CIM Distributed Collaborative Management and Control System for Wafer Manufacturing

Abstract

This invention belongs to the field of semiconductor intelligent manufacturing technology, providing a CIM distributed collaborative management and control system for wafer manufacturing. It includes: an edge access layer, deploying multiple distributed edge nodes, dynamically adjusting the acquisition window and sampling step size based on the load status of the factory collaborative management and control layer; a factory collaborative management and control layer, employing a microservice architecture to deploy microservice groups for production execution, equipment management, and quality yield; the quality yield microservice group performs statistical process control analysis on real-time data and adds priority tags to abnormal events, with the data bus prioritizing routing of abnormal events to the intelligent decision-making unit based on these tags; and an intelligent decision-making unit, with a built-in AI-native intelligent decision-making module, performing multi-objective optimization simulation and deduction of wafer yield anomaly root cause localization and production scheduling schemes, feeding the deduction results back to the factory collaborative management and control layer to generate control commands for execution. This invention improves the real-time response capability, anomaly handling efficiency, and autonomous controllability of the wafer manufacturing system.

Description

Technical Field

[0001] This invention relates to the field of semiconductor intelligent manufacturing technology, specifically to a CIM distributed collaborative management and control system for wafer manufacturing. Background Technology

[0002] Wafer manufacturing is a core link in the semiconductor industry. Its process is complex, with a wide variety of equipment and a huge amount of data, which places extremely high demands on the real-time performance, reliability and intelligent decision-making capabilities of the manufacturing execution system. Traditional CIM systems mostly adopt a centralized architecture, which couples data acquisition, production execution, quality control and scheduling decisions to the central server. This has the following inherent defects: (1) The data volume of edge devices is large, and centralized processing leads to network bandwidth bottlenecks and computing pressure on central nodes; (2) The detection and response delay of abnormal events is high, which makes it difficult to meet the real-time control requirements of wafer manufacturing at the second or even millisecond level; (3) The system has poor scalability, and a lot of customized development is required for the addition of new equipment or process changes.

[0003] Patent CN119151381A discloses an AI edge computing platform, method, terminal, and system for Fab systems. This patent proposes an AI edge computing platform specifically designed for Fab systems, deployed at the edge of multiple equipment in a wafer manufacturing plant. It can collect and filter process parameter data that meets high-quality wafer standards, analyze and process it using a specific algorithm model, and dynamically adjust the process parameters of the equipment and compensate the Fab system index based on the analysis results. However, this patent has the following shortcomings: (a) Its data acquisition window and sampling step size are preset fixed values, which cannot be dynamically adjusted according to the real-time load status of the system. This can easily lead to data congestion and loss under high load and waste of resources under low load; (b) The platform only focuses on data acquisition and analysis at the edge and does not build a three-layer collaborative architecture from the edge to the factory control layer and then to the intelligent decision-making unit. Abnormal events and normal data are not differentiated in routing, and real-time performance is not guaranteed; (c) Its processing of equipment communication protocols only mentions "specific algorithm models" and does not solve the problem of adaptive parsing of equipment private protocols that exist in large numbers in wafer manufacturing. New equipment access still requires manual configuration.

[0004] Patent CN118444637B discloses a method and system for acquiring data from semiconductor devices that supports standardized communication modifications. This patent proposes a method and system for acquiring data from semiconductor devices that supports standardized communication modifications. It classifies devices with different communication capabilities (standard-compliant devices, programmable communication devices, and devices without communication capabilities) through a cluster device controller system, achieving protocol conversion from device data to standard semiconductor device communication interfaces. However, this patent has the following shortcomings: (a) Its core idea is to "modify" the device's communication to match the standard interface, essentially making the device conform to the standard protocol, rather than the system actively adapting to the device. For devices that do not fully conform to the standard, "development" is required. (a) The program implements communication, and for devices without communication capabilities, it is necessary to "collect the device program interface". In essence, it still relies on manual intervention and targeted driver development, and cannot achieve plug-and-play adaptive parsing; (b) The solution is mainly aimed at protocol conversion at the device communication level, and does not involve the data semantic unification in the wafer manufacturing field (such as standardized modeling of core objects such as work orders, batches, and wafers), making it difficult to support collaborative decision-making across microservices; (c) Its collection method is centralized cluster control, and does not adopt a distributed edge node architecture. In the scenario of large-scale device access, it is easy to encounter single-point performance bottlenecks and data transmission delays; (d) It does not provide a smooth switching mechanism from the existing CIM system to the new system. In actual production line deployment, it requires long-term downtime and has a high switching risk.

