Semiconductor information provision method and semiconductor information provision system

Abstract

An input unit (11) of a semiconductor knowledge provision system (100) receives input of question information relating to a semiconductor-related device. A search unit (12) references a knowledge DB (13) in which semiconductor-related information regarding semiconductor-related devices is stored to acquire semiconductor-related information similar to the question information, and generates an extended prompt in which the question information and the semiconductor-related information are associated with each other. An answer LLM (14) refers to the semiconductor-related information regarding the semiconductor-related device on the basis of the generated extended prompt, and generates answer information for the question information. An output unit (15) outputs the generated answer information.

Description

Semiconductor information providing method and semiconductor information providing system

[0001] The present disclosure relates to a semiconductor information providing method and a semiconductor information providing system.

[0002] Manufacturing apparatuses for manufacturing semiconductor devices and apparatuses for measuring and inspecting semiconductor wafers manufactured by such manufacturing apparatuses have many setting parameters. Such setting parameters should be selected and set with high accuracy in order to appropriately manufacture a semiconductor wafer as a product. For example, a semiconductor manufacturing apparatus or the like is controlled by an operation program (hereinafter sometimes referred to as a recipe) set to appropriately operate the semiconductor manufacturing apparatus, and the recipe requires setting and adjustment by an experienced person with advanced expertise.

[0003] It would be ideal if a recipe could always be set by an experienced person with advanced knowledge, but there may be cases where such personnel are not available. On the other hand, there is also a large language model (LLM) that outputs an answer from existing knowledge when a question is input. By fine-tuning (additional learning) the LLM with related information on semiconductor-related apparatuses, the LLM can also answer semiconductor-related questions from beginners (Patent Document 1).

[0004] International Publication No. 2023 / 223535

[0005] Techniques using fine-tuning such as Patent Document 1 have high learning costs and there are also concerns about overlearning and forgetting of learning data. Furthermore, there may be cases where it is difficult to expand domain knowledge.

[0006] The present disclosure has been made in view of such a situation, and the main problem is to reduce the training burden when using an artificial intelligence-based solution configured to output information about semiconductor-related apparatuses.

[0007] To solve the above problems, the semiconductor knowledge provision system of this disclosure has the following features. The present invention is a semiconductor knowledge provision system that is executed by a semiconductor knowledge provision system, wherein the semiconductor knowledge provision system comprises an input subsystem, a search subsystem, an LLM, and an output subsystem, wherein the input subsystem receives input of question information relating to semiconductor-related equipment, the search subsystem queries a database storing semiconductor-related information about the semiconductor-related equipment to obtain semiconductor-related information similar to the question information, and generates an augmented prompt by a RAG (Retrieval Augmented Generation) engine that associates the question information received by the input subsystem with the obtained semiconductor-related information, the LLM generates answer information to the question information by referring to the semiconductor-related information about the semiconductor-related equipment based on the generated augmented prompt, and the output subsystem outputs the generated answer information. Other features will be described later.

[0008] According to this disclosure, the training burden when using an artificial intelligence-based solution configured to output information about semiconductor-related equipment can be reduced.

[0009] This is a diagram illustrating the configuration of the semiconductor knowledge provision system according to this embodiment. This is an explanatory diagram of the main part of the semiconductor knowledge provision system according to this embodiment. This is a diagram illustrating the configuration of the computer system on which the semiconductor knowledge provision system according to this embodiment operates. This is a diagram illustrating the configuration of each semiconductor-related device that provides knowledge documents to be registered in the knowledge database by the update unit according to this embodiment. This is an explanatory diagram showing an example of the registered contents of the knowledge database according to this embodiment. This is an explanatory diagram illustrating the new registration process in the knowledge database by the update unit according to this embodiment. This is an explanatory diagram illustrating the update process of the knowledge database by the update unit according to this embodiment. This is an explanatory diagram illustrating the registration process of semiconductor-related information regarding device differences according to this embodiment. This is a screen diagram of the database construction field according to this embodiment. This is a screen diagram of the question input field according to this embodiment. This is a screen diagram of the response output field according to this embodiment. This is a screen diagram of the language model specification field according to this embodiment. This is a screen diagram of the context information confirmation field according to this embodiment. This is a screen diagram of the template input field according to this embodiment. This is an explanatory diagram showing an example of use in recipe creation as a first example of use according to this embodiment. This is an explanatory diagram showing an example of use in recipe error analysis and troubleshooting as a second example of use according to this embodiment. This is a screen diagram of the database maintenance field according to this embodiment. This is an explanatory diagram showing an example of use in analyzing and reducing differences between multiple measuring devices, as a third example of application of this embodiment. This is an explanatory diagram showing an example of use in defect inspection, as a fourth example of application of this embodiment. This is a configuration diagram of the semiconductor knowledge provision system according to this embodiment. This is a flowchart showing the process of outputting an answer based on the input of a question, according to this embodiment. This is a diagram showing an example of a GUI screen having a question input field, according to this embodiment.

[0010] Hereinafter, an embodiment of this disclosure will be described with reference to the drawings. This disclosure describes a method, system, program, and computer-readable medium that can receive knowledge from experienced individuals in written form for setting recipes, etc. Specifically, this disclosure proposes a method for outputting desired information by systematically identifying, aggregating, and / or acquiring setting information related to manufacturing equipment, measuring equipment, and / or inspection equipment, and related data such as semiconductor wafers that are products manufactured by the above manufacturing equipment, etc. More specifically, this disclosure proposes a method using an artificial intelligence system that outputs information related to the setting conditions of semiconductor-related equipment or generates a report based on semiconductor-related information of semiconductor-related equipment.

[0011] Figure 1 is a diagram of the configuration of the semiconductor knowledge provision system 100. The semiconductor knowledge provision system 100 is an AI system for supporting the operation of equipment for semiconductor manufacturing processes. The semiconductor knowledge provision system 100 has an input unit 11, a search unit 12, a knowledge database 13, an answer LLM 14, an output unit 15, and an update unit 16. The input unit 11 receives input of question information (prompt) regarding semiconductor-related equipment from the questioner 31 (S31) and notifies the search unit 12 of the question (S32). The search unit 12 is composed of a Retrieval Augmented Generation (RAG) engine (hereinafter referred to as the "RAG engine"). The RAG engine may be machine learning based or may be a search engine that does not use machine learning.

[0012] The search unit 12 retrieves semiconductor-related information similar to the question information by querying the knowledge database 13, which stores semiconductor-related information about semiconductor-related equipment (S33). The search unit 12 generates an extended prompt that associates the question information received by the input unit 11 with the acquired semiconductor-related information, and inputs the extended prompt into the answer LLM 14 (S34). A "prompt" refers to an instruction or command used when interacting with the generating AI, such as "Please tell me about XX" or "Please summarize the contents of XX." In this disclosure, the question information is a prompt, and the "extended prompt" is a prompt that associates semiconductor-related information with the question information.

[0013] The response LLM 14 generates response information (answer text) to the question information by referring to semiconductor-related information about semiconductor-related equipment based on the generated extended prompt. The response LLM 14 notifies the output unit 15 of the response information (S35). The output unit 15 outputs the generated response information. For example, the output unit 15 replies to the questioner 31 with the response text (operation instructions, etc.) from the response LLM 14 (S36), and also outputs control commands (automatic operation commands, etc.) based on the response text to the measuring device 21 (S37). The update unit 16 registers log information of the measuring device 21, etc., in the knowledge DB 13.

[0014] In this way, the input unit 11 receives questions from the questioner 31 in the form of text data, image data, etc., indicating requirements and operations related to semiconductor-related equipment. The answer LLM 14 outputs a response to the question. "Semiconductor-related equipment" refers to equipment (manufacturing equipment, measuring equipment, inspection equipment, etc.) used in each process related to semiconductor devices (research process, design process, manufacturing process, measurement and inspection process, etc.), and also includes the semiconductor device itself. Semiconductor measurement and inspection equipment among semiconductor-related equipment is, for example, a critical dimension-scanning electron microscope (CD-SEM) that applies scanning electron microscope (SEM) technology. The search unit 12 acquires semiconductor-related information for at least one of the following devices as semiconductor-related equipment: semiconductor manufacturing equipment, semiconductor devices manufactured by semiconductor manufacturing equipment, semiconductor measuring equipment for measuring semiconductor devices, and semiconductor inspection equipment for inspecting semiconductor devices.

