Method and system for automatically inspecting semiconductor factory by intelligent agent

Abstract

The invention provides a method and a system for automatically inspecting a semiconductor factory by an intelligent agent, which are applied to the technical field of semiconductor manufacturing, and the method comprises the following steps: firstly, constructing and configuring an inspection intelligent agent at least comprising an equipment interface, a knowledge base and a workflow engine through an intelligent agent configuration platform; then, in a 2.5 D on-line semiconductor factory where physical factory layout and equipment are mapped, the intelligent agent is driven to execute the defined inspection process, and sensor data are obtained in real time; and finally, the workflow engine controls the intelligent agent to carry out anomaly recognition and root cause analysis on the data, and an inspection report is automatically generated in combination with a knowledge base, so that 7 * 24-hour uninterrupted intelligent and automatic factory inspection is realized, a large amount of vigor of engineers originally invested in repeated inspection is effectively released, and the inspection efficiency is improved. According to the method, engineers are enabled to concentrate on key areas such as high-value decision analysis and process optimization, the labor cost is remarkably reduced, the operation efficiency is improved, and stable operation of a semiconductor factory is ensured.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing technology, specifically to a method and system for an intelligent agent to automatically inspect a semiconductor factory. Background Technology

[0002] In modern intelligent manufacturing, the stable operation of factory equipment is crucial to ensuring production efficiency and product quality. Therefore, continuous and efficient inspection of production equipment is essential.

[0003] Traditional inspection methods mainly rely on manual inspections. While they can complete basic inspections, their limitations are becoming increasingly apparent: they are highly dependent on the personal experience, sense of responsibility, and physical condition of engineers, and are prone to missed inspections and misjudgments due to fatigue, emotions, and other factors. Furthermore, the recording standards are inconsistent, and the high labor costs result from investing a large number of senior engineers in repetitive, low-value-added tasks, making it difficult to achieve the goal of high-frequency, comprehensive inspections. In addition, manual inspections rely heavily on sensory observation and simple instrument readings, making it difficult to conduct multi-parameter correlation analysis and trend prediction, which makes it difficult for individual experts' tacit knowledge to be accumulated and passed on.

[0004] To improve inspection efficiency, the industry has gradually introduced semi-automated inspection systems. Although these systems can automatically collect data and trigger alarms, they are still at the level of passive response. They often experience a flood of alarms but lack effective context judgment and root cause analysis, requiring a lot of manual intervention for identification. In addition, semi-automated systems are often information silos with rigid inspection logic, making it difficult to flexibly adjust with changes in production processes, equipment, or formulas. The required knowledge base and rule base also rely on manual maintenance and updates, resulting in slow response and a lack of self-learning capabilities.

[0005] Therefore, a new solution is needed for intelligent agents to automatically inspect semiconductor factories. Summary of the Invention

[0006] In view of this, embodiments of this specification provide a method and system for intelligent agents to automatically inspect semiconductor factories. Through an artificial intelligence large model system, real-time dynamic inspection of semiconductor smart factories is realized, helping engineers to achieve automated processes of rapid online inspection, rapid discovery, rapid handling, and post-event attribution. This frees up a lot of engineers' energy, allowing them to focus on key areas, improving shift efficiency and ensuring the stable operation of semiconductor smart factories.

[0007] The embodiments in this specification provide the following technical solutions: This specification provides an embodiment of a method for an intelligent agent to automatically inspect a semiconductor factory, including: Based on the intelligent agent configuration platform, an inspection intelligent agent is built and deployed, and a device interface plugin, a knowledge retrieval base and an inspection workflow engine are configured for the inspection intelligent agent. The inspection task process is defined through the inspection workflow engine. The inspection agent is driven to perform inspection tasks in an online semiconductor factory based on a 2.5D three-dimensional model, wherein the online semiconductor factory maps the layout, equipment location, and equipment data interface of the physical factory. The device data interface is invoked through the device interface plugin to obtain real-time sensor data; The inspection workflow engine controls the inspection agent to perform anomaly pattern recognition and root cause analysis on the real-time sensor data, and generates an inspection report by combining the knowledge retrieval base.

