A semiconductor wafer manufacturing closed-loop control method, system, device and medium based on OODA cycle

Abstract

The present application relates to the technical field of semiconductor manufacturing, in particular to a semiconductor wafer manufacturing closed-loop control method, system, device and medium based on OODA cycle; by constructing a unified semantic system adapting to multiple manufacturer equipment and multiple processes, accurate mapping of original data to business semantics is realized, and data islands are broken; by establishing an event-driven real-time sensing mechanism, combined with edge computing technology, the abnormal response time is reduced from hours to minutes; by configuration deployment technology, the landing cycle of process optimization scheme is shortened, different manufacturer equipment and process nodes are adapted, and the overall optimization of wafer manufacturing process is realized, and the efficiency, yield and safety of wafer manufacturing are improved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and more specifically, to a closed-loop control method, system, device, and medium for semiconductor wafer manufacturing based on the OODA cycle. Background Technology

[0002] Semiconductor wafer manufacturing is a typical industrial scenario characterized by high complexity, high precision, and long process. Existing technologies are all focused on optimizing the wafer manufacturing process, but none of them have formed a systematic closed-loop control system. They cannot meet the core requirements of high precision and high safety in the entire wafer manufacturing process, and have obvious defects and shortcomings.

[0003] (a) Its core pain points include: 1. Data silos and semantic fragmentation hinder data connectivity across all processes: In existing technologies (especially single-process automation solutions and AI-assisted solutions), the data formats of various devices (such as ASML lithography machines and KLA inspection machines) are inconsistent, and there is a lack of unified semantic definitions. For example, the "alignment error" of lithography equipment and the "linewidth deviation" of etching equipment do not have a unified semantic relationship, resulting in data not being able to be linked across processes. The core reason is that existing technologies only focus on data acquisition from a single device or a single process, without building a unified semantic system covering "wafer batch-process-parameter-defect," which makes it impossible to achieve data connectivity in the "observation-orientation" stage of the OODA loop. This is also one of the core technical problems that this invention needs to solve.

[0004] 2. Delayed event response and low closed-loop efficiency: Traditional manual control solutions rely on manual inspections, resulting in anomaly response times of up to several hours; single-process automation solutions can only detect anomalies in a single process and cannot link with other processes to trace the cause; AI-assisted solutions can predict anomalies, but lack a real-time triggering mechanism. The core reason is that existing technologies have not established an "event-driven" real-time perception mechanism, and do not use defects, parameter deviations, and other anomalies as trigger points, leading to a passive "post-event detection - manual traceability" mode, which can easily cause the scrapping of entire batches of wafers. This corresponds to the "delayed observation" problem that this invention needs to solve.

[0005] 3. AI decision-making carries high risks and ambiguous responsibilities, failing to meet industry compliance requirements: Existing AI-assisted solutions directly output control commands through AI models, lacking a human (process engineer) decision-making process. This makes them susceptible to "model illusions" (such as misjudging the cause of defects) leading to process anomalies. Furthermore, the lack of a decision log traceability mechanism makes it impossible to clearly identify the responsible party after problems arise. The core reason is that existing technologies rely excessively on independent AI decision-making, failing to balance "AI efficiency" with "human controllability," and ignoring the semiconductor industry's stringent requirements for compliance and traceability in process decisions. This corresponds to the "AI proposal-human decision-making" mechanism that this invention aims to construct.

[0006] 4. Long deployment cycle, lack of security, and poor adaptability: In existing technologies, adjusting new process parameters requires custom development of equipment interfaces, taking 1-3 months from proof-of-concept (POC) to production deployment, making it impossible to quickly adapt to different process nodes and equipment from different manufacturers. Furthermore, wafer process data (such as lithography parameters for 3nm chips) is core sensitive information, and existing systems lack end-to-end security protection, making them prone to data leakage. The core reason is that existing technologies have not built a "configuration-based deployment" delivery model and have not embedded security mechanisms throughout the entire process, resulting in high deployment costs and significant security risks. This corresponds to the "configuration-as-deployment" delivery model and end-to-end security mechanism of this invention.