[0015] "Semiconductor-related information" refers to domain knowledge (context information) related to semiconductor-related equipment, such as the following, and is stored in Knowledge DB13: • Semiconductor domain knowledge (educational materials, papers) • Measurement and inspection equipment domain knowledge (instruction manuals, specifications) Furthermore, domain knowledge related to semiconductor-related equipment includes, for example, the following: • Error information (events that have occurred in the past) and their causes or countermeasures (such as replacing faulty parts that caused the error) • Instruction manuals • Functional specifications, recipe information, log information (usage history, anomaly resolution history, process anomaly resolution history, maintenance history) • Individual identification information

[0016] The search unit 12 acquires at least one of the following as semiconductor-related information obtained from the knowledge DB 13: the operating program of the semiconductor-related equipment, the setting conditions set for the semiconductor-related equipment, the documentation describing how to handle the semiconductor-related equipment, the instruction information for the semiconductor-related equipment, the information of the semiconductor devices handled by the semiconductor-related equipment, the output information from the semiconductor-related equipment, and the events that have occurred in the past with the semiconductor-related equipment and the corresponding information for those events.

[0017] The search unit 12 searches for the similarity between the semiconductor-related information stored in the knowledge database 13 and the question, and extracts one or more semiconductor-related information items in descending order of similarity (hereinafter referred to as "extracted related information"). The search unit 12 then generates an extended prompt that links the extracted related information with the question. The answer LLM 14 refers to the extracted related information based on the extended prompt and outputs a response to the question. It is preferable that the answer LLM 14 is a local model that does not connect to an external server.

[0018] Furthermore, the answer LLM 14 may use an LLM (LLM without additional learning) in which semiconductor-related information has not been fine-tuned (additional learning LLM) at all. Alternatively, the answer LLM 14 may use an LLM (LLM with additional partial learning) in which semiconductor-related information has been fine-tuned to a extent that does not incur a large learning cost. In this way, whether using an LLM without additional learning or an LLM with additional partial learning, it is not necessary to fine-tune all of the semiconductor-related information into the answer LLM 14. This allows for effective utilization of the domain knowledge in the knowledge DB 13 and prevents confidential information from being leaked to the outside from the answer LLM 14.

[0019] The additional partial learning LLM used as the answer LLM 14 is pre-fine-tuned with the following prerequisite knowledge for interpreting the question information received by the input unit 11 or semiconductor-related information in the knowledge DB 13: knowledge about semiconductor devices, knowledge about the manufacturing of semiconductor devices, knowledge about the evaluation (measurement or inspection) of semiconductor devices, and knowledge about basic terminology in the semiconductor field. On the other hand, the update unit 16 stores case knowledge, such as log data, past recipes, customer conditions, and manuals, which are frequently updated, in the knowledge DB 13 as data that has not been fine-tuned. As a result, the roles are divided so that case knowledge is included in the extended prompt and prerequisite knowledge is included in the answer LLM 14, saving capacity in the knowledge DB 13 and reducing the update load on the update unit 16.

[0020] The search unit 12 generates a composite data as an extended prompt by combining the input data received by the input unit 11 with data pre-prepared in the knowledge database 13. When the answer LLM 14 receives an extended prompt from the search unit 12, it generates information related to semiconductor-related equipment as an answer to the extended prompt. In this configuration, the search unit 12 acquires semiconductor-related information from the knowledge database 13, allowing the questioner 31 to generate a detailed extended prompt with a large amount of data by inputting only a small amount of data from the input unit 11. Because this extended prompt complements the input question with specialized semiconductor-related information, the answer LLM 14 can output more relevant information compared to when the question is input as text data.

[0021] Figure 2 is an explanatory diagram of the main parts of the semiconductor knowledge provision system 100. Explanations of the symbols (S30 to S36) corresponding to each arrow, which have already been explained in Figure 1, are omitted in Figure 2. The update unit 16 processes the knowledge document 13D, which contains semiconductor-related information of semiconductor-related equipment, as appropriate to facilitate processing (details in Figures 6 and 7), and registers the results in the knowledge database 13 (S30). The search unit 12 uses the RAG engine to search the knowledge database 13 for semiconductor-related information similar to the question received by the input unit 11 (S33A), and retrieves the search results (Retrieval) (S33B). The search unit 12 generates a new augmented prompt (Augmented) that links the acquired semiconductor-related information with the question from the input unit 11. The answer LLM 14 generates (Generates) an answer to the question (operation instructions for the equipment and error analysis results) based on the augmented prompt input from the search unit 12.

[0022] Thus, instead of acquiring semiconductor-related information from a finely tuned LLM, the semiconductor knowledge provision system 100 acquires semiconductor-related information from the knowledge DB 13 using the RAG engine. This provides the following benefits: - Additional learning of the LLM is unnecessary, and the acquired semiconductor-related information can be provided to the questioner 31. - Since semiconductor-related information is acquired directly from the knowledge DB 13, accurate and consistent answers can be generated. - Even when the frequency of updates to semiconductor-related information is high, such as when new products are released, additional learning of the LLM is unnecessary, and the latest semiconductor-related information can be provided to the questioner 31 simply by the update unit 16 updating the knowledge DB 13.

[0023] Figure 3 is a diagram showing the configuration of a computer system 10 on which the semiconductor knowledge provision system 100 operates. The computer system 10 consists of one or more computers, each having one or more processors 19P and one or more memories 19M. This computer may further have an HDD, a communication interface, an input / output interface, and a media interface. The communication interface is connected to an external communication device via a network 18 provided in a semiconductor device design department or semiconductor manufacturing plant. The input / output interface is connected to an input / output device. The media interface reads and writes data to and from a recording medium.

[0024] Furthermore, the processor 19P executes the processes (S30 to S36) shown in Figures 1 and 2 by controlling each processing unit through the execution of program instructions 19C loaded into memory 19M. Program instructions 19C are, for example, instructions for generating information related to semiconductor-related equipment using the search unit 12 (RAG model of the RAG engine) and the answer LLM 14. These program instructions 19C can also be distributed via a communication line or by recording them on a recording medium such as a CD-ROM. The search unit 12 of the computer system 10 is equipped with a RAG engine. The input / output device 31G is a device that handles the input unit 11 and the output unit 15.

[0025] Program instruction 19C includes the following instructions: - An instruction (S11) in which the input unit 11 prompts the questioner 31 to input a question via the input / output device 31G. - An instruction (S12) in which the search unit 12 searches the knowledge database 13 for documents related to the question in S11. - An instruction (S13) in which the knowledge database 13 returns semiconductor-related information that is similar (highly relevant) to the question in S11. - An instruction (S14) in which the search unit 12 inputs an extended prompt, which is composite information of the question in S11 and the semiconductor-related information in S13, to the answer LLM 14. - An instruction (S15) in which the answer LLM 14 generates an answer to S14. - An instruction (S16) in which the output unit 15 displays the answer to S15.

[0026] Figure 4 is a configuration diagram of each semiconductor-related device that provides knowledge documents 13D to be registered in the knowledge database 13 by the update unit 16. Each semiconductor-related device is a test device 24A, a first measurement device 21A, a second measurement device 21B, a design database 26, a first process device 27A, a second process device 27B, and a simulator 28. Each of these devices is connected to the devices in Figure 3 (computer system 10, knowledge database 13, response LLM 14) via a network 18. By registering the knowledge documents 13D acquired from each of the devices in Figure 4 in the knowledge database 13, the update unit 16 can respond to new tasks and improve response information.

