconductor LLM 232 (or the article manufacturing model 115 after reinforcement learning) to be able to output a process that can achieve a target value at which the evaluation device evaluates the semiconductor device, by performing reinforcement learning on the semiconductor LLM 232. Thereby, in addition to being able to evaluate the evaluation device by string manipulation, it is also possible to obtain a process and other specification data that can achieve a desired target value by string instruction.

[0103] For example, in a case where there is an abnormality in the observation image acquired by the evaluation device, it is necessary to correct the measurement instruction or the process. However, making the measurement instruction and the process sometimes requires a large cost and time for the user. According to Embodiment 2, even in such a case, as long as the LLM is instructed to make a process and a measurement instruction by a string, it is possible to obtain a process and a measurement instruction that achieve a desired target value. Thereby, it is possible to reduce the burden on the user.

[0104] <Embodiment 3>

[0105] In Embodiments 1 to 2, an example in which an article or design data is input to the LLM, and a process is acquired as output is explained. The input to the LLM and the output from the LLM are not limited to this, and the LLM can be caused to learn to achieve other input / output. In Embodiment 3 of the present disclosure, an example of such input / output is explained. The structure of the computer system 101 (or the device vendor side computer system 220 and the device user side computer system 230) is the same as in Embodiment 1.

[0106] Figure 13Examples of input to the LLM and output from the LLM. The LLM here can be any one of the general-purpose LLM 210, the evaluation device LLM 222, the semiconductor LLM 232, and the article creation model 115.

[0107] In a case where the LLM is instructed to improve a program (source code, executable code, machine learning model, etc., hereinafter the same), an instruction sentence (article) for requesting the meaning is input to the LLM, and the program before the improvement is input. The LLM is caused to learn in advance so that a program after the improvement in which the instructed improvement is implemented on the input program is output. The program output from the LLM can be used in the evaluation device. For example, the image quality improvement model or the pattern emphasis model explained in Embodiment 1 can be acquired as the output from the LLM.

[0108] The LLM can also be caused to learn in advance so that a program is generated. At this time, an article that specifies a desired function is input to the LLM, and the LLM outputs a program in which the function is installed. For example, the image quality improvement model or the pattern emphasis model explained in Embodiment 1 can be output.

[0109] The LLM can be caused to learn in advance so that a process most suitable for evaluating an input image (observation image obtained by photographing a shape pattern formed on a semiconductor device) is output. In this case, the image is input to the LLM, and an article for instructing output of a process most suitable for evaluating the shape pattern of the image is input together. The LLM outputs the optimal process in accordance with the instruction.

[0110] As explained in Embodiment 2, the LLM can also be configured to output a process of a semiconductor device suitable for evaluating design data of the semiconductor device by inputting the design data and an instruction sentence to the LLM. In this case, in addition to the LLM directly outputting the process, a measurement instruction can also be output as an intermediate output.

[0111] As explained in Embodiment 2, the LLM can also be configured to output an improvement countermeasure when an observation image is abnormal. In this case, an observation image including an abnormality is input to the LLM, and an instruction sentence for instructing output of a process capable of improving the abnormality is input together. The LLM is caused to learn in advance so that a corrected process capable of improving the observation image of the portion (circuit pattern) in which the abnormality has occurred is output. Also, the cause of the abnormality and the correction site within the process can be output together. As for the cause of the abnormality and the correction site within the process, they do not themselves participate in the automatic operation of the evaluation device, but in a case where the meaning, cause, or the like of the process correction is explained to the user, these information can be used.

[0112] <Embodiment 4>

[0113] In Embodiment 4 of the present disclosure, another example of fine tuning of LLM is explained. Figure 6 Figure 14 A process of associating information other than article data and design data information, which is related to a semiconductor manufacturing process, with the information related to the request given to the evaluation device, is exemplified as a process of learning data of LLM. Fine tuning is preferably performed by associating article data and at least one data other than the article data.

[0114] The information related to the request given to the evaluation device is, for example, article data. The information related to the semiconductor manufacturing process other than design data information is, for example, a SEM image obtained by a measurement tool such as a CD-SEM or a defect inspection tool of SEM type. In addition, it can be an image obtained from an optical inspection device. It can be used for adjustment for making a process of the optical inspection device, or for setting a device condition of SEM so as to be close to an image obtained by the optical inspection device.

[0115] The SEM image or the like is obtained in a semiconductor manufacturing process for evaluating completion of the device. For example, image data can be input together with article data such as "please set a shooting condition so that contrast between a pattern and a substrate is greater than an input image". At this time, the LLM can be adjusted in association with action program (process) information of the evaluation device that has obtained the image.