[0007] Existing technologies do not deeply integrate the OODA loop with semiconductor manufacturing scenarios, and cannot simultaneously meet the five major requirements of "real-time performance, semantic uniformity, decision controllability, rapid deployment, and security compliance". There is an urgent need for a systematic closed-loop control solution. Summary of the Invention

[0008] This invention addresses the problems of data silos and semantic fragmentation, delayed anomaly response, high AI decision-making risks, long implementation cycles, and security deficiencies in existing semiconductor wafer manufacturing technologies. It proposes a closed-loop control method, system, equipment, and medium for semiconductor wafer manufacturing based on an OODA loop. By constructing a unified semantic system adaptable to multiple vendors' equipment and multiple processes, it achieves accurate mapping from raw data to business semantics, breaking down data silos. By establishing an event-driven real-time perception mechanism combined with edge computing technology, it reduces anomaly response time from hours to minutes. Through configurable deployment technology, it shortens the implementation cycle of process optimization solutions, adapts to different vendors' equipment and process nodes, and achieves comprehensive optimization of the wafer manufacturing process, improving wafer manufacturing efficiency, yield, and security.

[0009] The specific implementation details of this invention are as follows: A closed-loop control method for semiconductor wafer manufacturing based on an OODA loop specifically includes the following steps: Step S1: Obtain process data from the edge gateway deployed on the device, call the ontology modeling algorithm to construct semantic mapping rules, and map the process data into semantic context data; Step S2: Standardize and preprocess semantic context data, and call the sliding window-based anomaly detection algorithm to obtain effective semantic context data; Step S3: Input the effective semantic context data into the AI ​​model to obtain the process adjustment proposal, and adjust the process parameters according to the process adjustment proposal; Step S4: Convert the adjusted process parameters into executable instructions, and call the encryption algorithm to encrypt and transmit the executable instructions to the equipment.

[0010] To better realize the present invention, step S2 further includes the following steps: Step S21: Based on the semantic context data, obtain the device sensor data and the defect detection device image data; Step S22: Call the sliding window-based anomaly detection algorithm to filter invalid data in the device sensor data during the device start-up and shutdown phases, and obtain valid data from the device sensors; Step S23: Detect image data from defect detection equipment according to the set threshold to obtain valid image data from defect detection equipment.

[0011] To better realize the present invention, the specific operation of step S1 is as follows: based on the constructed ontology semantic model and the set semantic mapping rules, the process data is mapped into semantic information to obtain semantic context data.

[0012] To better implement this invention, the semantic mapping rule is further defined as follows:

[0013] in, For valid data The corresponding semantic information, For wafer batch body objects, For process parameter body object, For defective ontology objects, , , The set association weight.

[0014] To better realize the present invention, step S3 further includes the following steps: Step S31: Integrate the CNN defect detection model and the RNN process prediction model to obtain the AI ​​detection model; Step S32: Based on semantic context data and AI detection models, obtain a process adjustment plan through multi-model fusion reasoning.

[0015] To better implement this invention, the specific operation of multi-model fusion inference in step S32 is further as follows:

[0016] in, This is a collection of final process adjustment proposals. Adjustment suggestions for CNN model output based on defective data. This is an optimization scheme for the RNN model based on historical process parameters. , For model weights, .

[0017] To better implement this invention, the specific operation of the encryption algorithm in step S4 is further as follows:

[0018] in, These are encrypted executable instructions. AES-256 encryption key, Executable instructions This is an XOR operation.

[0019] Based on the above-mentioned closed-loop control method for semiconductor wafer manufacturing based on the OODA cycle, in order to better realize the present invention, a closed-loop control system for semiconductor wafer manufacturing based on the OODA cycle is further proposed, including a semantic mapping unit, an anomaly detection unit, an adjustment unit, and an encryption unit. The semantic mapping unit is used to obtain process data from the edge gateway deployed on the device, call the ontology modeling algorithm to construct semantic mapping rules, and map the process data into semantic context data. The anomaly detection unit is used to standardize and preprocess semantic context data, and call a sliding window-based anomaly detection algorithm to obtain effective semantic context data. The adjustment unit is used to input semantic context data into the AI ​​model to obtain a process adjustment proposal, and adjust the process parameters according to the process adjustment proposal. The encryption unit is used to convert the adjusted process parameters into executable instructions and call the encryption algorithm to encrypt and transmit the executable instructions to the device.