[0027] The update unit 16 may also create a database (register information in the knowledge DB 13) by associating information on multiple parameter fluctuations with countermeasures, as illustrated below. • The database may be constructed based on actual parameter fluctuations and countermeasures (document information). • The relationship between measurement conditions for the simulator (parameters such as beam acceleration voltage and beam current) and measured values ​​obtained based on those inputs may be created in the database. • Fluctuations in measured values ​​due to fluctuations in the process conditions of the process apparatus (monitored based on sensor output attached to the process apparatus) and conditions for readjusting the process conditions may be created in the database. • Parameters or CD values ​​output from the process apparatus that indicate that CD (Critical Dimension) value fluctuations are due to circumstances on the process apparatus side and not the measurement device may be created in the database.

[0028] Furthermore, the update unit 16 may create a database by associating the occurrence of an event with its countermeasures (for example, a countermeasure of not taking any measures regarding the adjustment of the measuring device), as illustrated below. • The correlation between the sensor output of the manufacturing device and the measurement and inspection results of the measurement and inspection device may be created in the database. • Information from multi-databases and multi-search engines may be created in the database. • Process conditions for at least two or more steps may be created in the database. • Documents regarding the characteristics of the object being measured and inspected (material, node, process conditions, similarity of shape and / or arrangement, pattern density, etc.), phenomena occurring during measurement and inspection (causes), and countermeasures corresponding to those phenomena may be created in the database. • Additional questions (e.g., "Is it correctly positioned?", "What is the beam current value?", etc.) may be included in the countermeasures and created in the database to prompt additional questions (to supplement the information provided to the LLM).

[0029] Figure 5 is an explanatory diagram showing an example of the contents registered in the knowledge database 13. The knowledge database 13 is configured as a tree structure 110 in which semiconductor-related information indicated by each node is linked from higher-level nodes to lower-level nodes. For example, the parent node "Measurement-related" is connected to the child node "Operation Method," etc. The child node "Operation Method" is connected to the grandchild node "Recipe Creation Procedure Manual," etc. When the search unit 12 extracts semiconductor-related information from a higher-level node by search, it may also acquire semiconductor-related information from its descendant nodes and include it in the extended prompt.

[0030] Figure 6 is an explanatory diagram of the new registration process in the knowledge database 13 by the update unit 16. The update unit 16 receives input of a knowledge document 201 (knowledge document 13D in Figure 2) to be updated (new registration target) and generates multiple document fragments (chunks 202) by dividing the knowledge document 201 into predetermined sizes. The predetermined size is set appropriately according to the purpose, etc. Then, the update unit 16 vectorizes each chunk 202 to facilitate similarity search and generates a chunk vector 203 as a result. In other words, one chunk vector 203 is generated from one chunk 202. The chunk vector 203 is, for example, data indicating the frequency of word occurrences within the chunk 202 (the larger the value of the chunk vector 203, the higher the application frequency). The update unit 16 newly registers each chunk vector 203 in the knowledge database 13.

[0031] Figure 7 is an explanatory diagram of the update process of the knowledge DB 13 by the update unit 16. Assume that an update location 211N occurs in a knowledge document 211 that has already been registered in the knowledge DB 13. The update unit 16 identifies the chunk 212N to which the update location 211N belongs and generates a new chunk vector 213N for that chunk 212N. The update unit 16 updates the chunk vector 213N of the update location 211N in the knowledge DB 13, and omits updating the chunk vector 213 of chunks 212 to which the update location 211N does not belong in the knowledge DB 13. By performing a differential update only on the update location 211N, the update unit 16 can achieve an efficient update process to the knowledge DB 13. Alternatively, instead of performing a differential update of the chunk vector 213 as a differential update by the update unit 16, a differential update of chunk 212 may be performed without vectorization. In this way, the update unit 16 generates chunks of semiconductor-related information divided into predetermined sizes, and updates the contents of the knowledge DB 13 differentially on a chunk-by-chunk basis. Alternatively, the update unit 16 generates chunk vectors by vectorizing each generated chunk, and updates the contents of the knowledge DB 13 differentially on a chunk vector basis instead of performing differential updates on a chunk-by-chunk basis.

[0032] Figure 8 is an explanatory diagram showing the registration process of semiconductor-related information regarding device differences. As shown by reference numeral 221, even if the same parameter items (a, b, c, ...) are used between measuring devices A and B, some of the parameter values ​​may differ (in Figure 8, parameters c and f differ between the two devices). In other words, device differences occur between measuring devices A and B. As shown by reference numeral 222, there are various combinations of measuring devices A, B, C, ... As shown by reference numeral 223, the combinations of measuring devices A, B, C, ... can be managed in a table format. The update unit 16 records the device differences (for example, parameters c and f differ between the two devices) for each combination in this table (for example, between measuring devices A and B). The update unit 16 then registers this table containing the device differences as semiconductor-related information in the knowledge database 13.

[0033] The following describes the screen diagrams of the semiconductor knowledge provision system 100 with reference to Figures 9 to 17. Figure 9 is a screen diagram of the database construction area 231. The database construction area 231 is used by users who register knowledge documents 13D in the knowledge DB 13, such as administrators 32 and workers 33. The database construction area has a UI for opening knowledge document 13D files, allowing users to select individual files such as "CV6300_Instruction Manual.pdf" as the target for registration in the knowledge DB 13. Note that the database construction area in Figure 9 is an example of a GUI for constructing the knowledge DB 13, and is a GUI for reading documents (device instruction manuals, manuals, device log information, recipes). Alternatively, a GUI for specifying the production management system of a linked customer or the design data of the knowledge document 13D may be used as a source for obtaining knowledge documents 13D.

[0034] Figure 10 is a screen view of the question input field 232. The input unit 11 displays the question input field 232, which includes a text box for the questioner 31 to enter a question, and obtains question information from the content entered in the question input field 232. For example, in Figure 10, the question information entered in the question input field 232 is a combination of the sentence "Please tell me the procedure for creating a measurement recipe for the XXX pattern under the following conditions" and the conditions "THP: less than or equal to XXX, skew: less than or equal to XXX". In this way, the question input field 232 is provided as a GUI that allows text-based communication as a chat system between the questioner 31 and the answering LLM 14, so the questioner 31 does not have to incur the cost of learning a GUI.

[0035] Figure 11 is a screen view of the response output field 233. The output unit 15 displays the response output field 233 for displaying the response to the question from the answering LLM 14 to the questioner 31. The response output field 233 outputs steps such as "1. Click the □□ button in the ○○ window" in response to the question information in Figure 10, "Please tell me the steps to create ~".

[0036] Figure 12 is a screen view of the language model specification field 234. The search unit 12 prompts the questioner 31 to specify the language model to be used as the answer LLM 14 from among several LLM candidates via the language model specification field 234. The questioner 31 inputs the identification information of the language model, such as "Meta-Llama-3.1-70B, RakutenAI-7B-chat, shisa-gamma-7b-v1", into the language model specification field 234.

[0037] Figure 13 is a screen view of the context information confirmation area 235. The output unit 15 displays the basis (information source) of the response output in the response output area 233 via the context information confirmation area 235 as semiconductor-related information used by the answer LLM 14 to generate the response. In other words, the following is displayed in the context information confirmation area 235: - Context information extracted in response to the question. In Figure 13, it is "Chapter ○ ~~~". - Information indicating where the context information was extracted. In Figure 13, it is "from:CV6300_Instruction Manual.pdf". In this way, the answer LLM 14 generates the response information to the question information, including information on the source of the referenced semiconductor-related information. When the output unit 15 outputs the generated response information "Chapter ○ ~~~", it also outputs the information source "from:CV6300_Instruction Manual.pdf", which is the basis of the response information.

[0038] Figure 14 is a screen view of the template input field 241. In the template input field 241, some of the questions are pre-set as templates for the question input field 232 in Figure 10, and questions can be entered efficiently by filling in the blanks in the input fields within the template. In the template input field 241, input fields are inserted in the middle of the displayed text, for example, "An error code (input field) occurred in the process (input field). Recipe name (input field), wafer number (input field)."