[0008] This specification also provides an embodiment of a system for intelligent agents to automatically inspect semiconductor factories, the system comprising: The inspection intelligent agent construction module is used to build and deploy inspection intelligent agents based on the intelligent agent configuration platform, and to configure the inspection intelligent agents with device interface plugins, knowledge retrieval base and inspection workflow engine; The workflow definition module is used to define the inspection task process through the inspection workflow engine. The task execution module is used to drive the inspection agent to perform inspection tasks in an online semiconductor factory based on a 2.5D three-dimensional model, wherein the online semiconductor factory maps the layout, equipment location and equipment data interface of the physical factory. The data acquisition module is used to call the device data interface through the device interface plugin to acquire real-time sensor data; The report generation module is used to control the inspection agent through the inspection workflow engine to perform abnormal pattern recognition and root cause analysis on the real-time sensor data, and generate an inspection report in conjunction with the knowledge retrieval base.

[0009] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least: By constructing a 2.5D online digital twin environment that precisely maps to the physical semiconductor factory, and deploying an intelligent agent integrating equipment interfaces, knowledge bases, and workflow engines, 24 / 7 automated and intelligent inspection of production equipment is achieved. The system can proactively identify abnormal patterns and perform root cause analysis, generating structured inspection reports. This significantly improves inspection efficiency and consistency, effectively avoids human error and omissions, and reduces repetitive labor, allowing human resources to focus on high-value decision analysis and process optimization. Furthermore, through a continuous learning and optimization mechanism, the system is able to evolve with the business, fundamentally solving the inherent defects of traditional manual and semi-automated inspection methods in terms of reliability, efficiency, and intelligence. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a method for an intelligent agent to automatically inspect a semiconductor factory, as described in this application; Figure 2 This is a flowchart illustrating the application of the intelligent agent in semiconductor factory inspection in this application. Detailed Implementation

[0012] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0013] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0015] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0016] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that practice can be carried out without these specific details.

[0017] Although the overall automation level of semiconductor factories has improved significantly, bottlenecks still exist in the critical link of equipment inspection. On the one hand, semiconductor manufacturing has extremely high requirements for equipment stability, and even minor anomalies can lead to the scrapping of an entire batch of wafers, resulting in significant losses. On the other hand, traditional manual inspection methods rely heavily on the personal experience of engineers, resulting in low efficiency, poor consistency, and a tendency to miss or misjudge. Existing semi-automated inspection solutions are still at the stage of passive alarm response and lack root cause analysis capabilities. As a result, experienced engineers have to devote a lot of energy to repetitive inspection and alarm identification work, and cannot focus on high-value decision analysis and process optimization.

[0018] In light of this, the inventors, through research and improvement, discovered that the expert experience possessed by senior engineers—such as how to correlate alarms, trace root causes, and handle rare faults—is stored in their minds and is difficult to effectively replicate and systematically pass on. Furthermore, an expert cannot work 24 / 7 or be present in multiple workshops simultaneously; their abilities cannot be replicated or expanded, thus becoming a key bottleneck restricting further increases in production capacity.

[0019] The inventors further discovered in their research on inspection efficiency that although the semi-automated inspection solutions currently used in the industry have improved efficiency, they are mostly information silos. They cannot collaboratively process multi-source information and make in-depth reasoning and decisions like human experts. Moreover, the inspection logic is rigid, unable to flexibly respond to new faults, and lacks the ability to learn autonomously from historical experience, causing the system's capabilities to gradually age with technological iterations.