[0020] Based on the above-mentioned closed-loop control method for semiconductor wafer manufacturing based on the OODA cycle, in order to better realize the present invention, an electronic device is further proposed, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, the above-mentioned closed-loop control method for semiconductor wafer manufacturing based on the OODA cycle is implemented.

[0021] Based on the above-mentioned closed-loop control method for semiconductor wafer manufacturing based on the OODA cycle, in order to better realize the present invention, a computer-readable storage medium is further proposed, wherein computer instructions are stored on the computer-readable storage medium; when the computer instructions are executed on the above-mentioned electronic device, the above-mentioned closed-loop control method for semiconductor wafer manufacturing based on the OODA cycle is realized.

[0022] The present invention has the following beneficial effects: (1) This invention reduces the response time of abnormal events from hours to minutes by using edge computing and anomaly detection algorithms, and reduces the wafer batch scrap rate by 30%-50%.

[0023] (2) This invention achieves 100% semantic consistency of data in each process through semantic modeling technology, improves process traceability efficiency by 60%, and completely breaks down data silos.

[0024] (3) In terms of decision reliability, the "AI proposal-human decision" collaborative decision-making algorithm of this invention improves the decision accuracy to over 95%, solves the "illusion" problem of independent decision-making by existing AI models, and the decision-making process is traceable and meets the compliance requirements of the semiconductor industry.

[0025] (4) In terms of implementation efficiency, the configurable deployment technology of this invention shortens the cycle from POC to production deployment of process optimization scheme from "1-3 months" to "≤1 hour", improves R&D efficiency by 80%, and can quickly adapt to different manufacturers' equipment and multiple process nodes.

[0026] (5) In terms of security performance, the present invention has a full-process security protection technology that meets ISO27001 and semiconductor industry data security standards, reducing the risk of sensitive process data leakage to 0 and ensuring the security of core technology secrets. Attached Figure Description

[0027] Figure 1 This is a schematic flowchart of the closed-loop control method for semiconductor wafer manufacturing based on the OODA loop provided by the present invention. Detailed Implementation

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and therefore should not be regarded as a limitation on the scope of protection. 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.

[0029] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set up," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0030] Example 1: This embodiment proposes a closed-loop control method for semiconductor wafer manufacturing based on the OODA loop, which specifically includes the following steps: Step S1: Obtain process data from the edge gateway deployed on the device, call the ontology modeling algorithm to construct semantic mapping rules, and map the process data into semantic context data; The specific operation of step S1 is as follows: based on the constructed ontology semantic model and the set semantic mapping rules, the process data is mapped into semantic information to obtain semantic context data.

[0031] The semantic mapping rules are specifically as follows:

[0032] in, For valid data The corresponding semantic information, For wafer batch body objects, For process parameter body object, For defective ontology objects, , , The set association weight.

[0033] Step S2: Standardize and preprocess semantic context data, and call the sliding window-based anomaly detection algorithm to obtain effective semantic context data; Step S2 specifically includes the following steps: Step S21: Based on the semantic context data, obtain the device sensor data and the defect detection device image data; Step S22: Call the sliding window-based anomaly detection algorithm to filter invalid data in the device sensor data during the device start-up and shutdown phases, and obtain valid data from the device sensors; Step S23: Detect image data from defect detection equipment according to the set threshold to obtain valid image data from defect detection equipment.

[0034] Step S3: Input semantic context data into the AI ​​model to obtain process adjustment proposals, and adjust process parameters according to the process adjustment proposals; Step S3 specifically includes the following steps: Step S31: Integrate the CNN defect detection model and the RNN process prediction model to obtain the AI ​​detection model; Step S32: Based on semantic context data and AI detection models, obtain a process adjustment plan through multi-model fusion reasoning.

[0035] The specific operation of multi-model fusion inference in step S32 is as follows:

[0036] in, This is a collection of final process adjustment proposals. Adjustment suggestions for CNN model output based on defective data. This is an optimization scheme for the RNN model based on historical process parameters. , For model weights, .

[0037] Step S4: Convert the adjusted process parameters into executable instructions, and call the encryption algorithm to encrypt and transmit the executable instructions to the equipment.

[0038] The specific operation of the encryption algorithm in step S4 is as follows:

[0039] in, These are encrypted executable instructions. AES-256 encryption key, Executable instructions This is an XOR operation.