[0039] Questioner 31 selects their desired use case from the list of use cases displayed to the left of the template input field 241. [Use Case 1] Use case in recipe creation (see Figure 15) [Use Case 2] Use case in recipe error handling (see Figure 16) [Use Case 3] Use case in equipment difference analysis (see Figure 18) [Use Case 4] Use case in defect inspection (see Figure 19)

[0040] The input unit 11 displays a template suitable for the selected usage example to the right of the template input field 241 and prompts the user to enter information in the input field (blank). The answer LLM 14 then communicates with the questioner 31, as follows, between the entered question information and the answer information output from the extended prompt for that question information: Questioner 31: "I can't measure the length of the image of process X. Why?" Answer LLM 14: "Is the image clear?" Questioner 31: "Yes, it is." Answer LLM 14: "Is the length measurement box correctly positioned?" Questioner 31: "It's in an area unrelated to length measurement." Answer LLM 14: "Please adjust the xxx parameter related to the length measurement box position." The search unit 12 may also evaluate whether appropriate decisions are being made at each phase. An example of such evaluation is outputting the top 5 answers with high confidence and comparing them with the correct answer value.

[0041] Figure 15 is an explanatory diagram showing an example of use in recipe creation as the first example of application. Below, an example of the process by which the recipe management device 22 creates a recipe file for semiconductor-related equipment will be described. The semiconductor knowledge provision system 100 has the recipe management device 22 as an update unit 16. Recipes are stored in a storage medium built into the recipe management device 22. Semiconductor-related equipment is controlled by an operating program called a recipe. A recipe is prepared in advance, for example, to automatically control a CD-SEM that constitutes part of the measuring device 21. A recipe is set based on a measurement instruction sheet (a document describing the user's measurement conditions) in which the administrator 32 (operator) has set the desired measurement conditions in advance. The measurement instruction sheet may include identification information (ID) of the semiconductor wafer to be measured, identification information and coordinate information of the pattern to be measured on the semiconductor wafer.

[0042] Furthermore, the recipe management device 22 may be configured to allow reference to semiconductor wafer design data written in GDS (Graphic Design System) format or OASIS (Open Artwork System Interchange Standard) format. The recipe management device 22 may be configured to allow extraction of internal structure and placement conditions, for example, where measurement locations are described in the design instructions, by comparing them with the design information. The recipe management device 22 receives semiconductor wafer information and pattern information described in the measurement instructions, and further receives additional conditions for measuring the pattern, and then generates a recipe. Specifically, the recipe management device 22 generates a file (hereinafter sometimes referred to as data defining pattern placement) that specifies the chips, etc., containing the pattern to be measured, by specifying the chips, shot (the area exposed in one shot of the exposure device) formed on the semiconductor wafer, placement information within multiple chips, etc.

[0043] Furthermore, the recipe management device 22 selects a chip that includes a pattern (measurement point) to be measured. Also, in order to place the chip including the measurement point at an appropriate position for measurement, it selects a pattern (alignment pattern) for visual field position identification using template matching, the irradiation conditions of the beam with respect to the alignment pattern, and the like. The alignment may be performed using an optical microscope instead of an electron beam. After registering, the recipe management device 22 actually performs alignment and checks whether the alignment is correctly performed.

[0044] Next, the recipe management device 22 selects a unique pattern (addressing pattern) that has a known positional relationship with the measurement point (measurement target pattern), the beam irradiation conditions with respect to the addressing pattern, and the like. The addressing pattern is a unique pattern included in a wide-field image including the measurement point and is used for specifying the position of the measurement point by template matching by image processing. Furthermore, the recipe management device 22 sets the beam conditions for generating an image of the measurement point, the position of the length measurement box, and the like, actually performs length measurement, and checks whether the measurement conditions are properly set. Note that the focus width for performing auto-focus of the beam is set together and the conditions are stored. Based on the above condition setting, the recipe management device 22 generates a recipe file, which is an operation program of the device.

[0045] The knowledge database 13 stores, for example, semiconductor-related information (at least one piece of information) as exemplified below. The search unit 12 can refer to this semiconductor-related information and generate appropriate questions (extended prompts) corresponding to the question information contained in the input unit 11. ・Output information of the recipe management device 22 equipped with a storage medium that stores past recipes ・Past measurement instructions (such as optical conditions of the electron microscope if the measurement device 21 is a CD-SEM) ・Information from a measurement instruction sheet (measurement requirements specification) containing instructions for the operator to create a recipe ・Recipes generated based on the measurement specifications in the measurement instruction sheet ・Instruction manual for the measurement device 21 ・Wafer design document (such as design data written in GDS or OASIS format) ・SEM images (for example, images generated by past recipes) ・Device operation screen In this way, the update unit 16 stores recipe files that show the operation of semiconductor-related equipment for manufacturing semiconductor devices in the knowledge database 13 as semiconductor-related information.

[0046] Next, we will explain the process from the construction of the knowledge database 13 to the generation of answers to measurement questions using the constructed knowledge database 13 by the answer LLM 14. First, the update unit 16 stores data related to measurement instructions and design documents in the knowledge database 13. At this time, it is desirable for the update unit 16 to construct the knowledge database 13 by pre-storing instruction manuals, wafer design documents, recipes, etc., with similar conditions to the equipment conditions described in the measurement instruction manual. The search unit 12 notifies the knowledge database 13 of a search request in response to the question information received from the input unit 11, for example, to search for semiconductor-related information such as instruction manuals, wafer design documents, recipes, etc. with similar conditions. It is also possible to include in the search request whether there are any points to note before creating a recipe.

[0047] The search unit 12 inputs an extended prompt associating the search result of semiconductor-related information obtained from the knowledge DB 13 with the question information received from the input unit 11 to the answering LLM 14, and the answering LLM 14 outputs an answer. By asking questions as described above, it is possible to expect the answering LLM 14 to output explanations such as parameter settings that should be noted and precautions for measurement conditions. The questioner 31 can obtain preliminary knowledge for recipe generation by receiving an answer as described above.

[0048] Next, the questioner 31 gives the following questions regarding the operation procedures and parameter settings necessary for recipe setting via the input unit 11. This next question includes at least one of the following questions. - Questions about the method of setting data for defining pattern placement, alignment patterns, and addressing patterns. - Questions about beam irradiation conditions (such as the fluctuation range during beam autofocus) for image generation. - Questions about the position conditions of the length measurement box, other parameter candidates, etc. - Questions about the pattern information to be measured and the device conditions for measuring the pattern information.

[0049] Based on the semiconductor-related information obtained from the knowledge DB 13, the answering LLM 14 outputs operation instructions and setting candidates. Since this output is set after referring to precautions, it is more refined compared to the case of directly inputting questions about recipe setting information. The answering LLM 14 outputs an answer based on the above input. For example, if the question includes content asking about the influence on measurement reproducibility and throughput, information related to the degree of influence can be obtained as an answer. Therefore, the update unit 16, for example, stores a past measurement instruction manual and a recipe generated based on the instruction manual in the knowledge DB 13 in advance. The search unit 12 can generate an extended prompt of recipe-related information (only fragmentary information when searching, as context) referring to past semiconductor-related information by searching the knowledge DB 13 based on the input of a new instruction manual. The answering LLM 14 can output an answer that has broken down the extended prompt.

[0050] Figure 16 is an explanatory diagram showing a second application example, specifically in recipe error analysis and troubleshooting. The semiconductor knowledge provision system 100 includes a shaping LLM 16 as an update unit 16, in addition to the recipe management device 22 shown in Figure 15 (explained later). As described above in the explanation of Figure 15, semiconductor-related equipment is controlled by an operating program called a recipe. However, semiconductor-related equipment may operate correctly in the initial stages of operation, but the equipment conditions may become inappropriate due to changes in the equipment's timing, etc. Therefore, semiconductor-related equipment needs to detect errors from information obtained from the output of sensors installed within the equipment or from images, and perform troubleshooting to resolve those errors.

[0051] For example, if the semiconductor-related equipment is a CD-SEM, the following steps are performed: - Introduction of the semiconductor wafer into the vacuum sample chamber before the measurement process - Alignment to position the chip including the pattern to be measured - Addressing to determine the position of the measurement point Here, error analysis and troubleshooting can be performed in a flowchart format (hereinafter referred to as the "inspection flowchart") in the order of the steps, and the analysis (confirmation) and correction of settings that are deemed inappropriate as a result of the analysis can be performed. In this way, the update unit 16 stores the troubleshooting history, which shows the content of past troubleshooting, as semiconductor-related information in the knowledge DB 13.