[0005] Existing CIM systems, while attempting to incorporate edge computing nodes to cope with massive amounts of data, typically use them only as static data filters or protocol conversion gateways, with fixed data acquisition and processing strategies. This lack of dynamic adjustment based on real-time production conditions, equipment health status, or business priorities leads to low utilization of edge node computing resources and difficulty in effectively supporting complex AI inference tasks. Furthermore, the absence of efficient collaboration mechanisms between edge nodes and the cloud or factory control layers results in a single path for data uploads and command issuance, failing to fully leverage the advantages of edge computing in real-time and localized processing, and hindering the organic integration of global optimization decisions and real-time edge control. In addition, existing systems lack targeted encryption transmission and access control strategies for sensitive process data and intellectual property information collected at the edge, posing a potential risk of data leakage. These issues collectively result in significant performance bottlenecks and functional shortcomings in existing CIM systems when facing the increasing demands for intelligence and flexibility in wafer manufacturing.

[0006] Therefore, how to build a distributed collaborative management and control system for wafer manufacturing that has the capabilities of dynamic adaptive acquisition, protocol self-adaptation, multi-feature fusion root cause localization, and smooth migration has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] This invention provides a distributed collaborative management and control system for CIM in wafer manufacturing, which addresses the technical problems of bandwidth bottlenecks, insufficient real-time response, reliance on manual development for protocol adaptation, and inability to achieve layered collaborative management and control in the centralized architecture of existing technologies.

[0008] This invention provides a CIM distributed collaborative management and control system for wafer manufacturing, comprising: The edge access layer, deployed at the wafer manufacturing equipment, includes multiple distributed edge nodes. Each edge node is used to receive and process real-time data from at least one semiconductor manufacturing equipment that is connected to the edge node, and dynamically adjusts the size of its data acquisition window and sampling step size, as well as the sliding step size of its feature extraction window, according to the load status from the factory collaborative management layer. The load status includes at least one of CPU utilization or task queue depth. The edge node also has a built-in protocol adaptive parsing module based on device fingerprints, which is used to: extract the device fingerprint of the access semiconductor manufacturing equipment; match the standard communication protocol template according to the device fingerprint; if the match fails, start the dynamic learning mode, listen to the communication messages between the device and the host computer, and reverse-engineer a private protocol parser for the semiconductor manufacturing equipment by analyzing the semantic features of the message frame structure, start character, end character and data length field; the private protocol parser is stored in association with the device fingerprint and uploaded to the unified protocol repository of the factory collaborative management layer; The factory collaborative management and control layer communicates with the edge access layer via a data bus. The factory collaborative management and control layer adopts a microservice architecture, deploying production execution microservice groups, equipment management microservice groups, and quality yield microservice groups. It is configured to receive real-time data from the edge access layer. The quality yield microservice group performs statistical process control analysis on the real-time data and adds event priority tags to the analyzed abnormal event data. Based on the event priority tags, the data bus prioritizes routing abnormal event data to the intelligent decision-making unit. The intelligent decision-making unit communicates with the factory collaborative management and control layer and has a built-in AI native intelligent decision-making module. The AI ​​native intelligent decision-making module is used to receive abnormal event data, locate the root cause of wafer yield abnormalities, and perform multi-objective optimization simulation and deduction of production scheduling schemes. The simulation and deduction results are fed back to the factory collaborative management and control layer, which generates control commands and sends them to the corresponding semiconductor manufacturing equipment through the edge access layer.