[0116] Further, as the information related to the semiconductor process, there is process information. A semiconductor device is manufactured through a plurality of manufacturing processes, but measurement conditions of devices manufactured in the same manufacturing process are similar even if generations or types of devices are changed. Therefore, article data such as "please make a measurement instruction related to a shooting condition frequently used in an input process" is considered to be input. At this time, if the LLM is learned or adjusted in advance in units of manufacturing processes based on process information of past generations or the like, a measurement instruction or the like including a shooting condition (device condition of a measurement device) suitable for the manufacturing process can be received.

[0117] Further, as other information related to the semiconductor process, there is device information. The device information is, for example, a device name or identification information of a device. If types of devices are the same, there are many cases where measurement conditions or the like for measuring the devices are the same or similar, and therefore by performing fine tuning or the like based on the device information, a measurement instruction or the like including a shooting condition suitable for the device can be received.

[0118] ​As explained above, even if the devices are of different generations, if the manufacturing processes, types are the same, the setting conditions of the measurement processes used are similar in many cases, and therefore learning and adjustment of the LLM can be performed by the data set of the process information, equipment information, and / or process information, article data, a loss function is calculated between the measurement instruction sheet and the correct measurement instruction sheet output by the LLM on which adjustment and the like have been performed, and the LLM is supervised-learned by back propagation to make the difference small.

[0119] <Embodiment 5>

[0120] In Embodiment 5 of the present disclosure, a process setting method for reducing mechanical errors among a plurality of CD-SEM or the like measurement tools (semiconductor evaluation tools) is explained. The recent semiconductor manufacturing process requires several hundred processes, and therefore there are sometimes multiple production lines. In this case, the manufacturing conditions need to be controlled to avoid manufacturing errors among the production lines and among the devices. The measurement devices also need to suppress mechanical errors among the devices. More specifically, when the measurement result of the pattern size in the measurement device A (reference device) is 20 nm, it is preferable to adjust the device conditions so that the same measurement result is obtained in the measurement device B that measures the same pattern. With the recent miniaturization of semiconductor device patterns, the required performance value of the mechanical error is becoming smaller year by year.

[0121] In Embodiment 5, a method and system for making a process in which the mechanical error is reduced, a correction process, or proposing a device adjustment condition using the LLM are explained. By using the LLM, the adjustment work required so far for mechanical error reduction can be reduced or eliminated, and as a result, the cost and time required for the work can be reduced.

[0122] Figure 15 A method of using the LLM suitable for suppressing the mechanical error of the processes of the reference device A (CD-SEM_A) (first evaluation device), the measurement device B (CD-SEM_B) (second evaluation device), and the measurement device C (CD-SEM_C) is exemplified. In Figure 15 the LLM is input with an article including a required specification, and a process (that satisfies the required specification) is output.

[0123] Figure 15 An example in which an article "Please make a process of the following specification. However, please adjust the existing process so that the length measurement error becomes OO nm. Required specification: difference in length measurement value among devices OO nm or less, throughput OO or less" is input to the LLM as a required specification is shown. In addition, in Figure 15In the example of FIG. 6, an example in which the recipe of the reference device A is input to the LLM together with the article data is shown, but it can not be input in the absence of an existing recipe. In addition, the comparison result (for example, the comparison result of the images obtained by the reference device and the other devices) can be input to the LLM together with the article data to adjust the LLM.

[0124] Next, the plurality of devices (the reference device A, the measurement device B, the measurement device C) read in the recipe output from the LLM and perform measurement. In the case where the plurality of devices are CD-SEMs, it is possible to output image data, a size measurement result of a pattern, and the like from each CD-SEM. The computer system 101 or other arithmetic device outputs the result by comparing the output results of the CD-SEMs. The computer system 101 or the like can also be configured to be capable of transmitting and receiving information through a communication medium or the like so as to be capable of receiving the output of each CD-SEM.

[0125] The recipe is an action program for operating the CD-SEM, and is set with a device condition of the electron microscope, an image processing condition of the output image, and the like. The recipe is preferably set so that the mechanical error of each CD-SEM is small. Therefore, in Embodiment 5, the optimization of the recipe is performed in the following steps.

[0126] First, learning is performed by inputting to the LLM, together with the recipe information of the reference device A, article data of "Please create a recipe of the following specifications. However, please adjust the existing recipe so that the length measurement error becomes OO nm. Required specifications: difference in length measurement value between devices OO nm or less, throughput OO or less". With such input, the LLM outputs a recipe or a set specification for the device. In the case where a set specification is output, the specification is input manually or automatically as a set condition of the recipe, and thereby a recipe is generated.