[0040] Working principle: This embodiment is based on the OODA (Observe-Orient-Decide-Act) decision loop, integrating ontology modeling, edge computing, multi-model fusion reasoning, configurable deployment and AES-256 encryption, to achieve real-time anomaly detection, semantic context construction, AI collaborative decision-making and human adjudication, configurable execution of the solution and full-process security protection, forming a complete closed loop.

[0041] This embodiment utilizes semantic mapping formulas, sliding window anomaly detection algorithms, and multi-model fusion inference formulas to achieve seamless data flow throughout the entire wafer manufacturing process, rapid anomaly response, controllable and traceable decision-making, rapid solution implementation, and data security compliance. This significantly improves the efficiency, yield, and safety of wafer manufacturing and is applicable to the entire process of manufacturing 12-inch and smaller wafers, demonstrating significant industrial application value.

[0042] Example 2: This embodiment is based on the above embodiment 1, such as... Figure 1 As shown, a specific embodiment will be described in detail.

[0043] (1) Observe: Event-driven real-time perception Trigger event definitions include "parameter out-of-tolerance events" (such as photolithography alignment error > 0.2μm), "defect events" (such as the detection of linewidth unevenness defects > 1%), and "equipment abnormality events" (such as etching machine pressure fluctuation > 5%). Data acquisition method: Real-time acquisition of device sensor data (sampling frequency 100Hz) and defect detection device image data (300 frames per wafer) via edge gateway, and synchronization to the data access module; Key technologies: Employing edge computing to preprocess data and filter redundant noise, such as eliminating invalid data during equipment start-up and shutdown, to ensure the accuracy of event triggering.

[0044] (2) Orient: Semantic-driven context modeling Wafer Ontology Definition: Constructs core business objects and their relationships, as shown in the example below:

[0045] Semantic mapping logic: Automatically map the original equipment data (such as "alignment error 0.3μm") to semantic information such as "wafer batch A - lithography process - alignment error out of tolerance (threshold 0.2μm)", and associate it with the substrate temperature data of the epitaxial stage of the batch, such as "substrate temperature 900℃±5℃", to form a complete context.

[0046] (3) Decide: AI-enhanced + human-in-the-loop collaborative decision-making AI Proposal Generation: The AI ​​intelligence layer invokes the "Defect-Parameter Association Model" in the Catalog model to generate 2-3 solutions based on semantic context. Example: Option 1: Adjust the stage displacement parameter to +0.2μm, and simultaneously reduce the exposure dose by 5%; Option 2: Stop the machine and check the photoresist coating thickness (target value 300nm±10nm). If the deviation exceeds 5%, recoat. Human decision-making process: Process engineers can view the basis of AI proposals through a visual interface, such as "In 30 similar defects in history, Solution 1 has a success rate of 92%", and can choose: ① directly adopt the solution; ② modify parameters, such as adjusting the displacement parameter to +0.18μm; ③ reject and re-trigger AI analysis; Responsibility Anchoring: The system automatically records "AI proposal content - engineer's decision result - operation time" to form an unalterable decision log.

[0047] (4) Act: The execution of the configuration and deployment plan. Configurable execution: After the engineer makes a decision, the system automatically converts the solution into executable instructions for the equipment, such as issuing the instruction "displacement parameter +0.18μm" to the ASML lithography machine. There is no need to customize the development of the equipment interface, and the deployment cycle is ≤10 minutes. Effect verification: After execution, the equipment parameters and defect detection data are collected in real time and fed back to the Observe stage to trigger the next round of OODA loop. If the verification shows that the alignment error has dropped to 0.15μm and the defect rate is <0.3%, the closed loop ends; if the target is not met, the Orient stage analysis is re-entered.

[0048] 3. Core Support Mechanism Embedded security mechanisms: running throughout the entire OODA process, including: ① Data encryption (process parameters are transmitted using AES-256 encryption); ② Access control (engineers can only view / review data for their assigned process); ③ Operation traceability (each parameter adjustment corresponds to a unique operator ID); Model and deployment neutrality: The AI ​​intelligence layer supports model replacement (such as replacing the CNN defect detection model with the Transformer model), and the operation and maintenance base supports cross-vendor equipment deployment (such as simultaneously connecting ASML lithography equipment and Tokyo Electron etching equipment), avoiding customer lock-in.