[0020] Based on this, the embodiments of this specification propose a method for intelligent agents to automatically inspect semiconductor factories. This method embeds expert experience into the system, making the system itself a replicable and scalable expert. The overall approach is as follows: First, an inspection agent with at least device interfaces, a knowledge base, and a workflow engine is built and configured through an intelligent agent configuration platform. Then, in a 2.5D online semiconductor factory that maps the physical factory layout and equipment, the agent is driven to execute a defined inspection process and acquire sensor data in real time. Finally, the workflow engine controls the agent to perform anomaly identification and root cause analysis on the data, and automatically generates inspection reports based on the knowledge base. This achieves 24 / 7 uninterrupted intelligent and automated factory inspection, effectively freeing up engineers from the significant energy previously spent on repetitive inspections. Engineers can focus on high-value decision analysis, process optimization, and other key areas, significantly reducing labor costs, improving operational efficiency, and ensuring the stable operation of the semiconductor factory.

[0021] The technical solutions provided by the various embodiments of this application are described below with reference to the accompanying drawings.

[0022] like Figure 1 As shown in the embodiments of this specification, a method for an intelligent agent to automatically inspect a semiconductor factory is provided, including: Step S100: Based on the intelligent agent configuration platform, build and deploy the inspection intelligent agent, and configure the equipment interface plugin, knowledge retrieval base and inspection workflow engine for the inspection intelligent agent.

[0023] Specifically, the intelligent configuration platform abstracts and encapsulates the capabilities of the agent into a series of reusable functional modules, including but not limited to "Tool Invocation", "RAG Knowledge Retrieval", and "Workflow Engine". Users can quickly build dedicated agents with specific inspection capabilities for different process areas in the semiconductor factory through the intelligent configuration platform.

[0024] Step S200: Define the inspection task flow through the inspection workflow engine.

[0025] Specifically, a logical sequence is set for the intelligent agent to perform tasks. This sequence includes at least a complete process of data collection, abnormal pattern recognition, root cause inference, and generation of preliminary handling suggestions, forming a standardized inspection path that can be executed autonomously.

[0026] Step S300: Drive the inspection agent to perform inspection tasks in the online semiconductor factory based on a 2.5D three-dimensional model, wherein the online semiconductor factory maps the layout, equipment location and equipment data interface of the physical factory.

[0027] In practice, in the complex environment of a semiconductor factory, a purely flat 2D map is difficult to intuitively express the spatial relationships and three-dimensional appearance of equipment, which is not conducive to rapid positioning and status recognition. In order to solve the efficiency problem of rapid inspection in semiconductor factories, for repetitive and target-specific inspection tasks, a 2.5D three-dimensional modeling method is used to construct an online semiconductor factory, realizing a complete mapping of the spatial layout of the physical factory, the physical location of equipment and its data interfaces. This enables the intelligent agent to move spatially along a preset inspection route in the virtual environment, and accurately locate and access the corresponding equipment data interface based on the equipment location information, realizing automated inspection operations that link the virtual and real worlds.

[0028] Step S400: Call the device data interface through the device interface plugin to obtain real-time sensor data.

[0029] During implementation, the Agent is given a professional persona and integrated with the RAG knowledge base to query SOPs and historical cases in real time. At the same time, it is equipped with device interface plugins to obtain real-time sensor data, providing a data foundation for subsequent anomaly pattern recognition and root cause analysis.

[0030] Step S500: Control the inspection agent through the inspection workflow engine to perform abnormal pattern recognition and root cause analysis on the real-time sensor data, and generate an inspection report by combining the knowledge retrieval base.

[0031] Specifically, the inspection workflow engine controls the inspection agent to automatically process the collected real-time sensor data. First, it performs anomaly pattern recognition to detect deviations in equipment operating status, and then initiates a root cause analysis process based on the recognition results. During this process, the agent accesses the knowledge retrieval base in real time to obtain relevant knowledge such as equipment manuals, process formulas, and historical failure cases. It then performs correlation analysis between real-time data and domain knowledge, and finally automatically generates a structured inspection report containing anomalies, root cause inferences, and handling suggestions.