[0127] In addition, as input information of the LLM, there are semiconductor manufacturing process information (the final manufacturing process before measurement of the semiconductor device to be measured), a device name, a design drawing, material information, an output image group of the reference device A, and the like. In the presence of these additional information, it is possible to describe within the article data how to handle the data within the article data. Furthermore, it is also possible to input other information related to the reference device A, the device B, and the device C. The other information includes a set parameter of the device, detailed information of a component constituting the electron microscope, adjustment content of the component, and device information.

[0128] In Embodiment 5, the recipe generated as described above is used to operate the reference device A, the device B, and the device C to perform length measurement of a pattern formed on a semiconductor device. The output of the LLM can be a recipe common to each device, or a recipe specific to each device adjusted to reduce the mechanical error.

[0129] The computer system 101 or the like receives an output image, a length measurement value, or the like of the reference device A, the device B, and the device C controlled by the output process of the LLM. The computer system 101 or the like calculates an index value, that is, a PSNR (Peak signal noise ratio), a contrast, a sharpness, or the like for output comparison of each device from each image, or compares between the devices based on the output of the computer mounted on the measurement device. In addition, image correlation can be performed. Furthermore, the length measurement value and an area value of a pattern portion can be compared. In addition, it can be another index value indicating a degree of deviation of the index value of each device. The comparison between images and the evaluation method of the comparison result can use a deep learning model, an evaluation model, a mathematical formula, or the like.

[0130] The computer system 101 or the like functions as a preprocessing device for learning the LLM based on the output information of the device. The computer system 101 or the like functions as an input device for inputting the above comparison result together with article data or the like. The computer system 101 or the like can generate a process related to a plurality of devices corresponding to various variation factors by repeatedly performing a process of inputting the process generated by the LLM and the comparison result again to the LLM.

[0131] Figure 16 The process generation step of the device B mainly indicates a process for reducing a mechanical error between the reference device A and the device B. One of the differences from the Figure 15 One of the differences from the above is that the process information of the device B or the like is input to the LLM. The information of the reference device A and the device B has, for example, an in-device parameter, detailed information of a component used for an electron microscope, adjustment content of the component, or the like. In addition, as the information input to the LLM, there are an image group of the reference device A, a measurement result of the reference device A, semiconductor manufacturing process information, a device name, a design drawing, material information, or the like. In a case where the device (measurement device) is a superposition error measurement device, the appearance of the lower layer pattern is different depending on the depth to the lower layer pattern and the material. Therefore, adjustment of changing the energy of the electron beam is performed depending on the material and the depth.

[0132] In the example of the above Figure 16 In the example of the above

[0133] The computer system 101 or the like controls the device B based on a process output from the LLM or a process adjusted based on a parameter output from the LLM or the like. The computer that controls the device B and the computer that generates the process can be one computer or a plurality of computers.

[0134] The computer system 101 or the like receives output data of the device B, performs comparison with output data of the reference device A, and thereby generates input data to the LLM. The input data is data indicating a result of comparison between the reference device A and the device B or indicating a divergence of outputs of the two devices, and the LLM estimates an adjustment parameter corresponding to a degree of divergence from the input. In addition, image data or the like of the reference device A can be stored in advance in a storage medium as a reference image.

[0135] By performing the above processing or repeating it two or more times, it is possible to estimate a parameter of the device B in which a mechanical error between the reference device A and the device B is suppressed.

[0136] Figure 17 An example in which a process generation process of the device B for reducing a mechanical error between the reference device A and the device B, that is, a manual process adjustment according to an output of the LLM is performed. In the example of Figure 17 , by inputting an article containing a required specification and information of the device to the LLM, an adjustment content is output. According to the adjustment content, the device is manually adjusted. An article containing a meaning that the adjustment is completed is input again to the LLM, and a verification process related to a mechanical error is output. The device reads the verification process, causes the device B to automatically operate, and captures an image. Then, the captured image is compared with an image captured by the device serving as a reference, an evaluation value is calculated, and an evaluation result is fed back to the LLM.

[0137] As a manual operation, there are an operation of the adjustment content output from the LLM, an instruction input or a GUI operation, a character input or the like to the LLM, and the like. In the information of the reference device A and the device B, there are included a parameter within the device, component information used in the device, and an adjustment history of the component. In addition, the adjustment content output from the LLM includes a parameter within the device and a correction of the process.

[0138] Figure 18 An example of a GUI screen in which a display field of an article (prompt) input to a first LLM and a display field of an adjustment content output from the LLM are provided is shown. An article input in the prompt field is the same as an article input to the LLM in Figure 16 . In Figure 18 , an example in which the LLM outputs an adjustment content of "A manual operation is required. Please adjust the 〇〇 component. After the adjustment, please instruct. Perform a mechanical error evaluation here" to a comment field is shown.