[0009] Furthermore, in dynamic learning mode, the protocol adaptive parsing module reverse-engineers a private protocol parser by: statistically analyzing the occurrence patterns of start and end characters in communication messages to determine message boundaries; analyzing the range of value changes of bytes at fixed offset positions to infer length fields, command codes, or data identifiers; and using entropy analysis to assist in identifying integer, floating-point, or string data types.

[0010] Furthermore, the data bus includes a semantic translation layer and a dynamic routing layer; The semantic transformation layer is pre-set with a core object data model for the wafer manufacturing field. It is used to semantically annotate and standardize heterogeneous data from the production execution microservice group, equipment control microservice group, quality yield microservice group and edge nodes according to the core object data model to obtain standardized data. Among them, the core objects include work orders, batches, wafers, equipment, process steps, measurement parameters and defects. The dynamic routing layer receives standardized data and distributes it to the production execution microservice group, equipment management microservice group, and quality yield microservice group that subscribe to the standardized data, according to the semantically labeled data type, data source, and preset priority rules, in descending order of priority. The dynamic routing layer also prioritizes routing abnormal event data to the intelligent decision-making unit based on the event priority tags.

[0011] Furthermore, the AI-native intelligent decision-making module is equipped with a wafer yield anomaly root cause localization model based on hybrid time-domain and frequency-domain feature fusion; the wafer yield anomaly root cause localization model includes: The time-domain feature extraction submodule is used to calculate the statistical features of real-time data over time, including time-domain statistical features. The frequency domain feature extraction submodule is used to transform the time series of real-time data into the frequency domain through fast Fourier transform and extract the spectral energy distribution features. The feature fusion and attention submodule is used to concatenate time-domain features and frequency-domain features to form a fused feature vector, and apply a multi-head self-attention mechanism to calculate the contribution weights of different measurement parameters and different feature domains to the current yield anomaly. The root cause inference submodule is built on a graph neural network, where the nodes of the graph represent equipment, process steps and measurement parameters, and the edges of the graph represent physical connections or process flow relationships. The root cause inference submodule uses fused feature vectors and contribution weights as the initial features of nodes and edge weights, and infers the set of root cause nodes that lead to abnormal yield and the confidence level of each node through graph convolution operations.

[0012] Furthermore, it also includes a switching migration module for achieving a smooth transition from the existing CIM system to the system. The switching migration module includes: The data mirroring and dual-track parallel submodule is used to establish a two-way data mirroring channel with the existing CIM system database while maintaining the operation of the existing CIM system, and to synchronize the full production data in real time; and to simultaneously start the edge access layer, factory collaborative management and control layer and intelligent decision-making unit of the system, so that they receive the same real-time production data stream as the existing CIM system and perform no-load verification. The gradual switching submodule is used to periodically compare the key control instructions output by the system and the existing CIM system for the same batch of wafers, and calculate the instruction matching degree. The instruction matching degree is the percentage of instructions with completely identical content out of the total number of compared instructions. When the instruction matching degree reaches more than 95% for three consecutive batches, the gradual switching is initiated, and a preset proportion of the control flow of semiconductor manufacturing equipment is gradually switched from the existing CIM system to the system. The full switchover and rollback submodule is used to switch all control flows of the semiconductor manufacturing equipment to the system within a switching time of no more than 30 seconds in response to a switchover command after a preset time of gradual switching to stable operation; at the same time, the existing CIM system is kept running as a backup, and when an anomaly is detected in the system, it supports rolling back to the existing CIM system within a rollback time of no more than 60 seconds.