[0052] The knowledge database 13 stores information indicating at least one of the following errors, along with troubleshooting information (troubleshooting history, inspection flowchart, etc.) to resolve that error: - Error log at the time the error occurred - SEM image acquired at the time the error occurred - Device operation screen - SEM image (for example, an image generated by a past recipe) In addition, the knowledge database 13 may also store the formatted troubleshooting information response obtained by inputting troubleshooting information from the operator 33 into the formatting LLM 16. The following describes the process flow for obtaining troubleshooting information as a response using the knowledge database 13 constructed in this manner.

[0053] First, the questioner 31 inputs the items to be checked according to the inspection flowchart into the input unit 11. The search unit 12 retrieves semiconductor-related information similar to the items to be checked from the knowledge database 13 and generates an extended prompt. The answer LLM 14 outputs answer information from the output unit 15 that indicates possible solutions to the error in response to the extended prompt. The answer LLM 14 proposes possible solutions based, for example, on the error log (error number) and the pattern matching score. In addition, the answer LLM 14 can propose error handling information by comparing the SEM and measurement images at the time of the error and referring to past solutions.

[0054] Figure 17 is a screen view of the database maintenance area 300. This database maintenance area 300, like the database construction area 231 in Figure 9, is a screen for registering semiconductor-related information in the knowledge DB 13. The operator inputs the errors that have been identified and troubleshooting information showing the actions taken to resolve those errors into the database maintenance area 300. The operator can also instruct the next operation from the database maintenance area 300, which can display the output of the response LLM 14.

[0055] The database maintenance section 300 includes the following elements: • Item input section 301 is for inputting various information related to maintenance work on semiconductor-related equipment. • Event section 302 is for inputting the occurrence of an event (mainly information indicating the type of error). • Cause section 303 is for inputting the cause of the occurrence of an event (if known). • Maintenance Task / Remarks section 304 is for inputting the details of the maintenance work performed by the expert (operator) to address the occurrence of the event. • Cost section 305 is for inputting the cost of the maintenance work, such as the cost of replacement parts and the time taken. • Log Data (before Maintenance) section 306 is for inputting log data of the semiconductor-related equipment before maintenance work. • Log Data (after Maintenance) section 307 is for inputting log data of the semiconductor-related equipment after maintenance work. By recording log data before and after maintenance in this way, the effectiveness of the maintenance can be quantified. The button group 308 is a collection of buttons that accept operations on the edited content of each field in the database maintenance section 300, and includes a New button for creating a new entry and a Save button for saving data.

[0056] The following information will be entered in item input field 301, for example: • "Operator ID" will contain the ID of the operator (expert, etc.) performing the maintenance work. • "Date / Time" will contain the time of the maintenance work performed, which will be used to record and associate time information such as the time of best condition occurrence and worst condition occurrence with the error log and task details. • "Equipment 1,2,3" will contain identification information (name, ID, etc.) for semiconductor-related equipment.

[0057] Thus, the database maintenance section 300 is structured to facilitate the organized input of semiconductor-related information using a templated input form. This makes it easy for even beginners unfamiliar with handling semiconductor-related equipment to input well-organized semiconductor-related information. With this database maintenance section 300, the operator can register error information and countermeasures for resolving errors together in the knowledge DB 13. Therefore, the questioner 31 can obtain appropriate information on countermeasures for resolving errors based on their questions about errors.

[0058] Figure 18 is an explanatory diagram illustrating a third application example: the analysis and reduction of device differences between multiple measuring devices 21. The semiconductor knowledge provision system 100 has multiple measuring devices 21 as an update unit 16. One of the multiple measuring devices 21 is the master device. Among multiple measuring devices 21 that physically measure semiconductor devices (circuit patterns), "device differences" can occur, resulting in different measurement results. In a semiconductor manufacturing plant, there are hundreds of processes in one manufacturing line. On the other hand, there may be multiple manufacturing lines, and even within a single manufacturing line, there may be multiple identical manufacturing devices or measurement / inspection devices. Device differences between lines, or between identical devices within the same line, can cause performance differences in semiconductor elements, so reducing device differences is required.

[0059] As semiconductor device patterns become smaller, it is expected that there will be a greater need to reduce instrumentation errors. Semiconductor-related equipment has many setting parameters, and individual differences in the components that make up the equipment also contribute to instrumentation errors. For example, if there are multiple measuring devices 21 of the same type, it is necessary to adjust them so that the measurement results when measuring the same pattern are the same.

[0060] Therefore, the Knowledge DB 13 contains the following semiconductor-related information: • Equipment design drawings • Manufacturing process information for semiconductor wafers to be measured • Measurement results • Adjustment flow during equipment setup (introduction) • Equipment parameters • Equipment logs • SEM images • Feature quantities extracted from SEM images (contrast, sharpness, noise (S / N ratio), frequency characteristics, etc.) • Equipment operation screen • Customer specifications (target values) required for the semiconductor manufacturing line • Past machine difference adjustment history • Results • Reports • Results during adjustment • Maintenance reports • Equipment specification information. In this way, the update unit 16 stores information on machine differences, which indicates the contents of multiple measuring devices 21, in the Knowledge DB 13 as semiconductor-related information.

[0061] The semiconductor knowledge provision system 100 is configured to register the outputs of multiple measuring devices 21 (images, features extracted from images, device differences, maintenance information for each device, parameters before and after maintenance, etc.) as update information for the knowledge database 13. The input / output device 31G displays the database maintenance column 300 shown in Figure 17 on the screen. The input unit 11 receives input from the database maintenance column 300 regarding the equipment used for device difference adjustment and the actual work performed to reduce the device difference. The update unit 16 may also use the input data to the database maintenance column 300 as update data for the knowledge database 13.

[0062] The semiconductor knowledge provision system 100 provides information regarding the reduction of instrument differences to the questioner 31 according to the following procedure: - The input unit 11 receives input of the instrument status of each measuring device 21 that is subject to instrument differences and extracts the instrument log and adjustment flow of each measuring device 21. - The search unit 12 searches for semiconductor-related information (instrument log and adjustment flow) from the knowledge DB 13 or the storage medium of each measuring device 21 and generates an extended prompt from the search results. - The answer LLM 14 outputs the factors that worsen the instrument status (factors that increase instrument differences), or multiple estimated results of the factors, based on the extended prompt input from the search unit 12.

[0063] The following is the procedure for the second question: The questioner 31 or the semiconductor knowledge provision system 100 uses the proposed causes of machine difference deterioration output from the answer LLM 14 as input information (question) to the input unit 11 to query the system again. The search unit 12 searches the knowledge DB 13 for necessary information by referring to past machine difference adjustment history, results, reports, etc., and generates an extended prompt from the search results. The answer LLM 14 outputs proposed countermeasures based on the extended prompt input from the search unit 12. In this way, the questioner 31 can obtain information on machine difference reduction based on past information, etc.

[0064] If, for example, an existing recipe exists, questioner 31 inputs a question to the input unit 11 requesting the creation of a recipe that minimizes the machine difference compared to the target machine. This question might be something like, "Please create a recipe with the following specifications. However, please adjust it so that the measurement error is within XX nm compared to the existing recipe." Alternatively, if no existing recipe exists, questioner 31 inputs a question to the input unit 11 requesting feedback on the automatic operation and measurement results of devices A and B. This question might be something like, "We want to keep the machine difference relative to the reference machine within XX. Based on the current device information, what should we adjust?" As a result, the answering LLM 14 outputs the following: - Identify the factors causing machine differences from the adjustment flow at the time of device introduction - Present robust settings for parameters that are less prone to machine differences

[0065] Figure 19 is an explanatory diagram showing an example of application in defect inspection as the fourth application example. The semiconductor knowledge provision system 100 has a review SEM 23, an optical inspection device 24, and an etching / exposure device 25 as an update unit 16. With the miniaturization and complexity of semiconductor device patterns, the number of inspection points required of defect inspection devices (such as the review SEM 23 and the optical inspection device 24) is also increasing. As the number of inspection points increases, the cost required for semiconductor device manufacturing increases, so setting an appropriate number of inspection points is required. The semiconductor manufacturing equipment is, for example, an etching / exposure device 25 that exposes patterns on a semiconductor wafer. The occurrence of defects changes depending on the equipment conditions and the condition of the materials constituting the semiconductor wafer. Therefore, it is desirable to perform appropriate defect inspection according to the semiconductor manufacturing conditions.