[0139] Further, Figure 19 A GUI screen in which a parameter adjustment field is provided in addition to a prompt field and a comment field in which the output of the LLM is displayed is exemplified. The computer system 101 or the like is configured to display the parameter adjustment field in the GUI screen in which the prompt field and the comment field are displayed. Figure 19 The "current value" of the exemplified parameter adjustment field displays a parameter registered in the process of the device B, and the "estimated value" displays a parameter recommended by the LLM. As for the parameter displayed in the "estimated value" field, for example, a portion corresponding to each parameter is retrieved from the text data output from the LLM, which is extracted and displayed. The computer system 101 or the like is configured to apply the parameter selected in the check box provided for each "estimated value" field as update data of the process of the device B.

[0140] As Figure 17 As exemplified, an article that prompts the operator to make adjustment is made, and the LLM is used as an application for making a process that reflects the adjustment condition set by the operator, whereby the update condition of the process recommended by the LLM can be selected by the judgment based on the operator's experience.

[0141] Figure 20 A processing order when the adjustment content output from the LLM is divided into adjustment content that requires manual operation (adjustment) and adjustment content that is sufficient with the update of the process by the LLM is indicated. The LLM is configured to output a measurement instruction sheet required for the process of the device B and manual operation. The computer system 101 or the like compares the output of the device that operates according to the process output from the LLM and the device that operates according to the process that is manually set based on the measurement instruction sheet output based on the LLM, and feeds back the result of the comparison, whereby the LLM can be adjusted according to the operator's intention.

[0142] <Embodiment 6>

[0143] With recent miniaturization and complication of semiconductor device patterns, the manufacturing difficulty is increasing. Due to the complication of semiconductor devices and the like, there is a tendency for the number of manufacturing processes and inspection points to increase. Due to the increase in the number of inspection and measurement points, there is a tendency for the cost to also increase.

[0144] In Embodiment 6 of the present disclosure, a method of predicting a hot spot of a semiconductor device is described. The hot spot in Embodiment 6 refers to a portion of a wiring line in which disconnection, short circuit, or the like is likely to become a defect due to the semiconductor device. The defect is likely to occur at a portion of the semiconductor device in which the process margin is small, and can also be defined as a layout portion in which the pattern is disconnected, the patterns are in contact with each other, the desired size condition is not satisfied, or the like, in which the risk of becoming a defect is high.

[0145] In Embodiment 6, a method of predicting a hot spot using an LLM is described. By appropriate selection of a hot spot, a portion in which the defect occurrence rate is high can be selectively measured and inspected, and thus the production cost can be reduced.

[0146] Figure 21 A prediction method indicating a hot spot using an LLM. The LLM stored in a storage medium of one or more computers is configured to output a predicted position of a defect based on input of article data and design data. More specifically, one or more computer systems output a defect prediction position by inputting an article and a design drawing (whether 2D or 3D) to the LLM. Based on the output result and the correct defect generation position, a loss (Loss) value is calculated and fed back to the LLM. The calculation of the loss value uses a loss function such as cross-entropy error.

[0147] As input to the LLM, it is preferable to also input material information, process information (exposure device, etching device information), process window (degree of surplus of focus (Focus), dose (Dose)), SEM images, and the category of defects. In addition, it is preferable to also input defect position information generated in an old version of the device. Furthermore, in Embodiment 6, as layout data of a semiconductor device, an example of input design data is described, but instead, an SEM image in which a defect is displayed, position information of the defect, and the like can be input as input information to the LLM.

[0148] Defects in a semiconductor process can be generated due to a focus and a dose of an exposure device deviating from a process window. In addition, defects can be generated due to over-etching of an etching device and the like. In addition, defects can also be generated based on phenomena occurring in a process prior to a manufacturing process that becomes a measurement target, and thus this information (internal parameters of a manufacturing device of a previous process, irradiation time, and the like) can also be input as input information.

[0149] Furthermore, by inputting material information (resist or other semiconductor materials) used in the current process or a previous process as input information, it is possible to expect an improvement in the accuracy of the estimated information. Regarding design data input to the LLM, not only the current process, but also by inputting layout information of a previous process and the like, it is possible to expect an improvement in hot spot prediction performance. Furthermore, it is also possible to input output information of a measurement device (CD-SEM and the like) (SEM images of a pattern in which a defect is generated, measurement results), position information of a wafer in which a defect is generated (defect position information determined by a defect inspection device and the like), material information of a defect determined by an X-ray inspection device and the like built into a defect inspection device, and the like.