[0032] This application achieves standardized, 24 / 7 operation without discrimination through intelligent agents, strictly follows preset logic and standards, completely eliminates human error, and ensures extremely high consistency and reliability of inspection results.

[0033] This application frees human resources from repetitive labor, allowing them to focus on high-value decision analysis and process optimization. At the same time, the coverage and speed of a single inspection far exceed those of manual labor, achieving significant reductions in labor costs and increased operational efficiency.

[0034] This application possesses cognitive and reasoning capabilities, enabling it to proactively identify abnormal patterns, perform correlation analysis, and provide preliminary diagnostic suggestions. It achieves a leap from "passive alarm" to "proactive diagnosis," greatly reducing the cognitive load on personnel.

[0035] In some embodiments, such as Figure 2 As shown, the construction and deployment of the inspection intelligent agent includes: Configure a large language model for the agent, wherein the method for constructing the large language model includes: An evaluation system for the semiconductor manufacturing field is constructed. Based on process control logic, equipment interaction instructions and measurement data analysis capabilities, multiple base-based large language models are quantitatively scored, and the target base-based large language model is selected based on the scoring results. By utilizing a domain corpus covering the entire wafer manufacturing process, the target base large language model is pre-trained and fine-tuned to obtain a domain expert model for the semiconductor manufacturing field.

[0036] Specifically, in the process of constructing a dedicated Large Language Model (LLM) for semiconductor manufacturing, a base LLM model evaluation system specifically designed for semiconductor manufacturing is first built. This system not only evaluates the performance of general tasks, but also focuses on quantitatively scoring the accuracy and reliability of highly specialized tasks such as process control logic, the rigor of equipment interaction instructions, and measurement data analysis. By performing the above benchmark tests on multi-source public network base models, the model that performs best in structured data understanding and logical chain reasoning is finally selected as the base for domain adaptation, providing a reliable foundation with a high starting point for subsequent domain-specific injection.

[0037] Subsequently, deep domain knowledge fusion was performed on the selected base LLM to create a domain corpus covering the entire wafer manufacturing process, including equipment manuals, process recipes, SOPs, fault code libraries, and yield analysis reports. A strategy combining continuous pre-training and instruction fine-tuning was adopted to enable the model to deeply understand the professional terminology, causal logic, and decision-making processes within the semiconductor fab, ultimately obtaining a domain expert model that can accurately understand and respond to professional questions.

[0038] In the above embodiments, the high-quality base models selected through a professional evaluation system, after being deeply integrated with domain knowledge covering the entire wafer manufacturing process, significantly improve their cognitive and decision-making capabilities in semiconductor manufacturing scenarios. This enables them to accurately understand equipment control commands, analyze process data, and perform reliable logical reasoning, providing the intelligent inspection system with core capabilities of semantic understanding and problem analysis at the level of domain experts.

[0039] In some embodiments, the instruction fine-tuning includes: By introducing and training a small number of additional parameters, the domain expert model can master the specialized terminology and decision-making logic of the semiconductor manufacturing field.

[0040] Specifically, an efficient parameter fine-tuning method is introduced. During the fine-tuning process, the original massive parameters of the base language model are kept frozen, and only a very small number of newly added parameters are trained and updated. By minimizing parameter adjustments, while fully preserving the original general capabilities of the model, knowledge from professional fields such as semiconductor manufacturing is superimposed. This ensures that the model achieves expert-level skill performance in professional fields, while fully maintaining its original general language understanding and logical reasoning capabilities.

[0041] In some embodiments, constructing the inspection agent on the agent configuration platform includes: Receive configuration instructions for defining the role description of the inspection agent and the inspection task objectives; In response to the user's modular selection operation, at least one functional module is configured for the inspection agent, the functional module including one or more of the following: a tool invocation module and a knowledge retrieval base module; Receive user orchestration operations for workflow nodes and configure the inspection workflow engine; Based on the role description, the inspection task objectives, the functional modules, and the inspection workflow, the inspection intelligent agent is generated and deployed.

