Conversational Recipe establishment system and method based on artificial intelligence

Abstract

The invention provides a dialogue type Recipe establishing system and method based on artificial intelligence, and belongs to the field of nanoscale semiconductor processing and wafer detection, the dialogue type Recipe establishing system comprises a user operation interface and a tool interface which support information input of dialogue content, model parameter editing and interactive operation; the machine table module comprises a wafer carrying table and an optical machine unit; and the AI model control module comprises a knowledge retrieval enhancement module, a text embedding model and a large language model and is used for establishing a Recipe, querying a Recipe establishment state based on an embedded structured instruction interface and confirming execution of an instruction by adopting dual verification. According to the method, Recipe system logic established by operators can be simplified, Recipe can be established efficiently and quickly, the influence of expert experience of the operators on Recipe quality is reduced, and the method is convenient to popularize and apply in the fields of nanoscale semiconductor manufacturing, optical wafer detection and the like.

Description

Artificial Intelligence-Based Conversational Recipe Creation System and Method Technical Field

[0001] This invention belongs to the field of nanoscale semiconductor processing and optical quantity detection, specifically relating to an artificial intelligence-based conversational recipe creation system and method. Background Technology

[0002] In the field of semiconductor recipe creation, the expert experience of operators is crucial. How to reduce the impact of expert experience, use AI models to guide users, and improve operator efficiency is a key challenge in recipe creation. In recent years, with the development of large language models, deep learning-based natural language processing models have become an important foundation for intelligent agents, demonstrating excellent language capabilities in areas such as text generation. However, the knowledge of large language models mainly comes from training data, which can lead to factual errors and outdated updates when dealing with content based on expert experience.

[0003] The core idea of ​​knowledge retrieval enhancement is "external knowledge" + "model reasoning," so knowledge retrieval enhancement combined with a large language model can provide an efficient and stable technical solution for intelligent information systems. However, knowledge retrieval enhancement combined with a large language model can only provide professional knowledge question answering and cannot execute code or control system execution. Therefore, there are problems such as the recipe's over-reliance on user expert experience, frequent camera movements and switching of camera magnifications during the routine process of creating a recipe, and difficulties in finding Mark point coordinates. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide an artificial intelligence-based conversational recipe creation system and method that can solve the above-mentioned problems.

[0005] Design principle: By combining Function Calls, the model can automatically perform tasks, making it a system with perception, reasoning, and action.

[0006] An AI-based conversational recipe creation system includes: an interactive terminal, comprising a user interface and tool interface, supporting information input for dialogue, editing of model parameters, and interactive operations; a machine module, comprising a wafer stage and an optomechanical unit; and an AI model control module, comprising a knowledge retrieval enhancement module, a text embedding model, and a large language model, used to create recipes and query the recipe creation status based on an embedded structured instruction interface and use double verification to confirm the execution of instructions.

[0007] Furthermore, the user interface on the interactive end provides a graphical user interface (GUI) or a web interface, supporting natural language interaction between users and the system. It displays the Recipe creation status, image acquisition results, and system operation logs in real time. Through the tool interface, it supports input information including text, voice, images, and video. The user interface provides visual operation prompts, including parameter editing and selection guidance, operation step prompts, and error message feedback. It also supports users to confirm or reject the system's suggested operations, realizing human-machine collaborative decision-making.

[0008] Furthermore, the machine module responds to the interactive terminal and the AI ​​model control module to control the movement of the wafer stage and the switching of optomechanical parameters.

[0009] Furthermore, the knowledge retrieval enhancement module includes multiple operation units: Mark point features, Test mode selection, Alignment, SmartCA, Premap, WaferMap, LightTrain, and PostProcess, which clean and slice the UserManual document into structured text fragments chunks.

[0010] Furthermore, the text embedding model converts structured text fragments into vector representations and stores them in a vector library. It generates query vectors based on user queries and performs nearest neighbor retrieval in the vector library. After reordering the vectorized text fragments, it adds them to the prompt and inputs them into the large language model.

[0011] Furthermore, the large language model includes a semantic understanding unit, a call parameter production unit, and an external triggering unit. The semantic understanding unit performs retrieval and inference based on vectorized text fragments, the call parameter production unit generates formatted call parameters based on the inference results, and the external triggering unit triggers the invocation of external functions or service execution instructions based on the call parameters.