[0150] Furthermore, the learning model is minimized by minimizing the size of the deviation of the defect prediction position estimated by the LLM from the correct value. For example, the output of a process simulator can be used as initial correct data, and a model optimized in such a way that the loss value is minimized is used to predict a defect position.

[0151] Figure 21An example is shown in which an article "Please predict the defect position for the OO process. However, please within OO points. In addition, want to be set to throughput OO within" is input to the LLM. In addition, Figure 21 An example in which design data is input in addition to the article data is illustrated. According to Figure 21 The illustrated LLM is able to refer to the design data to derive a defect position candidate suitable for the number of measurements and inspections.

[0152] Figure 22 An example is shown in which the position information of the hot spot is predicted using the first LLM, and the order of the processes is generated based on the input to the second LLM using the prediction result. In addition, a plurality of LLMs are shown in the figure, but the processes can also be performed by one LLM Figure 22 The illustrated processing. The input article data represents the same case as Figure 21 According to Figure 22 The illustrated method is able to generate a process on the basis of the update condition of the process recommended by one LLM selected by the judgment based on the operator's experience rule.

[0153] In addition, by feeding back the output result from the measurement / inspection device to the LLM, it is possible to improve the prediction accuracy of the LLM. As the output of the measurement / inspection device, there is presence / absence information (presence / absence probability, etc.) of defects at the predicted defect position, classification information of defects, and the like. By inputting the classification information of defects and the like, it is possible to obtain an output corresponding to an instruction for predicting the position of a specific defect type.

[0154] In addition, it is also possible to provide the processes output from the LLM to a plurality of measurement devices. The computer system 101 or the like receives the output of the defect inspection device or the like, that is, the inspection result of the defect position, the evaluation value of the throughput of the device, and the like, and generates feedback information to the LLM. The evaluation result for the required performance value is included in the article input to the LLM, and the evaluation can also be performed by an AI model for evaluation, manual judgment, software in which an algorithm is incorporated, and the like. Furthermore, as the input to the LLM, it is also possible to input material information, process information (exposure device, etching device information), process window (degree of excess of focus, dose), SEM image, and category of defects. In addition, it is also possible to input defect position information generated in an old version of the device.

[0155] <Embodiment 7>

[0156] In recent large-scale semiconductor manufacturing, there are several hundreds of processes in one production line. In order to evaluate the completion of the device in each process or the like, or to inspect defects, foreign matter, and the like, measurement and inspection are performed. For example, an electron microscope, an optical inspection device for semiconductor evaluation and inspection performs measurement and inspection by acquiring an image, a signal waveform, and the like.

[0157] On the other hand, in the process of measurement, inspection, due to the bad condition of the measurement inspection device, the bad condition of the semiconductor manufacturing device side (process abnormality), sometimes an abnormal image (too much noise, low contrast, pattern defect) is generated, and the measurement or inspection fails. If the production line is stopped for a long time in order to determine the cause of such an abnormality, the manufacturing efficiency is reduced, and therefore early determination of the cause is desired.

[0158] In Embodiment 7 of the present disclosure, an LLM capable of performing determination of the cause of an abnormality within a semiconductor manufacturing process and prompting of evidence considered to be a cause is described. Figure 23 A step of outputting an abnormality cause and its evidence in accordance with an instruction to the LLM (article input) is described. As input data to the LLM, in addition to the above, there are a SEM image, a measurement value as an output of a measurement device, design data, and / or device information (error message, probe current value of an electron microscope, etc.).

[0159] In Figure 23 The LLM illustrated includes three models, a cause determination model, a cause analysis model, and an abstract model. The computer system 101 or the like is configured to receive data related to a problem of a semiconductor manufacturing process to the LLM. The computer system 101 or the like receives data related to a cause of an abnormality from the LLM by inputting the received article data to the LLM as a cause determination model. The cause analysis model is configured to output process, device information (probe current, device parameter of an electron microscope, etc.) as evidence. The process information has an image, a measurement value, a device condition at that time, etc. obtained as a result of automatic shooting of a device.

[0160] Figure 23 The abstract model illustrated is configured to input a cause and evidence as input and output an article or an image. The abstract model calculates an evaluation value for the output result by a machine learning model or a manual evaluation, and feeds back to the LLM.