[0042] Specifically, the professional role and specific inspection task objectives of the intelligent agent are first clarified through configuration instructions. Then, the core functional modules such as tool invocation and knowledge retrieval are flexibly configured for the inspection intelligent agent through modular selection, and the execution logic and node sequence of the inspection workflow are defined. Finally, based on the above configuration, an inspection intelligent agent that can be put into operation immediately is automatically generated. It can seamlessly collaborate with semiconductor manufacturing CIM software such as MES, EAP, RTD and other systems. Moreover, its inspection logic can be quickly and flexibly adjusted through the configuration platform, which has high adaptability and scalability, breaking down system barriers.

[0043] In some embodiments, the user's modular selection and arrangement operations are completed by dragging and dropping modules and nodes in the graphical interface.

[0044] In conjunction with the above embodiments, users can independently build exclusive agents for different process areas through a graphical interface, such as dragging and dropping, and checking. They can flexibly configure the role, task objectives, execution logic and permissions without coding, which greatly reduces the technical threshold and time cost of developing semiconductor intelligent applications. This allows non-programmers (such as engineers and operations managers in semiconductor factories) to quickly build and manage these intelligent agents, significantly lowering the threshold for developing intelligent manufacturing applications and giving them the potential for large-scale and rapid deployment.

[0045] In some embodiments, such as Figure 2As shown, the method for intelligent agents to automatically inspect semiconductor factories further includes: a feedback optimization process. Record the dialogue and decision data generated by the inspection agent during its interaction with the user and the execution of tasks; Based on the dialogue and decision data, at least one of the following is optimized for the inspection agent: its knowledge retrieval base, anomaly attribution model, or role persona.

[0046] Specifically, through end-to-end dialogue tracking capabilities, the system can structurally record and analyze every user interaction process, including not only dialogue history, but more importantly, capturing the user's query intent, high-frequency questions, and unmet potential needs. This provides precise data support for system bottleneck identification, knowledge base blind spot supplementation, and interaction logic optimization, thereby enabling the system to evolve itself.

[0047] In some embodiments, the method for the intelligent agent to automatically inspect a semiconductor factory further includes: The interaction logs of multiple inspection agents are aggregated by a super agent coordinator. Perform cross-functional and / or cross-process correlation analysis on the interaction logs to obtain correlation analysis results used to describe complex problems; Based on the correlation analysis results, at least one of the following is optimized: the knowledge retrieval base, the anomaly attribution model, or the role persona of the inspection agent.

[0048] In implementation, a SuperAgent is used to collect dialogue logs, serving as the core coordinator of the system and aggregating the interaction logs of all sub-agents. By analyzing these cross-functional and cross-process dialogue logs, the system can discover the correlations between complex problems.

[0049] Based on feedback from dialogue tracking and task execution, the "persona" of the Agent can be dynamically optimized through data-driven methods. If an Agent is found to be too passive or aggressive in the interaction, the role description and response strategy in its prompt word template can be adjusted.

[0050] In some embodiments, after the intelligent inspection agent performs its tasks, it automatically generates a structured inspection report. This report integrates anomaly detection, preliminary attribution, and priority assessment into a comprehensive output.

[0051] In some embodiments, based on the deep mining and learning capabilities of historical inspection data, the system continuously corrects and optimizes its attribution model by performing pattern recognition and root cause analysis on accumulated abnormal data on a regular basis (e.g., weekly). This enables the system to process and correlate massive amounts of sensor data and historical records in real time, and to perform deep mining and root cause tracing through the model. The system transforms raw data into in-depth insights with decision support value, and allows expert experience to be solidified, inherited, and continuously optimized within the system.