[0049] Working Principle: This embodiment first addresses the technical challenges of data silos and semantic fragmentation by employing ontology modeling technology to construct a unified semantic model of "wafer batch-process-parameter-defect". Specifically, it first acquires various types of process data from mainstream manufacturers such as ASML, Applied Materials, and KLA through data acquisition interfaces. This includes wafer batch information (batch ID, size, material), process parameters (lithography alignment error, etching pressure, epitaxial temperature), defect data (defect type, linewidth deviation, location), and equipment information (equipment model, operating status). All types of data undergo standardized preprocessing to unify data units and formats. For example, the linewidth deviation unit of the etching equipment is converted from nm to μm to maintain consistency with the alignment error unit of the lithography equipment. Based on this, an ontology modeling algorithm is used to define the attributes and relationships of five core business objects, constructing semantic mapping rules. The core mapping formula is as follows:

[0050] in, For raw data The corresponding semantic information, For wafer batch body objects, For process parameter body object, For defective ontology objects, , , The association weight (set according to the importance of the process, with a value range of 0-1, and...) This formula allows for the automatic mapping of raw equipment data into semantic context, enabling the interconnection of data across multiple processes and providing technical support for the "directional" phase of the OODA loop.

[0051] Secondly, addressing the technical challenge of delayed anomaly response, this embodiment establishes an event-driven real-time perception mechanism, combining edge computing technology to achieve rapid data preprocessing and anomaly triggering. The specific technical process is as follows: Edge gateways are deployed at each device to collect real-time sensor data (sampling frequency set to 100Hz) and defect detection device image data (300 frames per wafer). An anomaly detection algorithm based on a sliding window is used at the edge to preprocess the collected data in real time, filtering out invalid redundant noise during device start-up and shutdown. The noise filtering formula is as follows:

[0052] in, For the preprocessed valid data, The value is 50, which is determined by optimization based on the sampling frequency. This is the original collected data. This parameter represents the normal threshold range. Data exceeding this range is considered abnormal and triggers an anomaly event. Anomalies are categorized into three types: parameter deviation events (e.g., lithography alignment error > 0.2 μm), defect events (e.g., linewidth unevenness defect rate > 1%), and equipment anomaly events (e.g., etching machine pressure fluctuation > 5%). After an event is triggered, the edge gateway synchronizes the pre-processed valid data and semantic information to the core processing module, enabling real-time anomaly detection and reducing response time from hours to minutes.

[0053] Furthermore, addressing the technical challenges of AI decision-making risks and accountability, this embodiment designs an "AI proposal-human decision-making" collaborative decision-making algorithm to balance AI decision-making efficiency with human controllability. The specific technical implementation is as follows: The AI ​​intelligence layer integrates multiple AI models (CNN defect detection model, RNN process prediction model), and based on semantic contextual data, generates process adjustment proposals through multi-model fusion reasoning. The multi-model fusion reasoning formula is as follows:

[0054] in, This is a collection of final process adjustment proposals. Adjustment suggestions for CNN model output based on defective data. This is an optimization scheme for the RNN model based on historical process parameters. , Model weights (dynamically adjusted based on model accuracy) The defect data here refers to the portion of y that is judged to be defective or abnormal. The AI ​​model will generate 2-3 sets of differentiated proposals and output supporting data such as the historical success rate and implementation time of each proposal for process engineers to make decisions. Engineers can directly adopt, modify parameters, or reject and recalculate. The system records "AI proposal content - engineer decision result - operation time - parameter adjustment details" through blockchain technology, forming an immutable decision log to achieve accountability and comply with semiconductor industry compliance requirements.

[0055] Finally, addressing the technical challenges of long deployment cycles and security vulnerabilities, this embodiment develops configurable deployment technology and end-to-end security protection technology. The configurable deployment technology, through the construction of a standardized equipment interface protocol library, is compatible with equipment from mainstream manufacturers such as ASML, Applied Materials, and KLA. Process plans approved by engineers can be automatically converted into executable instructions for the equipment via configuration files, eliminating the need for custom interface development and achieving a deployment cycle of ≤10 minutes. The end-to-end security protection technology permeates all stages of the OODA loop, employing AES-256 encryption algorithm for the transmission and storage of process data. The encryption formula is as follows:

[0056] in, For encrypted data, AES-256 encryption key, For the original process data, It performs an XOR operation; at the same time, it adopts the RBAC (Role-Based Access Control) mechanism to divide the permissions of different roles. Only process engineers have the authority to make decisions and adjust parameters, while ordinary operators can only view data. Combined with operation traceability technology, it ensures that every parameter adjustment is traceable, reducing the risk of data leakage and operation error.