[0013] Furthermore, the edge access layer also performs node self-healing and local caching, which includes: when the edge node detects a network interruption with the factory collaborative management layer, it switches to a degraded operation mode, caches the collected data in local non-volatile storage, and continues to perform real-time control of the semiconductor manufacturing equipment based on the rules of the local cache; and after the network is restored, it resumes the transmission of the cached data with added timestamps to the factory collaborative management layer. The factory collaborative management and control layer is configured with a three-active-node cluster architecture. The three-active-node cluster architecture includes: each microservice deploys at least three instances, and the three instances are distributed in three independent physical racks or three independent availability zones; the data bus uses the Raft consensus algorithm to realize metadata synchronization and master node election, and when any instance or network partition fails, the remaining instances complete the master node re-election within no more than 5 seconds. The intelligent decision-making unit is configured with a stateful hot standby architecture. The stateful hot standby architecture includes: the model parameters and real-time running status of the AI ​​native intelligent decision-making module are synchronized to the standby node every 1 second in the form of memory checkpoints, and the standby node loads the most recent checkpoint and takes over the service within 5 seconds when the master node fails; the normalized root mean square error of the model prediction results output by the standby node after taking over is less than 0.1% relative to the output of the master node.

[0014] Furthermore, the wafer yield anomaly root cause localization model based on hybrid time-domain and frequency-domain feature fusion is also configured to output a predictive health degradation trajectory of the semiconductor manufacturing equipment associated with the root cause nodes, while outputting the root cause node set and node confidence. The predictive health degradation trajectory is a health index sequence for the next 30 time units predicted by a Long Short-Term Memory (LSTM) network based on the historical fault characteristics and current fused feature vector of the semiconductor manufacturing equipment. Each time unit is 1 minute, and the health index is a dimensionless index normalized to the [0,1] interval. The AI-native intelligent decision-making module also includes a production scheduling simulation and deduction unit based on multi-objective optimization. The simulation and deduction unit works in collaboration with the root cause localization model. The simulation and deduction unit uses the predictive health decline trajectory as a dynamic constraint and substitutes it into the multi-objective optimization model. The objective function of the multi-objective optimization is to minimize the weighted sum, where the weight coefficients are adjustable. The terms of the weighted sum include: total completion time, semiconductor manufacturing equipment load balancing degree, and expected downtime penalty term calculated based on the predictive health decline trajectory. The simulation and deduction unit solves the multi-objective optimization model through reinforcement learning agents and outputs a production scheduling scheme. The production scheduling scheme is used to avoid assigning production tasks to bottleneck semiconductor manufacturing equipment, which refers to semiconductor manufacturing equipment whose health index will be lower than 0.3 within the next 30 minutes.

[0015] Furthermore, the system includes a switching and migration module and a three-active-node cluster architecture and a stateful hot standby architecture. During the gradual switching process, the three-active-node cluster in the three-active-node cluster architecture remains fully active, and the stateful hot standby architecture continuously performs memory checkpoint synchronization. When an abnormal instance is detected in any microservice group of the system during the gradual switching process, the gradual switching submodule triggers a rollback to the existing CIM system, and the three-active-node cluster automatically removes the abnormal instance and the remaining instance takes over.

[0016] Furthermore, the root cause node set and node confidence output by the root cause localization model are fed back to the simulation inference unit in real time, serving as the basis for dynamic adjustment of constraints in the multi-objective optimization model; after the optimized production scheduling scheme output by the simulation inference unit is executed, its execution result is fed back to the root cause localization model, serving as incremental samples for the next round of model training.

[0017] Furthermore, it also includes an adaptation layer for providing hardware and basic software adaptation for the edge access layer, factory collaborative management and control layer, and intelligent decision-making unit; the adaptation layer includes: The hardware adaptation submodule is used to adapt to CPU servers based on proprietary instruction sets or permanently licensed instruction sets, as well as NVMe solid-state drives based on control chips. The operating system adaptation submodule is used to adapt to the operating system, which integrates real-time optimization kernel patches to control interrupt response latency to within 50 microseconds. The database adaptation submodule is used to adapt to distributed time-series databases and implement a storage engine specifically for wafer manufacturing time-series data. The storage engine adopts a composite index structure based on time partitioning. The middleware adapter submodule is used to adapt to web application server middleware.