[0161] Figure 24 A summary of the cause determination LLM is shown. The cause determination LLM is preferably learned by backpropagation based on a loss (Loss) calculation between the output of the LLM and a correct article related to a cause. More specifically, an article describing a problem generated in the LLM is input, and its cause is output as an article (the output can also be multiple). The error of the correct cause and the output cause is calculated as a loss value, and the result is fed back to the LLM. The article can be any one of manual input via the computer system 101 or the like, and automatic input of associated data in a database. The calculation method of the loss value can use an error function such as an AI model for evaluation or cross-entropy error.

[0162] Figure 25This section outlines the use of an LLM (Limited Linear Modulation) for root cause analysis. The primary root cause analysis LLM learns through backpropagation based on loss calculations between the LLM output and correlation data from inspection devices, etc. More specifically, the LLM is input with a document describing the cause, and outputs correlation data from inspection devices used to prove the cause (multiple outputs are possible). The correct correlation data from the inspection devices is compared with the output data, a loss value is calculated, and the result is fed back to the LLM. Correspondence data from the inspection devices includes internal parameters such as process parameters, probe voltage, and current, as well as data associated with device information.

[0163] exist Figure 23 The illustrated steps (system) include, for example, performing inquiries related to anomalies that can be confirmed through images, adjusting the relationship between the causes of the anomalies to be optimized, and adjusting the relationship between the causes and the evidence (a second cause that is the reason for the cause (first cause)) to be optimized. By inputting inquiries related to the anomalies into one or more such adjusted models, responses containing evidence can be obtained.

[0164] Figure 24 , Figure 25 The illustrated models can be the same model or different models. As explained in Implementation 7, by adjusting the correlation information (information related to the evidence of the main cause) of the article data and the output of the measurement and inspection device respectively, appropriate learning can be carried out by experts who oversee the semiconductor process.

[0165] Figure 26 This describes the reinforcement learning steps for the cause determination model and the cause analysis model. Specifically, the cause determination model is configured to input articles containing the problems that have occurred, and output their causes as articles. The cause analysis model, based on the articles output from the cause determination model, outputs inspection correlation data, and uses this data to automatically activate the inspection device.

[0166] Computer system 101 evaluates data output from inspection devices and feeds it back to the cause determination model and cause analysis model. Data output from the inspection device includes images and information within the device. In the case of images, metrics such as measured values ​​of circuit patterns, sharpness, and noise levels within the image are used as evaluation values. Information within the device includes probe current values, voltage values, etc., and can also be set as values ​​obtained by evaluating this information. Through reinforcement learning that uses the evaluation results as rewards, the optimization of each model can be achieved.

[0167] Further, in the fine-tuning of the LLM, the reinforcement learning phase using the actual machine, the SEM image, the measurement result, the device information, the material information, the process information, the process window (focus, degree of excess of dose, margin), the past abnormal case, the coordinates (position) of the wafer, the design drawing, and the like can be input.

[0168] For the LLM, as the article data, for example, an article such as "an abnormal image (input together with the article data) occurred. Please tell the cause thereof" is input. By being set as a composite model capable of outputting the cause and the evidence including the cause determination model and the cause analysis model, an answer such as "the cause is 'an abnormality of the previous process'. The vicinity of the abnormal image generation site was photographed, but the following abnormal images occurred frequently... From the results of the abnormality, it is considered that..." can be obtained.

[0169] <Variations of the Present Disclosure>

[0170] The present disclosure is not limited to the above-described embodiments, and includes various variations. For example, the above-described embodiments are embodiments that are described in detail in order to easily understand the present disclosure, and it is not necessarily required to have all the structures described. In addition, a part of one embodiment can be replaced with a structure of another embodiment. In addition, a structure of another embodiment can be added to a structure of one embodiment. In addition, a part of a structure of each embodiment can be added, deleted, or replaced with a part of a structure of another embodiment.

[0171] In the above-described embodiments, the learning module 113, the reinforcement learning module 114, and the article creation module 111 (or the learning module 221 for the evaluation device LLM, the learning module 231 for the semiconductor LLM, the reinforcement learning module 233, and the article creation module 234) can be constituted by a hardware such as a circuit device in which these functions are installed, or can be constituted by a software in which these functions are installed being executed by a computing device such as a CPU (Central Processing Unit).

[0172] In the above-described embodiments, all or a part of the learning module 113, the reinforcement learning module 114, and the article creation module 111 (or the learning module 221 for the evaluation device LLM, the learning module 231 for the semiconductor LLM, the reinforcement learning module 233, and the article creation module 234) can be integrally constituted, or can be constituted so as to act in linkage on the basis of these modules being separately constituted. In any case, these modules implement the method of operating the evaluation device by a string via the LLM.

[0173] Explanation of Reference Signs

[0174] 101: Computer system

[0175] 102: Semiconductor evaluation tool

[0176] 111: Article creation module

[0177] 113: Supervised / unsupervised learning module

[0178] 114: Reinforcement learning module.