[0012] Furthermore, the external functions called include image similarity matching algorithms, controlling the movement of the machine, switching camera magnification, and querying the recipe creation status, all deployed via WebAPI.

[0013] Furthermore, the external triggering unit supports dual verification for function calls that control the movement of the machine and switch camera magnification, and the operation is only executed upon user confirmation.

[0014] This invention also provides an AI-based conversational recipe creation method, which includes: a knowledge retrieval enhancement module creating a UserManual document based on an operation unit, cleaning and slicing it into structured text fragments (chunks); a text embedding model vectorizing and storing the structured text fragments, reordering the vectorized text fragments and merging them into a prompt, which is then input into a large language model; the large language model constructing the recipe, querying the recipe creation status based on the embedded structured instruction interface, and using double verification to confirm the execution of the instructions.

[0015] Furthermore, the large language model constructs a recipe by inputting the basic parameters of the recipe, selecting a point for alignment, drawing a Wafer Map, and establishing the region to be detected.

[0016] Compared to existing technologies, the advantages of this invention are as follows: This application simplifies the dialogic recipe creation system based on operator expert experience, enabling efficient and rapid recipe creation and reducing the impact of operator expert experience on recipe quality. The solution provided by this application is suitable for widespread application in fields such as nanoscale semiconductor manufacturing and optical inspection. Attached Figure Description

[0017] Figure 1 is a schematic diagram of the process of establishing the dialogic Recipe of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] An AI-based conversational recipe creation system addresses the over-reliance on user expert experience in traditional recipe creation systems and the difficulties users face in frequently moving equipment, switching camera magnifications, and finding mark point coordinates during routine recipe creation. As shown in Figure 1, the conversational recipe creation system includes an interactive interface, an equipment module, and an AI model control module. The interactive interface includes a user interface and tool interfaces, supporting information input for dialogue, model parameter editing, and interactive operations. The equipment module includes a wafer stage and an optomechanical unit. The AI ​​model control module is used to create recipes and queries the recipe creation status based on an embedded structured instruction interface, employing double verification to confirm instruction execution.

[0020] The user interface on the interactive end provides a graphical user interface (GUI) or a web interface, supports natural language interaction between users and the system, displays the recipe creation status, image acquisition results, and system operation logs in real time, and supports input information including text, voice, images, and video through the tool interface. The user interface provides visual operation prompts, including parameter editing and selection guidance, operation step prompts, and error message feedback. It also supports users to confirm or reject the system's suggested operations, realizing human-machine collaborative decision-making.

[0021] The machine module responds to the interactive terminal and the AI ​​model control module, controlling the movement of the wafer stage and the switching of optomechanical parameters.

[0022] The AI ​​model control module includes a knowledge retrieval enhancement module, a text embedding model, and a large language model.

[0023] Specifically, the knowledge retrieval enhancement module includes multiple operation units: Mark point features, Test mode selection, Alignment, SmartCA, Premap, WaferMap, LightTrain, and PostProcess. These operation units are used to create User Manual documents for the Recipe, and to clean and slice the Recipe into structured text chunks.

[0024] Mark point features: Select feature points, record the coordinate information of the feature points and template image.

[0025] Test mode selection: In Test mode, different combinations can be set to detect different needs.

[0026] Alignment: Alignment is a key step in the semiconductor inspection process. It refers to the process of determining specific reference points (alignment sites) on the wafer. These reference points are used by the system to correctly position the wafer on the Chuck (wafer stage).

[0027] SmartCA: Automatically extracts ROIs from GDS files using custom rules, and imports the automatically generated ROIs into Recipe through coordinate system transformation.

[0028] Premap: Premap points are two selected positioning points on the wafer. During the inspection process, the random imaging computer of the inspection system uses these two positions to determine the precise coordinates and spacing parameters of the first two groups of chips within the sampling plan area.

[0029] WaferMap: WaferMap is used to define the layout, size, effective area, and numbering system of chips (dies) on a wafer.

[0030] LightTrain: LightTrain is an optical optimization module in semiconductor inspection systems, primarily used to optimize the contrast of the ROI (Region of Interest) during inspection. By training within a set grayscale range, the system can determine the optimal lighting conditions, thereby improving the accuracy and reliability of defect detection.