[0057] The five core technological innovations mentioned above are interconnected, forming a complete OODA closed-loop control process: the Observe stage achieves real-time perception through edge computing and anomaly detection algorithms; the Orient stage achieves data association and context construction through semantic modeling; the Decide stage achieves AI proposal and human adjudication through collaborative decision-making algorithms; the Act stage achieves solution execution and effect verification through configurable deployment; and security protection technologies are integrated throughout the entire process to ensure the stability, security, and traceability of the entire system.

[0058] The other parts of this embodiment are the same as those in Embodiment 1 above, so they will not be described again.

[0059] Example 3: Based on any one of Embodiments 1-2 above, this embodiment takes "processing of uneven linewidth defects after 12-inch wafer photolithography" as an example, and combines the core technical solution of this embodiment to explain the technical implementation process in detail, verifying the feasibility and effectiveness of this embodiment. The specific implementation steps are as follows: 1. Scene definition and preliminary preparation This implementation scenario involves handling linewidth unevenness defects after the lithography process on a 12-inch wafer using a 7nm process. The wafer batch ID is W20250930-01, and the wafer size is 12 inches. The lithography process uses an ASML NXE:3400C lithography machine, and the defect detection equipment is a KLA3950. The system deployment adopts a "local server + edge gateway" architecture. The operation and maintenance base layer is deployed on the local server and connects to ASML and KLA devices via the OPCUA protocol. The data and semantic layer uses a PostgreSQL database to store the ontology model and Elasticsearch to store real-time data. The AI ​​intelligence layer is deployed at the edge and integrates a CNN defect detection model (TensorFlow framework) and an RNN process prediction model (PyTorch framework). The security mechanism uses AES-256 encryption, and the access control is based on RBAC, with only the "process engineer" role having the authority to make decisions.

[0060] 2. Implementation steps of OODA loop technology (1) Orient: Semantic Context Construction The data and semantic layer invoke a pre-built ontology semantic model based on a semantic mapping formula. (in The raw data transmitted from the edge is automatically mapped into semantic information: "Wafer batch W20250930-01-Photolithography process-Linewidth deviation 50nm (out of tolerance)-Alignment error 0.3μm (out of tolerance)-Exposure dose 120mJ / cm² (upper limit)".

[0061] Meanwhile, the epitaxial stage data of this wafer batch is associated with the semantic model, including substrate temperature of 905℃ (threshold 900±5℃, which meets the standard) and epitaxial layer thickness of 100nm (which meets the standard). The influence of the epitaxial process on defects is eliminated by the semantic association algorithm, and the photolithography process is identified as the core source of anomalies, forming a complete semantic context to provide data support for subsequent decision-making. The whole process takes ≤1min.

[0062] (2) Observe: Exception event triggering and real-time data preprocessing The KLA3950 defect inspection equipment acquires image data of wafer W20250930-01 in real time (300 frames per wafer), while the edge gateway synchronously acquires real-time sensor data from the ASML NXE:3400C lithography machine at a sampling frequency of 100Hz. The acquired data includes parameters such as alignment error, exposure dose, and lithography stage displacement.

[0063] At the edge, an anomaly detection algorithm based on a sliding window (sliding window N=50) is used to preprocess the collected data and filter out invalid data during the equipment start-up and shutdown phases. According to preset thresholds (linewidth deviation ≤30nm, alignment error ≤0.2μm, exposure dose 115±5mJ / cm²), the wafer linewidth deviation of 50nm, defect rate of 1.2%, alignment error of 0.3μm, and exposure dose of 120mJ / cm² are detected, all of which exceed the normal threshold range, triggering "defect event" and "parameter out-of-tolerance event".

[0064] The edge gateway synchronizes the pre-processed valid data and anomaly identifiers to the data and semantic layer, completing the technical implementation of the Observe phase. The entire process takes ≤1 minute.