[0052] This application possesses continuous learning capabilities, enabling it to automatically optimize its diagnostic model, persona, and knowledge base through dialogue feedback and inspection results. This allows for the self-evolution of diagnostic strategies and continuous improvement in accuracy, demonstrating its ability to grow alongside the business.

[0053] This application, through dialogue tracking, feedback loops, and periodic attribution analysis, can identify knowledge blind spots, optimize diagnostic models, and adjust agent behavior. This transforms the system from a static tool into a dynamic, self-optimizing system that grows alongside the business, possessing autonomous capabilities. It frees engineers from repetitive inspection work, allowing them to engage in higher-value, truly expert work (such as agent optimization, handling extreme cases, and process innovation), while delegating standardized, routine diagnostic decision-making power to the agent system they configure themselves.

[0054] This application utilizes a low-code / no-code platform to instantly copy solidified expert capabilities (i.e., pre-configured Agent templates) with a single click and fine-tune them for different process areas, thereby instantly deploying countless "virtual expert clones." This enables the unlimited replication and large-scale application of expert capabilities, breaking the limitations of manpower and time.

[0055] Based on the same inventive concept, this application also provides a system for the intelligent agent to automatically inspect a semiconductor factory, comprising: The inspection intelligent agent construction module is used to build and deploy inspection intelligent agents based on the intelligent agent configuration platform, and to configure the inspection intelligent agents with device interface plugins, knowledge retrieval base and inspection workflow engine; The workflow definition module is used to define the inspection task process through the inspection workflow engine. The task execution module is used to drive the inspection agent to perform inspection tasks in an online semiconductor factory based on a 2.5D three-dimensional model, wherein the online semiconductor factory maps the layout, equipment location and equipment data interface of the physical factory. The data acquisition module is used to call the device data interface through the device interface plugin to acquire real-time sensor data; The report generation module is used to control the inspection agent through the inspection workflow engine to perform abnormal pattern recognition and root cause analysis on the real-time sensor data, and generate an inspection report in conjunction with the knowledge retrieval base.

[0056] In some embodiments, the inspection agent construction module further includes: The large language model construction module is used to build an evaluation system for the semiconductor manufacturing field. It quantitatively scores multiple base large language models based on process control logic, equipment interaction instructions and measurement data analysis capabilities, and selects the target base large language model based on the scoring results. By utilizing a domain corpus covering the entire wafer manufacturing process, the target base large language model is pre-trained and fine-tuned to obtain a domain expert model for the semiconductor manufacturing field.

[0057] In some embodiments, the system for automatically inspecting a semiconductor factory by an intelligent agent further includes: The feedback module is used to record the dialogue and decision data generated by the inspection agent during interaction with the user and execution of tasks; and to optimize at least one of the knowledge retrieval base, anomaly attribution model, or role persona of the inspection agent based on the dialogue and decision data.

[0058] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0059] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for an intelligent agent to automatically patrol a semiconductor factory, the method comprising: The method comprises the following steps: based on the agent configuration platform, constructing and deploying the inspection agent, and configuring the device interface plug-in, knowledge retrieval library and inspection workflow engine for the inspection agent; defining the inspection task process through the inspection workflow engine; driving the inspection agent to perform the inspection task in the online semiconductor factory based on the 2.5D three-dimensional model, wherein the online semiconductor factory maps the layout, device location and device data interface of the physical factory; calling the device data interface through the device interface plug-in to obtain real-time sensor data; controlling the inspection agent to perform abnormal pattern recognition and root cause analysis on the real-time sensor data through the inspection workflow engine, and generating an inspection report in combination with the knowledge retrieval library.

2. The method of claim 1, wherein, The method of constructing and deploying the inspection agent comprises: configuring a large language model for the agent, wherein the construction method of the large language model comprises: constructing an evaluation system for the semiconductor manufacturing field, quantitatively scoring a plurality of base large language models based on process control logic, device interaction instructions and measurement data analysis capabilities, and selecting a target base large language model according to the scoring results; using a domain corpus covering the whole wafer manufacturing process to pre-train and fine-tune the target base large language model to obtain a domain expert model in the semiconductor manufacturing field.