[0031] PostProcess: Primarily used for further analysis, classification, and optimization of detected defect data to improve the quality and usability of defect detection results. According to the knowledge base, this module is not mandatory, but it can significantly improve the efficiency and accuracy of defect analysis.

[0032] The text embedding model converts structured text fragments into vector representations and stores them in a vector library. It generates query vectors based on user queries and performs nearest neighbor retrieval in the vector library. After reordering the vectorized text fragments, it adds them to the prompt and inputs them into the large language model.

[0033] The large language model includes a semantic understanding unit, a call parameter production unit, and an external triggering unit. The semantic understanding unit performs retrieval and inference based on vectorized text fragments. The call parameter production unit generates formatted call parameters based on the inference results. The external triggering unit triggers the invocation of external functions or service execution instructions based on the call parameters.

[0034] The large language model is based on a structured instruction interface called by Function Calls, and it reasones and responds to interactive questions. Specifically, Function Calls enable the large language model to automatically generate call parameters in a specific format based on the results of semantic understanding, and trigger the execution of external functions or services to extend its reasoning and operational capabilities. The function calls in this invention mainly include image similarity matching algorithms, automatic control of machine movement, switching camera magnification, querying recipe creation status, etc., and are deployed via Web API. The large language model's responses: The retrieved content includes the user's question and fragments retrieved from documents. Based on the retrieved content and its own reasoning ability, if the query only involves professional knowledge, the large language model will directly output an answer that fits the context. However, if it involves operating the machine, such as marking points or automatically finding high-magnification points, the user can use dialogue to allow the AI ​​to automatically call the software's internal interface. The large language model will query the recipe creation status from the software system and then provide a response. The large language model performs double verification; only actions confirmed by the user will be automatically executed.

[0035] Specific function calls include: the stage movement function: StageMove(left); and the function to switch camera magnification: SwitchOptics(high / low). Of course, there are other related hardware control functions, which will not be listed here.

[0036] The external functions called include image similarity matching algorithms, controlling the movement of the machine, switching camera magnification, and querying the recipe creation status. The entire system is deployed via WebAPI.

[0037] The external trigger unit supports double verification for function calls that control the movement of the machine and switch camera magnification, and the operation is only executed after user confirmation.

[0038] The present invention also provides an artificial intelligence-based conversational recipe creation method, which includes the following:

[0039] The knowledge retrieval enhancement module creates UserManual documents based on operation units, and cleans and slices them into structured text fragments (chunks).

[0040] The text embedding model vectorizes and stores structured text fragments, then reorders the vectorized text fragments and adds them to the prompt, which is then input into the large language model.

[0041] The large language model constructs the recipe, and the status of the recipe is established by querying the embedded structured instruction interface and double-checking to confirm the execution of the instructions.

[0042] The process of constructing a large language model recipe includes inputting the basic parameters of the recipe, selecting a point for alignment, drawing a Wafer Map, and establishing the region to be detected.

[0043] The overall inventive points are as follows.

[0044] (1) Establish a knowledge retrieval enhancement system.

[0045] The recipe is documented as a UserManual, the chunks are cleaned, and the chunks are converted into vectors using Bge-Large-zh, which are then written to a vector database. Embeddings are generated for user queries, nearest neighbor searches are performed in the vector database, and then the vectors are reordered. Finally, the text fragments are incorporated into the prompt and passed to the large language model.

[0046] (2) Structured instruction interface Function Call.

[0047] Function Call is a structured instruction interface that enables large language models to automatically generate call parameters in a specific format based on the results of semantic understanding, and trigger the execution of external functions or services to extend their reasoning and operational capabilities. The function call implementation in this invention mainly includes image similarity matching algorithms, automatic control of machine movement, switching camera magnification, querying recipe creation status, etc., and is deployed via Web API.

[0048] (3) Large language model response.

[0049] The retrieved content includes user questions and document snippets. The large language model, based on the retrieved content and its reasoning ability, will directly output a context-appropriate answer if the query only involves professional knowledge. However, if it involves operating a device, such as marking points or automatically finding high-magnification points, the user can use dialogue to allow the AI ​​to automatically call the software's internal interfaces. The large language model will then query the recipe's creation status from the software system and provide a response. The large language model will perform double verification; only actions confirmed by the user will be automatically executed.

[0050] (4) Recipe creation process.