[0065] (3) Decide: AI Collaborative Reasoning and Human Ruling The AI ​​intelligence layer invokes the CNN defect detection model and the RNN process prediction model, and uses semantic contextual data to perform multi-model fusion inference formulas. (in Two process adjustment proposals were generated: Proposal 1: Adjust the lithography stage displacement parameter to +0.2μm, reduce the exposure dose to 118mJ / cm², and predict that the linewidth deviation can be reduced to 25nm based on historical data, with a success rate of 93% and an implementation time of 5min; Proposal 2: Stop the machine to check the photoresist coating thickness (current detection value 290nm, threshold 300±10nm), and if the thickness deviation is >5%, recoat, with a prediction success rate of 98% and an implementation time of 20min.

[0066] Process engineers can view the reasoning process and historical supporting data of the AI ​​proposal through a visual interface. Based on the urgency of the current batch delivery, they select proposal 1 and adjust the displacement parameter to +0.18μm (to avoid over-adjustment). The system automatically records "AI proposal content - engineer's decision result - operation time (T=3min) - parameter adjustment details" through blockchain technology, forming an immutable decision log and completing the technical implementation of the Decide stage.

[0067] (4) Act: Configurable execution and effect verification After the engineer's decision, the system automatically converts the adjusted process plan into executable instructions for the ASML NXE:3400C lithography machine through the configurable deployment module. No custom development interface is required, and the equipment completes the parameter update (displacement parameter +0.18μm, exposure dose 118mJ / cm²) within 2 minutes.

[0068] After the parameters are updated, the edge gateway continues to collect the device operating parameters and the defect detection data of KLA3950 in real time. At T=10min, the detection results show: line width deviation 22nm (compliant), alignment error 0.16μm (compliant), defect rate 0.2% (compliant). The system feeds back the verification results to the Observe link and determines that the closed loop is complete.

[0069] Meanwhile, the safety log records "Operator ID: ENG001-AI Proposal 1-Decision to adjust displacement parameters +0.18μm-Execution result meets the standard", enabling operation traceability. The entire Act process takes ≤7 minutes.

[0070] 4. Implementation effect verification This implementation, through its core technical solution, reduces the processing time for uneven linewidth defects from several hours in existing technologies to 10 minutes, achieving a 100% defect handling success rate. All wafers in this batch met standards, with no scrap incidents. Simultaneously, the decision-making process is traceable, and data transmission and storage are encrypted using AES-256, preventing data leakage. The configurable deployment mode eliminates the need for customized interfaces, and the entire process from solution verification to implementation takes only 30 minutes, significantly improving implementation efficiency compared to the 1-3 months of existing technologies. Experimental verification demonstrates that the technical solution of this invention is feasible and effective, effectively addressing the core pain points of existing technologies and improving the efficiency, yield, and security of wafer manufacturing. It can be widely applied to the entire process manufacturing of 12-inch and smaller wafers.

[0071] 3. System Deployment Instructions The operation and maintenance base layer is deployed on a local server and connects to ASML and KLA devices via the OPCUA protocol; The data and semantic layer uses a PostgreSQL database to store the ontology model and Elasticsearch to store real-time data. The AI ​​intelligence layer is deployed at the edge, and the model catalog integrates a CNN defect detection model (TensorFlow framework) and an RNN process prediction model (PyTorch framework). The security mechanism uses AES-256 encryption, and access control is based on RBAC (Role-Based Access Control), with only the "Process Engineer" role having decision-making authority.

[0072] The other parts of this embodiment are the same as any one of the above embodiments 1-2, so they will not be described again.

[0073] Example 4: Based on any one of Embodiments 1-3 above, this embodiment proposes a closed-loop control system for semiconductor wafer manufacturing based on the OODA loop, including a semantic mapping unit, an anomaly detection unit, an adjustment unit, and an encryption unit; The semantic mapping unit is used to obtain process data from the edge gateway deployed on the device, call the ontology modeling algorithm to construct semantic mapping rules, and map the process data into semantic context data. The anomaly detection unit is used to standardize and preprocess semantic context data, and call a sliding window-based anomaly detection algorithm to obtain effective semantic context data. The adjustment unit is used to input semantic context data into the AI ​​model to obtain a process adjustment proposal, and adjust the process parameters according to the process adjustment proposal. The encryption unit is used to convert the adjusted process parameters into executable instructions and call the encryption algorithm to encrypt and transmit the executable instructions to the device.