Claims

1. A method executed by a computer, characterized in that, The method comprises the following steps: inputting instructions for an evaluation device used to evaluate semiconductor devices into a pre-trained language model, thereby receiving specification data describing the operational specifications of the evaluation device from the language model. The language model is pre-tuned using at least one of the domain knowledge of the semiconductor device and the domain knowledge related to the evaluation device.

2. The method according to claim 1, characterized in that, The steps for receiving the specification data are as follows: The language model input describes an input string that indicates an instruction to the evaluation device; and Receive a response for the input string from the language model. The language model is pre-tuned by learning data describing technical knowledge related to the semiconductor device and data describing technical knowledge related to the evaluation device, respectively, so that when a request related to the semiconductor device or the evaluation device is received as the input string, an answer to the request is output as the response. In the step of inputting the input string, the string that requires the language model to output data describing the instructions for the evaluation device is input as the input string. In the step of receiving the response, the response is received from the language model based on the result of the fine-tuning of the language model, and the response reflects the technical knowledge related to the semiconductor device and the evaluation device respectively as the indication.

3. The method according to claim 1, characterized in that, The method also includes the step of performing the fine-tuning on the language model. In the fine-tuning step, as technical knowledge related to the evaluation device, the language model learns at least one of the user manual and the functional specifications of the evaluation device through unsupervised learning.

4. The method according to claim 1, characterized in that, The method also includes the step of performing the fine-tuning on the language model. In the fine-tuning step, as technical knowledge related to the evaluation device, the language model is either trained through unsupervised learning to learn data describing the specifications of the image processing performed by the evaluation device, or trained through supervised learning to learn data describing the results of the image processing performed by the evaluation device.

5. The method according to claim 1, characterized in that, The method also includes the step of performing the fine-tuning on the language model. In the fine-tuning step, as technical knowledge related to the semiconductor device, the language model is enabled to learn at least one of the following through unsupervised learning: The parameters that are constrained in the process of defining the operation of the evaluation device and the data that describe the content of their constraints are described. The specifications of the image processing performed by the evaluation device are described; as well as A measurement instruction manual for creating an evaluation process is described, wherein the evaluation process defines the processing performed by the evaluation apparatus when evaluating the semiconductor device.

6. The method according to claim 1, characterized in that, The method also includes the step of performing the fine-tuning on the language model. In the fine-tuning step, as technical knowledge related to the semiconductor device, the language model is made to learn at least one of the following through supervised learning: A measurement instruction for the sequence of an evaluation process is described, wherein the evaluation process defines the processing performed by the evaluation apparatus when evaluating the semiconductor device; Data on at least one of the shape, material, and manufacturing conditions of the semiconductor device are described; Data describing the layout of the pattern shape formed on the semiconductor device; and Data on the interaction between the material of the semiconductor device and the charged particle beam irradiating the semiconductor device when the evaluation device evaluates the semiconductor device are described.

7. The method according to claim 6, characterized in that, In the supervised learning, the supervised learning is implemented by at least any one of the following: The design data of the semiconductor device is used as input to the language model, and the measurement instruction is used as output from the language model. The measurement instructions are used as input to the language model, and the evaluation process is used as output from the language model. The evaluation device takes observation images of the semiconductor device using the shooting conditions and the three-dimensional design data of the semiconductor device, or the shooting conditions and information related to the material of the semiconductor device, as inputs to the language model, and the evaluation device takes observation images of the semiconductor device as outputs from the language model. as well as The image processing filtering conditions of the observed image of the semiconductor device obtained by the evaluation device are used as input to the language model, and the observed image is used as output from the language model.

8. The method according to claim 1, characterized in that, The language model is pre-learned so that it receives instructions for the evaluation device as strings and outputs specification data describing the action specifications of the evaluation device. The method also includes the following steps: At least one of the output data from the evaluation device and the text data generated based on the output data is used as learning data to perform relearning of the language model; Input a string of natural language into the language model after the relearning is performed, instructing the evaluation device to evaluate the semiconductor device; The specification data is received as output from the language model. The evaluation device is evaluated by using the specification data received from the language model to cause the evaluation device to operate. as well as The relearning process re-provides the output data or text data obtained by activating the evaluation device.

9. The method according to claim 8, characterized in that, In the relearning step, the relearning is performed such that, by taking the language model as the learning data, the target value of the evaluation result obtained by evaluating the semiconductor device through the evaluation device, the specification data, and the evaluation result as input, the language model outputs the specification data that makes the evaluation result close to the target value.