3. The method of claim 2, wherein, The fine-tuning comprises: introducing and training a small number of additional parameters to enable the domain expert model to master the professional terms and decision logic in the semiconductor manufacturing field.

4. The method of claim 1, wherein the intelligent agent automatically patrols the semiconductor factory. The construction of the inspection agent on the agent configuration platform comprises: receiving configuration instructions for defining the role description and inspection task target of the inspection agent; in response to the user's modular selection operation, configuring at least one functional module for the inspection agent, the functional module comprising one or more of a tool calling module and a knowledge retrieval library module; receiving the user's arrangement operation on the workflow node to configure the inspection workflow engine; based on the role description, the inspection task target, the functional module and the inspection workflow, generating and deploying the inspection agent.

5. The method of claim 4, wherein the intelligent agent automatically patrols the semiconductor factory by, The user's modular selection operation and arrangement operation are completed through the modules and nodes in the drag-and-drop graphical interface.

6. The method of claim 1, wherein the intelligent agent automatically patrols the semiconductor factory. The method of automatically inspecting the semiconductor factory by the agent further comprises: recording the dialogue and decision data generated by the inspection agent during interaction with the user and task execution; optimizing at least one of the knowledge retrieval library, the abnormal cause model or the role person of the inspection agent according to the dialogue and decision data.

7. The method of claim 6, wherein the intelligent agent automatically patrols the semiconductor factory by, The method of automatically inspecting the semiconductor factory by the agent further comprises: summarizing the interaction logs of a plurality of the inspection agents through a super agent coordinator; performing cross-functional and / or cross-process correlation analysis on the interaction logs to obtain correlation analysis results for describing composite problems; optimizing at least one of the knowledge retrieval library, the abnormal cause model or the role person of the inspection agent according to the correlation analysis results.

8. A system for automatically inspecting a semiconductor factory by an intelligent agent, characterized by, The system for automatically inspecting the semiconductor factory by the agent comprises: The inspection intelligent agent construction module is configured to construct and deploy an inspection intelligent agent based on an intelligent agent configuration platform, and configure a device interface plug-in, a knowledge retrieval library, and an inspection workflow engine for the inspection intelligent agent; The workflow definition module is configured to define an inspection task flow through the inspection workflow engine; The task execution module is configured to drive the inspection intelligent agent to perform an inspection task in an online semiconductor factory based on a 2.5D three-dimensional model, wherein the online semiconductor factory maps the layout, device position, and device data interface of a physical factory; The data acquisition module is configured to call the device data interface through the device interface plug-in to acquire real-time sensor data; The report generation module is configured to control the inspection intelligent agent to perform abnormal pattern recognition and root cause analysis on the real-time sensor data through the inspection workflow engine, and generate an inspection report in combination with the knowledge retrieval library.

9. The system for automatically inspecting a semiconductor factory by intelligent agents according to claim 8, wherein, The inspection intelligent agent construction module further includes: A large language model construction module is configured to construct an evaluation system for the semiconductor manufacturing field, quantitatively score a plurality of base large language models based on process control logic, device interaction instructions, and metrology data analysis capabilities, and select a target base large language model according to the score results; A domain corpus covering the whole process of wafer manufacturing is used to pre-train and fine-tune the target base large language model to obtain a domain expert model in the field of semiconductor manufacturing.

10. The system for automatically inspecting a semiconductor factory by intelligent agents according to claim 8, wherein, The system for automatically inspecting a semiconductor factory by the intelligent agent further includes: A feedback module is configured to record dialogue and decision data generated by the inspection intelligent agent during interaction with a user and task execution, and optimize at least one of a knowledge retrieval library, an abnormality attribution model, or a role person setup of the inspection intelligent agent according to the dialogue and decision data.

Images