[0018] Compared with the prior art, the present invention has the following beneficial effects: By performing real-time data preprocessing on the device side through the edge access layer, only key information is uploaded to the factory collaborative management and control layer. This solves the uplink bandwidth bottleneck problem of traditional centralized CIM systems from the architectural level, significantly reduces the computing pressure on the core layer, and improves the overall real-time response capability of the system.

[0019] To address the protocol adaptation requirements of different types of semiconductor manufacturing equipment, this invention implements automatic reverse parsing of proprietary protocols through a protocol adaptive parsing module, eliminating the need for manual development of adaptation code and reducing the cost and time required for new equipment integration.

[0020] The layered architecture, combined with a layered collaborative management and control mechanism, allows edge nodes to perform local real-time control and data preprocessing. The factory collaborative management and control layer relies on a microservice architecture to enable elastic expansion and independent upgrades of each business module. The intelligent decision-making layer relies on AI models to complete complex root cause analysis and scheduling optimization. Each layer has clear responsibilities and higher collaborative efficiency.

[0021] By employing a priority-based dynamic routing mechanism, abnormal events and critical control data are prioritized for processing. Combined with feature fusion and graph neural networks, more accurate root cause localization of yield anomalies is achieved. Coupled with dynamic constraint-based multi-objective production scheduling optimization, the anomaly handling time can be effectively shortened, and the overall yield of wafer manufacturing can be improved.

[0022] Furthermore, the smooth switching migration mechanism provided by this invention can upgrade and replace the existing CIM system without interrupting production. At the same time, the multi-layer high availability design greatly improves the overall reliability of the system and can meet the stringent requirements of 7×24-hour continuous production in wafer manufacturing. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the CIM distributed collaborative management and control system for wafer manufacturing according to the present invention.

[0024] Figure 2 This is a schematic diagram of the module structure included in the wafer yield anomaly root cause localization model.

[0025] Figure 3 This is a schematic diagram of the switching migration module structure. Detailed Implementation

[0026] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. This section describes the working principle and technical effects of the present invention by way of embodiments, but the scope of protection of the present invention is not limited to these embodiments.

[0027] A distributed collaborative management and control system for CIM (Computer Integrated Manufacturing) in wafer fabrication includes: An edge access layer, deployed at the wafer manufacturing equipment, includes multiple distributed edge nodes. Each edge node receives and processes real-time data from at least one semiconductor manufacturing equipment connected to it. Based on the load status from the factory collaborative management layer, it dynamically adjusts the size of its data acquisition window, sampling step size, and sliding step size of its feature extraction window. The load status includes at least one of CPU utilization or task queue depth. This dynamic adjustment strategy is based on a preset load status threshold: when the CPU utilization of the factory collaborative management layer is below 50%, the edge node is in high-precision acquisition mode; when the CPU utilization is between 50% and 80%, the edge node switches to balanced mode; and when the CPU utilization is above 80%, the edge node switches to derating mode to reduce upstream pressure. The edge node also has a built-in protocol adaptive parsing module based on device fingerprints, which is used to: extract the device fingerprint of the access semiconductor manufacturing equipment; match the standard communication protocol template according to the device fingerprint; if the match fails, start the dynamic learning mode, listen to the communication messages between the device and the host computer, and reverse-engineer a private protocol parser for the semiconductor manufacturing equipment by analyzing the semantic features of the message frame structure, start character, end character and data length field; the private protocol parser is stored in association with the device fingerprint and uploaded to the unified protocol repository of the factory collaborative management layer; The factory collaborative management and control layer communicates with the edge access layer via a data bus. The factory collaborative management and control layer adopts a microservice architecture, deploying a production execution microservice group, an equipment management microservice group, and a quality yield microservice group. It is configured to receive real-time data from the edge access layer. The quality yield microservice group performs statistical process control analysis on the real-time data and adds event priority tags to the analyzed abnormal event data. The data bus, based on the event priority tags, prioritizes routing the abnormal event data to the intelligent decision-making unit. The intelligent decision-making unit is communicatively connected to the factory collaborative management and control layer and has a built-in AI native intelligent decision-making module. The AI ​​native intelligent decision-making module is used to receive the abnormal event data, perform multi-objective optimization simulation and deduction of wafer yield abnormality location and production scheduling scheme, and feed the simulation and deduction results back to the factory collaborative management and control layer. The factory collaborative management and control layer generates control commands and sends them to the corresponding semiconductor manufacturing equipment through the edge access layer.