[0066] When an inspector performs defect inspection using the review SEM 23, they first use the optical inspection device 24 to detect defect coordinates from the entire surface of the semiconductor wafer and save them in a predetermined file format (such as a KRF file). Then, the inspector selectively uses the review SEM 23 to inspect defects that require particular inspection from among the defects detected by the optical inspection device 24. At this time, the inspector extracts hotspots using simulations or other methods to identify the defects that need to be inspected. The inspector then analyzes whether the defects are caused by the defect generation process or design by taking images of the identified hotspots with the review SEM 23 or by performing elemental analysis based on X-ray detection based on electron beam irradiation.

[0067] The following describes a method for identifying defect locations that are candidates for inspection by the review SEM 23 using the semiconductor knowledge provision system 100. First, the update unit 16 pre-registers the following semiconductor-related information in the knowledge DB 13: ・Wafer design document (completed design data, etc.) ・Material information ・Semiconductor manufacturing process ・SEM image ・Equipment operation screen ・Defect coordinate information (files with the extension KRF, etc.) ・Elemental analysis results ・Defect information and the process in which the defect occurred ・Pattern shape ・Manufacturing conditions ・Inspection conditions when the defect was inspected In this way, the update unit 16 stores information on the coordinates where defects are predicted in semiconductor devices as semiconductor-related information in the knowledge DB 13.

[0068] The input unit 11 accepts a question from the system regarding the prediction of defect locations in order to predict the location of defects. A question regarding the prediction of defect locations is, for example, "Please predict the locations where the probability of defect occurrence is high, as can be read from the wafer design document." Based on the above question, the search unit 12 searches the knowledge database 13 for information regarding locations where the probability of defect occurrence is high due to the pattern shape and material. As a result, the search unit 12 extracts information from the knowledge database 13 about locations that are prone to defects, such as the upper and lower layers of a multilayer semiconductor wafer with close pattern locations, pattern corners, and areas with narrow line widths.

[0069] Specifically, for example, the system extracts sections of code prone to defects from the programming language used to describe the design data (stored as text data in the knowledge database 13). It is desirable to refer to other data besides the design data to determine whether or not a section is prone to defects. Therefore, it is desirable for the update unit 16 to pre-register the following semiconductor-related information (information relating pattern shape and / or material to the state of defect occurrence) in the knowledge database 13: • A list of defect occurrence probabilities according to the pattern shape (a ranked list of occurrence probabilities) • Simulator results for predicting the pattern shape based on process conditions • A list of past defects

[0070] The search unit 12 then inputs the data retrieved from the knowledge database 13 and the question data as a composite data (extended prompt) into the answer LLM 14. The extended prompt is a composite data of an inquiry (question) requesting location information with a high probability of occurrence and information about defect occurrence probability, etc., retrieved by the search unit 12. In accordance with the extended prompt, the answer LLM 14 outputs data about the predicted location of the defect as information about a location with a high probability of defect occurrence from the output unit 15.

[0071] This allows the questioner 31 to obtain information about the process in which defects occur during the manufacturing process of multiple semiconductor devices. Specifically, the questioner inputs a question such as "Please predict the process in which the abnormality XX occurred" to the input unit 11 as input information. In the case of predicting the process in which defects occur, for example, actual defect images or elemental analysis results may be attached to the question. An abnormality may be an object that does not appear in an ideal manufacturing process, such as a defect or a specific elemental analysis result.

[0072] Knowledge DB13 stores the following semiconductor-related information: • Information on materials used in each manufacturing process • Past defect occurrence information stored at the design data level • Information on defect types such as pattern breaks, disappearances, and bridges

[0073] Therefore, the search unit 12 can extract information such as process data stored in relation to the input elemental analysis results. The answer LLM 14 receives the question and the retrieved process information and other composite data from the search unit 12 and outputs an estimated result of the abnormal process. In this way, the semiconductor knowledge provision system 100, which is optimized for abnormal process analysis, makes it possible to identify the abnormal process from limited information.

[0074] In the semiconductor knowledge provision system 100 described above, instead of fine-tuning the answer LLM 14 with semiconductor-related information, the search unit 12 generates an extended prompt containing semiconductor-related information and inputs it into the answer LLM 14. This eliminates the learning cost of fine-tuning and reduces the risk of overfitting and forgetting of the learning data. Furthermore, the update unit 16 performs update processing on the knowledge DB 13 to address the difficulty of expanding domain knowledge.

[0075] Thus, the semiconductor knowledge provision system 100 functions as an AI system that supports the operation of equipment for semiconductor manufacturing processes. For example, the semiconductor knowledge provision system 100 enables interactive equipment operation that does not require knowledge or know-how of semiconductors or measurement and inspection equipment for SEM products (measurement and inspection equipment) in general. Specifically, in the creation of measurement recipes, the semiconductor knowledge provision system 100 assists in setting and creating parameters for recipes that meet the measurement accuracy and throughput required by the user.

[0076] Furthermore, the semiconductor knowledge provision system 100 can address the increased operating costs associated with the increasing complexity of semiconductor manufacturing equipment, as listed below: • Measurement recipe creation: Various parameter settings and operating procedures are required, demanding advanced knowledge and experience for stable measurement. • Recipe error response: Error response in mass production requires quick root cause analysis and corrective action within a limited timeframe. • Training costs for new employees and customers: The increasing difficulty in mastering usage necessitates training for beginners. • Complex document management: Instruction manuals, equipment logs, etc., are managed individually, making them difficult to refer to during operation. In short, the semiconductor knowledge provision system 100 centrally manages documents, assists with information extraction and operation, and improves the efficiency of equipment operation.

[0077] The accompanying drawings constitute part of this specification and illustrate one or more embodiments, and together with this specification, illustrate the subject matter of this disclosure. Herein, we refer to the embodiments shown in the drawings and use specific terms to describe those embodiments. However, it will be understood that this does not limit the scope of this disclosure. Any modifications and further alterations of the features shown herein, as well as additional applications of the principles shown herein, that a person in possession of this disclosure and familiar with the relevant art might conceive of, will be considered within the scope of this disclosure.

[0078] Furthermore, this disclosure is not limited to the embodiments described above, and various other applications and modifications are possible, provided that they do not deviate from the gist of this disclosure as described in the claims. For example, the embodiments described above describe the configuration of the semiconductor knowledge provision system 100 in detail and specifically in order to explain this disclosure in an easy-to-understand manner, and are not necessarily limited to those comprising all the components described. Also, it is possible to replace parts of the configuration of one embodiment with components of another embodiment. It is also possible to add components of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, replace, or delete other components for parts of the configuration of each embodiment.

[0079] Furthermore, some or all of the above configurations, functions, and processing units may be implemented in hardware, for example, by designing them as integrated circuits. Broadly defined processor devices such as FPGAs (Field Programmable Gate Arrays) and ASICs (Application Specific Integrated Circuits) may be used as hardware. In addition, each component of the semiconductor knowledge provision system 100 according to the above embodiment may be implemented on any hardware, as long as the respective hardware can send and receive information from each other via a network. Moreover, the processing performed by a certain processing unit may be implemented by a single piece of hardware, or by distributed processing by multiple pieces of hardware.

[0080] The following describes other examples of LLM systems equipped with RAG. Specifically, we will describe an example that utilizes a larger database by connecting to an external system and having that system perform document searching and extraction, rather than simply using documents prepared in advance as the RAG database.