[0074] This embodiment also proposes an electronic device, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements the above-described closed-loop control method for semiconductor wafer manufacturing based on the OODA loop.

[0075] This embodiment also proposes a computer-readable storage medium storing computer instructions; when the computer instructions are executed on the aforementioned electronic device, the aforementioned closed-loop control method for semiconductor wafer manufacturing based on the OODA loop is implemented.

[0076]

[0077] The other parts of this embodiment are the same as any one of the embodiments 1-3 above, so they will not be described again.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A closed-loop control method for semiconductor wafer manufacturing based on an OODA loop, characterized in that, Specifically, the following operations are included: Step S1: Obtain process data from the edge gateway deployed on the device, call the ontology modeling algorithm to construct semantic mapping rules, and map the process data into semantic context data; Step S2: Standardize and preprocess semantic context data, and call the sliding window-based anomaly detection algorithm to obtain effective semantic context data; Step S3: Input the effective semantic context data into the AI ​​model to obtain the process adjustment proposal, and adjust the process parameters according to the process adjustment proposal; Step S4: Convert the adjusted process parameters into executable instructions, and call the encryption algorithm to encrypt and transmit the executable instructions to the equipment.

2. The closed-loop control method for semiconductor wafer manufacturing based on the OODA loop according to claim 1, characterized in that, The specific operation of step S1 is as follows: based on the constructed ontology semantic model and the set semantic mapping rules, the process data is mapped into semantic information to obtain semantic context data.

3. The closed-loop control method for semiconductor wafer manufacturing based on the OODA loop according to claim 1, characterized in that, Step S2 specifically includes the following steps: Step S21: Based on the semantic context data, obtain the device sensor data and the defect detection device image data; Step S22: Call the sliding window-based anomaly detection algorithm to filter invalid data in the device sensor data during the device start-up and shutdown phases, and obtain valid data from the device sensors; Step S23: Detect image data from defect detection equipment according to the set threshold to obtain valid image data from defect detection equipment.

4. The closed-loop control method for semiconductor wafer manufacturing based on the OODA loop according to claim 2, characterized in that, The semantic mapping rules are specifically as follows: ； in, For valid data The corresponding semantic information, For wafer batch body objects, For process parameter body object, For defective ontology objects, , , The set association weight.

5. The closed-loop control method for semiconductor wafer manufacturing based on the OODA loop according to claim 1, characterized in that, Step S3 specifically includes the following steps: Step S31: Integrate the CNN defect detection model and the RNN process prediction model to obtain the AI ​​detection model; Step S32: Based on semantic context data and AI detection models, obtain a process adjustment plan through multi-model fusion reasoning.

6. The closed-loop control method for semiconductor wafer manufacturing based on an OODA loop according to claim 5, characterized in that, The specific operation of multi-model fusion inference in step S32 is as follows: ； in, This is a collection of final process adjustment proposals. Adjustment suggestions for CNN model output based on defective data. This is an optimization scheme for the RNN model based on historical process parameters. , For model weights, .

7. The closed-loop control method for semiconductor wafer manufacturing based on the OODA loop according to claim 1, characterized in that, The specific operation of the encryption algorithm in step S4 is as follows: ; in, These are encrypted executable instructions. AES-256 encryption key, Executable instructions This is an XOR operation.

8. A closed-loop control system for semiconductor wafer manufacturing based on the OODA loop, characterized in that, It includes a semantic mapping unit, an anomaly detection unit, an adjustment unit, and an encryption unit; The semantic mapping unit is used to obtain process data from the edge gateway deployed on the device, call the ontology modeling algorithm to construct semantic mapping rules, and map the process data into semantic context data. The anomaly detection unit is used to standardize and preprocess semantic context data, and call a sliding window-based anomaly detection algorithm to obtain effective semantic context data. The adjustment unit is used to input semantic context data into the AI ​​model to obtain a process adjustment proposal, and adjust the process parameters according to the process adjustment proposal. The encryption unit is used to convert the adjusted process parameters into executable instructions and call the encryption algorithm to encrypt and transmit the executable instructions to the device.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements a closed-loop control method for semiconductor wafer manufacturing based on an OODA loop as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 9, they implement a closed-loop control method for semiconductor wafer manufacturing based on an OODA loop as described in any one of claims 1-7.

Images