[0028] The system described in this embodiment is deployed on a typical 12-inch wafer manufacturing line, which includes 87 semiconductor manufacturing devices for processes such as etching, deposition, lithography, and cleaning. Representative devices include the AMAT Centris® etching machine, the Lam Research 2300® deposition machine, and the ASML NXT:2000i lithography machine. The hardware environment is as follows: Edge access layer: Deployed at the edge nodes of each device, using Huawei Atlas 500 intelligent small stations (Kunpeng 920 processor, 8 cores, 2.6GHz clock speed, 16GB memory, and a built-in 256GB NVMe SSD for local caching). Factory collaborative management layer: Deployed in the data center, using 6 Huawei TaiShan 200 servers (Kunpeng 920 7260 processor, 64 cores, 2.6GHz clock speed, 256GB memory), of which 3 are used for a three-active node cluster (production execution, equipment management, and quality yield microservice group), and 3 are used for the data bus and database. The servers are interconnected via Huawei CloudEngine 8850 10 Gigabit Ethernet switches. The intelligent decision-making unit is deployed with two independent Kunpeng servers (master-slave mode) and configured with NVIDIA T4 GPU accelerator cards for AI model inference.

[0029] Software environment: The operating system is Kylin V10 (kernel version 5.4, with integrated PREEMPT_RT real-time patch), the database is DM8 (distributed time series engine), and the middleware is TongWeb 7.0. The AI ​​native intelligent decision-making module adopts a Transformer-based hybrid domain feature fusion model (8-layer encoder, 8-head attention mechanism, 512 hidden layer dimensions). The model is pre-trained using historical production line data covering the past 12 months (including 128-dimensional sensor features such as temperature, pressure, gas flow, RF power, vibration, etc., with a total data volume of approximately 15TB).

[0030] To verify the technical effectiveness of this invention, the system of this embodiment was compared with the existing centralized CIM system on the production line (Comparative Example 1, based on Oracle + x86 server, with a fixed data acquisition window of 1 second and no intelligent decision-making). The tests were conducted on the same production line under the same production task (same batch of wafers), and the following indicators were measured: Abnormal event response time: The time from when the quality yield microservice group detects an abnormal event (etching chamber pressure exceeds control limit) to when the intelligent decision-making unit outputs the root cause localization result.

[0031] Root cause localization accuracy: The degree of overlap between the set of root cause nodes output by the model and the manually labeled results for 100 manually labeled abnormal events (F1 score of precision / recall).

[0032] Production yield improvement: The yield change after using the system of this invention for 3 consecutive months, compared with the baseline yield of the existing CIM system before the switch.

[0033] System switching efficiency: The total time required for a smooth switch from the existing CIM system to this system and the wafer scrap rate during the switchover process.

[0034] The performance comparison is shown in the table below: Test results demonstrate that this embodiment, through distributed collaborative management and control and AI-native decision-making, significantly outperforms traditional centralized CIM systems in key indicators such as anomaly response speed (2.8 seconds vs. 47 seconds), root cause localization accuracy (92.4% automatic localization vs. complete manual reliance), production yield (improved by 1.8 percentage points), and system upgrade and switchover efficiency (30 seconds of uninterrupted production vs. 8 hours of downtime). This achieves the technical effects of "cost reduction and efficiency improvement" and "smooth replacement without production disruption." All the above test data can be reproduced in the same production line environment.