10. The method according to claim 9, characterized in that, In implementing the relearning step, at least one of the following is input as the target value or the evaluation result: The target value of the throughput of the evaluation device or its evaluation result, The evaluation device obtains a target value or evaluation result of the measured value by measuring the semiconductor device. The target value or evaluation result of the feature quantity of the observed image of the semiconductor device obtained by the evaluation device, and The target value or evaluation result of the probability of defects occurring in the semiconductor device.

11. The method according to claim 9, characterized in that, In the relearning step, at least one of the following is input as the specification data: The evaluation process is defined as the process performed by the evaluation device when evaluating the semiconductor device. The predicted location of the defect in the semiconductor device; The evaluation device obtains an observation image of the semiconductor device when it is abnormal, and the abnormality causes and improvement measures are described. A quality improvement model that transforms the observed image of the semiconductor device output by the evaluation device into a higher quality observed image; as well as A pattern emphasis model that emphasizes specific shape patterns contained in the observed image of the semiconductor device output by the evaluation device.

12. The method according to claim 9, characterized in that, The language model is configured to output an evaluation process that defines the processing performed by the evaluation device when evaluating the semiconductor device, as the specification data. In the relearning step, the evaluation process 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 process are input, thereby enabling the language model to learn how the evaluation result changes when the evaluation process is changed.

13. The method according to claim 9, characterized in that, The language model is configured to output a measurement instruction sheet describing the sequence of the evaluation process as an intermediate output, wherein the evaluation process defines the processing performed by the evaluation device when evaluating the semiconductor device. In the relearning step, the relearning is performed by re-inputting the measurement instructions, which are output by the language model as the intermediate output, into the language model, so that the language model outputs the specification data.

14. The method according to claim 2, characterized in that, The language model is configured to learn the frequency of occurrence of strings in data describing technical knowledge related to the semiconductor device and data describing technical knowledge related to the evaluation device, respectively, and when an input string associated with either the technical knowledge related to the semiconductor device or the technical knowledge related to the evaluation device is input, the response reflects at least one of the technical knowledge related to the semiconductor device and the technical knowledge related to the evaluation device.

15. The method according to claim 1, characterized in that, The language model is configured to implement at least one of the following: It receives a natural language string and a program as input, and outputs a result that improves the program according to the conditions specified in the natural language string. The program receives a natural language string as input and outputs a program that performs the processing specified by the natural language string. The evaluation device receives a natural language string and an image of the semiconductor device as input, and outputs parameters of the evaluation device that satisfy the conditions specified by the natural language string and are suitable for the evaluation device to evaluate the shape pattern of the semiconductor device represented by the image. The device receives a natural language string and the design data of the semiconductor device as input, and outputs parameters of the evaluation device that satisfy the conditions specified by the natural language string and are suitable for the evaluation device to evaluate the semiconductor device represented by the design data, or a measurement instruction that describes the order in which the parameters are made. The evaluation device receives a natural language string and an image of an abnormal shape pattern of the semiconductor device as input, and outputs parameters of the evaluation device that satisfy the conditions specified by the natural language string and are suitable for the evaluation device to correct the abnormal shape pattern represented by the image.

16. A program that specifies a method to be executed by a computer, characterized in that, The program is configured to cause the computer to perform the following steps: inputting instructions for an evaluation device used to evaluate semiconductor devices into a pre-trained language model, thereby receiving specification data from the language model describing the operational specifications of the evaluation device. The language model is pre-tuned using at least one of the domain knowledge of the semiconductor device and the domain knowledge related to the evaluation device.

17. A method executed by a computer, characterized in that, The method comprises the following steps: inputting specification data describing the operational specifications of a first evaluation device for evaluating semiconductor devices and instructions for a second evaluation device to a pre-trained language model, thereby receiving specification data describing the operational specifications of the second evaluation device from the language model. The language model was fine-tuned based on a comparison of the output information of the first evaluation device and the output information of the second evaluation device.

18. A method executed by a computer, characterized in that, The method comprises the following steps: inputting, for a pre-trained language model, information related to the layout of a semiconductor device and information indicating the predicted location of defects generated on the semiconductor device, thereby receiving data related to the defect location from the language model. The language model was fine-tuned based on comparison information of data related to the defect location and data related to the correct solution location of the defect, which were output from the language model.

19. A method executed by a computer, characterized in that, The method comprises the following steps: inputting instructions to investigate anomalies in the manufacturing process of semiconductor equipment for one or more pre-trained language models, thereby receiving data related to the causes of the anomalies. The language model was fine-tuned based on article data related to the cause of the occurrence and associated data that serve as evidence of the cause.

Images