[0081] Figure 20 shows another example of a semiconductor knowledge provision system. System 2001 is configured to receive data from semiconductor manufacturing plants (e.g., 1st fab 2002, 2nd fab 2003, 3rd fab 2004) where semiconductor device manufacturing equipment, measurement equipment, inspection equipment, etc. are installed, as exemplified in Figure 3. In semiconductor manufacturing plants, multiple semiconductor manufacturing equipment, measurement equipment, inspection equipment, etc. are in operation, and precise condition settings are required for each piece of equipment. Furthermore, even equipment with the same specifications may have subtle differences, and different settings may be required for each piece of equipment. In addition, the appropriate setting conditions differ depending on the materials and pattern size used in the semiconductor device.

[0082] In this embodiment, for example, equipment errors provided by multiple semiconductor manufacturing plants and countermeasures for those errors are stored in multiple databases 2005, and the answering LLM 14 selects the appropriate database from the multiple databases 2005 (for example, DB 1 to DB 5) according to the input question. The computer system 10 creates a prompt using the selected database and inputs the prompt to the answering LLM 14. The answering LLM 14 outputs an answer corresponding to the input prompt.

[0083] According to a system like the one illustrated in Figure 20, the computer system 10, for example, selects a database containing library data that includes the tool names of appropriate semiconductor measuring devices in response to a question, and then generates a prompt. This allows the computer system 10 to quickly output appropriate answers even if the database has a large capacity.

[0084] Furthermore, it is desirable to create multiple databases 2005 that store descriptions of events (e.g., errors) and the relationships and history of responses to those events, using various units (types) such as the following, and select the appropriate database according to the input of a question: • Type of event • Type of response • Type of equipment • Type of environment (temperature, region, atmospheric pressure, cleanroom class, etc.) • In the case of setting a recipe for a measuring device that measures semiconductors, the type of semiconductor device to be measured. The type of semiconductor device includes, for example, the process, the materials used in the semiconductor device, the type of manufacturing equipment for the semiconductor device, the identification information of the semiconductor device, the attributes of the design data, the type of pattern to be measured, the depth of the pattern to be measured, etc. • Type of measuring device • Type of measurement conditions (field of view size, beam energy, etc.)

[0085] Alternatively, the computer system 10 may pre-load necessary documents and other data into the context window of the answer LLM 14 and then generate answers to the questions. In this case, the computer system 10 can select data according to the words and attributes of the questions input from the input / output device 31G and selectively load it into the context window, thereby generating answers from appropriately filtered data.

[0086] The computer system 10 may, for example, selectively load data as exemplified below based on the input of a question for measuring pattern C of semiconductor device B using device A, and generate a prompt using the loaded data. • Operation manual for device A • Past operating programs for device A • Design data of the layer on which pattern C of semiconductor device B is formed

[0087] The computer system 10 should preferably select information on semiconductor devices with strong relevance and information on measurement and inspection equipment as load data. The computer system 10 may also select data according to the frequency of questions and data volume. Furthermore, the computer system 10 may select data according to the type of pattern to be measured. For example, the appropriate equipment conditions differ depending on whether the pattern is formed on the sample surface or at the bottom of a hole.

[0088] Therefore, it is desirable to obtain appropriate instrument conditions as answers, which vary depending on the type of object being measured, as illustrated below. • If the measuring instrument is an electron microscope, if the pattern is formed on the surface of the sample, set the instrument conditions so that the portion to be measured is included in the field of view. • If the pattern is formed at the bottom of a hole, set the instrument conditions so that the portion to be measured is included in the field of view, and also set the instrument conditions to irradiate with a pre-charging beam to extract secondary electrons from the bottom of the hole. To this end, if the question includes the type of pattern, it is desirable for the computer system 10 to load documents or past operating programs that describe the instrument conditions for that pattern so that the appropriate instrument conditions are set.

[0089] As described above, the computer system 10 can generate a response based on necessary and carefully selected data by selecting a database that stores semiconductor device information and equipment information such as measuring devices related to the desired measurement, and then generating a prompt. Furthermore, the computer system 10 may also select a database that stores information related not only to the measurement and inspection device, but also to the manufacturing equipment that produced the sample to be measured.

[0090] The computer system 10 may select necessary databases and information based on the input data included in the GUI screen exemplified in Figure 17. The input data includes operator ID, time, device name, event name, cause name, past maintenance response, cost and time spent on the response, and log data from past maintenance. Alternatively, a model may be prepared that associates keywords in the question with database IDs, and the computer system 10 may generate prompts using the database selected based on the input of the question into the model. In particular, the computer system 10 can obtain answers that refer to the knowledge of past experts by selecting databases and data based on the input of an expert operator's ID.

[0091] Furthermore, the computer system 10 may, regardless of the question, select a database or the like according to the type of measuring device to be used and the device to be measured, and load it into the LLM context window.

[0092] Furthermore, the computer system 10 monitors the chunked database for time, the number of times (frequency) similar operating programs are used, the volume of update data for existing recipes, and changes in question trends (for example, changes in the frequency of keywords used). When predetermined conditions are met, the computer system 10 selectively updates the differences with the existing database, enabling it to generate response statements using the selected database according to the changes in circumstances. In equipment for measuring semiconductor devices, control is performed by operating programs called recipes, but adjustments may be necessary to follow process fluctuations or temporal changes in the condition of the measuring equipment. Therefore, if there is a history of adjusting the equipment conditions to follow process fluctuations, the computer system 10 may selectively update the parts of those additional adjustments.

[0093] As described above, the computer system 10 can generate an answer using an appropriate database by performing a first-stage search based on the selection of databases and data, and a second-stage search based on prompt generation using the selected databases, etc.

[0094] Figure 21 is a flowchart showing the process of selecting a database corresponding to a question from a group of databases based on the input of a question, and generating an answer using the selected database. First, the computer system 10 receives the input of a question via the input / output device 31G (step S2101). The computer system 10 (processor) receives the input question and searches for a database (DB) corresponding to the question, or loads the necessary information from an external DB (step S2102).

[0095] The search engine used in step S2102 searches the target database based on, for example, the text data contained in the question. The computer system 10 may be configured to tokenize and vectorize at least a portion of the question and generate a vectorized query. In this way, the computer system 10 may perform the search in S2102 using the words contained in the question in S2101.

[0096] The question may also be phrased as follows: • "We want to match the measurement results of semiconductor device B pattern C measured by device D with the measurement results of device A." • "Please rewrite the recipe created by device A so that it can be used by device D (to eliminate device differences)." In such cases, the computer system 10 may use a search engine to select and load databases or data that store past recipes, past measurement results, and error logs from at least devices A and D. The computer system 10 may also search the search engine on a file-by-file basis and load the results into the context window.

[0097] Next, the computer system 10 and the response LLM 14 generate a prompt using data from the selected or retrieved database, or data loaded from an external DB (step S2103). The computer system 10 inputs the generated prompt to the response LLM 14, and the response LLM 14 outputs the response (step S2104). The computer system 10 displays the outputted response on a display device (display device) provided in the measuring device, or on a display device capable of transmitting information, so that the operator can refer to the response text.

[0098] As explained above in Figure 21, the semiconductor information provision system configured to output an answer based on the input of a question comprises one or more computer systems configured to execute program instructions, and the one or more computer systems perform the following processes: - A step of receiving the question (S2101) - A step of searching a database or data from a group of databases or data sets that store at least one of semiconductor-related information and semiconductor-related equipment information (information on semiconductor-related equipment) using the received question (S2102) - A step of generating a prompt using the retrieved database (S2103) - A step of inputting the generated prompt into the LLM and outputting an answer that is output from the LLM based on the input of the prompt into the LLM to a display device (S2104)

[0099] Furthermore, when searching the database, the computer system 10 may select and load databases based on weighting according to the following examples: • Frequency of recipe use • Time spent using recipes • Error rate of setting conditions • Similarity of semiconductor device generations and semiconductor device nodes • Similarity of materials used in semiconductor devices (similarity of physical phenomena such as static charge)

[0100] For example, a search could be conducted such that the more frequently a recipe is used, the higher the probability of it being selected. By performing such a first-stage search, the computer system 10 can generate prompts and response sentences using an appropriate database. In this way, one or more computer systems may perform the search from the database group or data group according to at least one of the frequency of use and the usage time of the database or data.

[0101] Furthermore, the computer system 10 may perform a first-stage search to increase the likelihood of selecting a recipe set by an expert operator compared with a recipe set by another operator.