[0035] The working principle of the above technical solution is as follows: After the system starts, each distributed edge node in the edge access layer connects to the corresponding semiconductor manufacturing equipment. First, the protocol adaptive parsing module extracts the device fingerprint and matches it with the standard or private protocol template stored in the unified protocol repository. If the match is successful, the corresponding parser is directly loaded to complete the device data parsing. If the match fails, the dynamic learning mode is started, and the reverse parsing of the private protocol is completed through traffic monitoring. The newly generated protocol parser is then associated with the device fingerprint and synchronized to the unified protocol repository of the factory collaborative management layer to complete the self-expansion of the protocol ecosystem. During data processing, each edge node continuously monitors its own CPU utilization and task queue depth. Combined with the global load adjustment instructions issued by the factory collaborative management layer, the data acquisition window size and the sampling and feature extraction step size are dynamically adjusted. Under high load, the sampling frequency of non-critical data is reduced to reduce node resource consumption. Under low load, the sampling accuracy is improved to ensure data integrity and achieve dynamic and efficient utilization of edge resources. After preprocessing and feature extraction, the equipment data is transmitted to the factory collaborative management layer via the data bus. Various microservice groups perform production execution, equipment status monitoring, and quality yield analysis tasks respectively. When the quality yield microservice group identifies abnormal events through statistical process control analysis, it assigns corresponding priority tags to the abnormal data. The data bus performs routing scheduling based on priority, with high-priority abnormal events being transmitted to the intelligent decision-making unit first. The intelligent decision-making unit calls the AI-native intelligent decision-making module to locate the root cause of yield anomalies based on the accessed multi-dimensional production data. Then, based on the location results, it conducts multi-objective optimization simulation of the production scheduling plan. Under multiple constraints such as capacity, yield, and equipment loss, it generates the optimal adjustment plan, which is fed back to the factory collaborative management layer. The corresponding microservice generates specific control instructions, which are then sent to the corresponding edge nodes via the data bus. Finally, the edge nodes transmit the instructions to the semiconductor manufacturing equipment to complete the action adjustment, realizing distributed collaborative closed-loop management of the entire wafer manufacturing process.

[0036] The technical benefits of the above solution are as follows: By using distributed deployment at the edge and self-learning protocol parsing, it solves the problem of unified access for heterogeneous protocols of semiconductor manufacturing equipment from different manufacturers and models. It eliminates the need for repeated development of adaptation interfaces for new equipment, enabling rapid device access and reducing system expansion costs. Simultaneously, it achieves global sharing of protocol parsing capabilities. By dynamically adjusting sampling parameters based on load status at edge nodes, it achieves elastic allocation of edge computing resources. While ensuring data acquisition accuracy meets control requirements, it avoids excessive resource consumption at edge nodes, improving the overall stability of the system. The factory collaborative control layer adopts a microservice architecture to decompose different business modules. Combined with the priority routing mechanism of the data bus, it ensures both independent iterative expansion of business modules and timely processing of high-priority abnormal data, meeting the real-time control requirements of wafer manufacturing scenarios. Through an AI-native intelligent decision-making module, it completes root cause localization and scheduling scheme optimization, replacing traditional decision-making methods that rely on human experience. This improves anomaly response efficiency and decision accuracy, ultimately achieving a stable increase in wafer manufacturing yield and capacity.

[0037] In one embodiment, the protocol adaptive parsing module in dynamic learning mode reverse-engineers a private protocol parser by: statistically analyzing the occurrence patterns of start and end characters in communication messages to determine message boundaries; analyzing the range of value changes of bytes at fixed offset positions to infer length fields, command codes, or data identifiers; and using entropy analysis to assist in identifying integer, floating-point, or string data types.