[0051] The main process of recipe creation can be divided into four parts: inputting the basic parameters of the recipe, selecting points for alignment, drawing the wafer map, and establishing the area to be detected. Point selection is highly dependent on the operator's expert experience and familiarity with the wafer. For example, the alignment module requires the user to mark four points (high and low magnification). This invention only requires the user to mark one point. Then, the machine automatically moves within the current field of view, uses an algorithm to automatically match the image of the user-marked point, finds an image with high similarity, stops moving, and the dialogue model pops up a link to prompt the user that the current node has been found and asks if the user wants to select that point for alignment, greatly simplifying the user's operation of moving the machine and finding points.

[0052] In summary, this application integrates external knowledge, model reasoning, and Function Calls, enabling the model to automatically execute tasks. This transforms the model into a system with perception, reasoning, and action capabilities, allowing for efficient and rapid recipe creation and reducing the impact of operator expertise on recipe quality. It is particularly suitable for widespread application in the semiconductor nanoscale measurement or manufacturing fields.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A conversational recipe creation system based on artificial intelligence, characterized in that, The conversational recipe creation system includes: an interactive terminal, comprising a user interface and tool interface, supporting information input for dialogue, model parameter editing, and interactive operations; a machine module, comprising a wafer stage and an optomechanical unit; and an AI model control module, comprising a knowledge retrieval enhancement module, a text embedding model, and a large language model, used to create recipes and query the recipe creation status based on the embedded structured instruction interface and use double verification to confirm the execution of instructions.

2. The dialogic Recipe creation system according to claim 1, characterized in that: The user interface on the interactive end provides a graphical user interface (GUI) that supports natural language interaction between users and the system. It displays the recipe creation status, image acquisition results, and system operation logs in real time. The tool interface supports input information including text, voice, and images. The user interface provides visual operation prompts, including parameter editing and selection guidance, operation step prompts, and error message feedback. It also supports users to confirm or reject the system's suggested operations, enabling human-machine collaborative decision-making.

3. The dialogic Recipe creation system according to claim 1, characterized in that: The instrument module responds to the interactive terminal and the AI ​​model control module, controlling the movement of the wafer stage and the switching of optomechanical parameters.

4. The dialogic Recipe creation system according to claim 1, characterized in that: The knowledge retrieval enhancement module includes multiple operation units: Mark point features, Test mode selection, Alignment, SmartCA, Premap, WaferMap, LightTrain, and PostProcess, which clean and slice the UserManual document into structured text fragment chunks.

5. The conversational Recipe creation system according to claim 4, characterized in that: The text embedding model converts structured text fragments into vector representations and stores them in a vector library. It generates query vectors based on user queries and performs nearest neighbor retrieval in the vector library. After reordering the vectorized text fragments, it adds them to the prompt and inputs them into the large language model.

6. The dialogic Recipe creation system according to claim 5, characterized in that: The large language model includes a semantic understanding unit, a call parameter production unit, and an external triggering unit. The semantic understanding unit performs retrieval and inference based on vectorized text fragments. The call parameter production unit generates formatted call parameters based on the inference results. The external triggering unit triggers the invocation of external functions or service execution instructions based on the call parameters.

7. The conversational Recipe creation system according to claim 6, characterized in that: The external functions called include image similarity matching algorithms, controlling the movement of the machine, switching camera magnification, and querying the recipe creation status. The entire system is deployed via WebAPI.

8. The dialogic Recipe creation system according to claim 7, characterized in that: The external trigger unit supports double verification for function calls that control the movement of the machine and switch camera magnification, and the operation is only executed after user confirmation.

9. A conversational recipe creation method based on artificial intelligence, characterized in that, The method includes: a knowledge retrieval enhancement module creates a UserManual document based on the operation unit, cleans and slices it into structured text fragments chunks; a text embedding model vectorizes and stores the structured text fragments, reorders the vectorized text fragments and adds them to the prompt, and inputs them into the large language model; the large language model constructs a recipe, queries the recipe's status based on the embedded structured instruction interface, and uses double verification to confirm the execution of the instructions.

10. The dialogic recipe creation method according to claim 9, characterized in that: The process of constructing a large language model recipe includes inputting the basic parameters of the recipe, selecting a point for alignment, drawing a Wafer Map, and establishing the region to be detected.

Images