[0102] Furthermore, in the case of a semiconductor manufacturing plant where the LLM has received sufficient training based on knowledge of semiconductor devices, the computer system 10 may perform the first stage of the search to retrieve necessary information from equipment data provided by the supplier of the measuring equipment.

[0103] Furthermore, the computer system 10 may visualize databases and data corresponding to the question and allow the operator to select from them. Figure 22 shows an example of a GUI screen having a question input field 232. The GUI screen includes a question input field 232 and a database candidate display field 2201. The question input field 232 is used to input questions such as, "Please rewrite the recipe created on device A so that it can be used on device B (to eliminate differences between devices)."

[0104] The database candidate display area 2201 displays the contents of the databases searched according to the keywords (device A, device B, etc.) included in the question entered in the question input area 232. For example, the database candidate display area 2201 displays the following: • The file name of the searched database • The score of the database found as a candidate • The keywords (search words) that formed the basis for the search of the searched database

[0105] The score indicates the degree to which the database displayed as a candidate is suitable as a prompt for the question, and can be determined by predetermined rules based on, for example, the number of keyword hits, the frequency of use of the recipes mentioned above, and the error rate. Alternatively, a model showing the relationship between keywords and scores may be prepared, and by inputting the question or keywords extracted from the question into this model, the computer system 10 may output a score (an evaluation index for the suitability of the database).

[0106] Furthermore, the computer system 10 can determine the appropriateness of selecting a query for an existing database by displaying the keyword information that served as the basis for the database search.

[0107] The GUI screen illustrated in Figure 22 is configured to allow the questioner to select the database that will serve as the basis for generating the prompt. However, to automate the selection process, a predetermined threshold may be set, and the database may be selected only when the score is equal to or exceeds the threshold.

[0108] Furthermore, the computer system 10 may not simply display the database ID and file name, but may also display the text that served as the basis for the selection.

[0109] Furthermore, the computer system 10 may be configured to add databases that were not found in the search but which it can determine, based on experience, should be used as a basis for generating prompts. For this purpose, the database candidate display area 2201 may be configured to allow users to drag and drop file icons, files, or text files.

[0110] The computer system 10 may categorize the prompt creation databases into levels, designating some databases to be searched and others to be excluded from the search.

[0111] Furthermore, the computer system 10 may also generate prompts and create response texts using a database that is entered into and registered in a maintenance report, as illustrated in Figure 17. This allows the user to receive advice based on past operational data. In particular, multiple related devices are registered, as illustrated in Figure 17, and the computer system 10 inputs measures to suppress differences between them. This allows the user to obtain a response text that describes how to configure the devices to reduce differences between different devices.

[0112] 10 Computer System 11 Input Unit (Input Subsystem) 12 Search Unit (Search Subsystem, RAG Engine) 13 Knowledge DB (Database) 13D Knowledge Documents 14 Answer LLM 15 Output Unit (Output Subsystem) 16 Update Unit (Update Subsystem) 18 Network 19P Processor 19M Memory 19C Program Instructions 21 Measurement Device (Semiconductor-related Device) 21A First Measurement Device (Semiconductor-related Device) 21B Second Measurement Device (Semiconductor-related Device) 24A Test Device (Semiconductor-related Device) 26 Design DB (Semiconductor-related Device) 27A First Process Device (Semiconductor-related Device) 27B Second Process Device (Semiconductor-related Device) 28 Simulator (Semiconductor-related Device) 100 Semiconductor Knowledge Provision System

Claims

1. A semiconductor knowledge provision method performed by a semiconductor knowledge provision system, wherein the semiconductor knowledge provision system comprises an input subsystem, a search subsystem, an LLM (Large Language Model), and an output subsystem; the input subsystem receives input of question information relating to semiconductor-related equipment; the search subsystem queries a database storing semiconductor-related information about the semiconductor-related equipment to obtain semiconductor-related information similar to the question information; a RAG (Retrieval Augmented Generation) engine generates an augmented prompt associating the question information received by the input subsystem with the obtained semiconductor-related information; the LLM generates answer information to the question information by referring to the semiconductor-related information about the semiconductor-related equipment based on the generated augmented prompt; and the output subsystem outputs the generated answer information.

2. The semiconductor knowledge provision method according to claim 1, characterized in that the search subsystem acquires semiconductor-related information for at least one of the following semiconductor-related devices: a semiconductor manufacturing device, a semiconductor device manufactured by the semiconductor manufacturing device, a semiconductor measuring device for measuring the semiconductor device, and a semiconductor testing device for inspecting the semiconductor device.

3. The semiconductor knowledge provision method according to claim 2, characterized in that the search subsystem acquires at least one of the following as semiconductor-related information to be obtained from the database: the operating program of the semiconductor-related device, the setting conditions set for the semiconductor-related device, the documentation describing how to handle the semiconductor-related device, the instruction information for the semiconductor-related device, the information of semiconductor devices handled by the semiconductor-related device, the output information from the semiconductor-related device, and the events that have occurred in the past with the semiconductor-related device and the corresponding information for those events.

4. The semiconductor knowledge provision method according to claim 1, characterized in that the LLM has pre-tuned its knowledge for interpreting question information or semiconductor-related information.

5. The semiconductor knowledge provision method according to claim 1, characterized in that the LLM generates information including information on the source of the semiconductor-related information referenced as answer information to the question information, and when the output subsystem outputs the generated answer information, it also outputs information on the source of the information that forms the basis of the answer information.

6. The semiconductor knowledge provision system further comprises an update subsystem, the update subsystem generates chunks of a predetermined size from the semiconductor-related information, and updates the contents of the database differentially on a chunk-by-chunk basis, as described in claim 1.

7. The semiconductor knowledge provision method according to claim 6, characterized in that the update subsystem generates a chunk vector by vectorizing each generated chunk, and updates the contents of the database differentially on a chunk vector basis instead of differential updating on a chunk basis.

8. The semiconductor knowledge provision system further comprises an update subsystem, wherein the update subsystem stores a recipe file indicating the operation of the semiconductor-related equipment for manufacturing semiconductor devices in the database as semiconductor-related information, characterized in that it is the semiconductor knowledge provision method according to claim 1.

9. The semiconductor knowledge provision method according to claim 1, wherein the semiconductor knowledge provision system further has an update subsystem, and the update subsystem stores a troubleshooting history showing the content of past troubleshooting as semiconductor-related information in the database.

10. The semiconductor knowledge provision system further comprises an update subsystem, the update subsystem storing device difference information indicating the contents of a plurality of measuring devices in the database as semiconductor-related information, characterized in that the semiconductor knowledge provision method according to claim 1.

11. The semiconductor knowledge provision system further comprises an update subsystem, the update subsystem storing information of coordinates where defects in a semiconductor device are predicted in the database as semiconductor-related information, characterized in that the semiconductor knowledge provision method according to claim 1.

12. A semiconductor information provision system comprising: an input subsystem that receives input of question information regarding semiconductor-related equipment; a search subsystem that queries a database storing semiconductor-related information about the semiconductor-related equipment to obtain semiconductor-related information similar to the question information, and generates an extended prompt by a RAG engine that associates the question information received by the input subsystem with the obtained semiconductor-related information; an LLM that generates answer information to the question information by referring to the semiconductor-related information about the semiconductor-related equipment based on the generated extended prompt; and an output subsystem that outputs the generated answer information.

13. A semiconductor information provision system configured to output an answer based on the input of a question, comprising one or more computer systems configured to execute program instructions, wherein the one or more computer systems receive the question, use the received question to search a database or data from a group of databases or data sets in which at least one of semiconductor-related information and semiconductor-related device information is stored, generate a prompt using the retrieved database, input the generated prompt to an LLM, and output an answer based on the input of the prompt to the LLM to a display device.

14. The semiconductor information provision system according to 13, characterized in that the one or more computer systems perform the search from the database group or data group according to at least one of the frequency of use and the usage time of the database or data.

15. The semiconductor information provision system according to claim 13, characterized in that the one or more computer systems perform the search using the words contained in the question.

Images