[0038] The working principle of the above technical solution is as follows: The protocol adaptive parsing module does not need to know the format rules of the private protocol in advance. It can autonomously deduce the protocol structure by relying on statistical analysis of actual communication messages. First, the boundaries of a single message are defined by the frequency and positional patterns of the start and end symbols, splitting the continuously transmitted messages into independent message units. Then, for each split message, the changing characteristics of byte values ​​at fixed offset positions are analyzed to locate the corresponding fields that identify message length, operation commands, and data types. Finally, the randomness of different data segments is analyzed by entropy calculation to help distinguish different data types, ultimately completing the reverse construction of a complete private protocol parser. The entire process can quickly complete the adaptation and parsing of unknown private protocols without the need for manual writing of protocol rules.

[0039] The technical effects of the above solution are as follows: it solves the problems of traditional private protocol parsing requiring manual format sorting and rule writing, resulting in low adaptation efficiency and high dependence on engineers' protocol experience. It can complete the reverse derivation of unknown private protocols in a short time, significantly shortening the docking cycle of different manufacturers' devices accessing the CIM system, reducing the technical threshold for multi-device heterogeneous network access, and ensuring the accuracy of protocol parsing, providing basic support for reliable data transmission between distributed nodes.

[0040] In one embodiment, the data bus includes a semantic translation layer and a dynamic routing layer; The semantic transformation layer is pre-set with a core object data model for the wafer manufacturing field. It is used to semantically annotate and standardize heterogeneous data from the production execution microservice group, equipment control microservice group, quality yield microservice group and edge nodes according to the core object data model to obtain standardized data. Among them, the core objects include work orders, batches, wafers, equipment, process steps, measurement parameters and defects. The dynamic routing layer receives standardized data and distributes it to the production execution microservice group, equipment management microservice group, and quality yield microservice group that subscribe to the standardized data, according to the semantically labeled data type, data source, and preset priority rules, in descending order of priority. The dynamic routing layer also prioritizes routing abnormal event data to the intelligent decision-making unit based on the event priority tags.

[0041] The working principle of the above technical solution is as follows: The data bus acts as the data hub within the system, connecting edge nodes, various microservice groups, and the intelligent decision-making unit. The semantic transformation layer incorporates a core object model based on JSON Schema or Protobuf. For example, a "Measurement Parameter" object contains: object_type="MeasurementParameter", attributes={parameter_id, equipment_id, value, unit, timestamp, lot_id}; when the raw data reported by the edge node is "Equipment A, Pressure, 12.3, kPa, 14:05:06", the semantic transformation layer converts it into a standardized message conforming to the core object model. Data exchanged between microservice groups (such as a "work order start" command issued by the production execution microservice) is also converted into standard objects. The dynamic routing layer maintains a content-based routing table, and each microservice subscribes to the object types of interest to the routing layer upon startup (e.g., the quality yield microservice subscribes to "Measurement Parameter" and "Defect" objects). When standardized data arrives, the dynamic routing layer extracts its object type, data source, and whether it carries an anomaly priority label. It then distributes the data according to the following priority order: highest priority is given to anomaly event data (directly sent to the intelligent decision-making unit); second is real-time control commands; and third is ordinary production data. The routing layer internally employs lock-free hash tables and zero-copy technology to ensure high throughput and low latency.

[0042] The technical effects of the above solution are as follows: semantic conversion unifies the format of heterogeneous data from different sources, eliminates the data semantic differences between microservice groups and edge nodes, avoids the problem of repeatedly developing data conversion interfaces when connecting point to point, and reduces the complexity of system integration; the priority-based dynamic routing mechanism can ensure that high-priority abnormal events and control commands are transmitted and processed first, meeting the real-time control requirements of wafer manufacturing.

[0043] In one embodiment, such as Figure 2 As shown, the AI-native intelligent decision-making module is deployed with a wafer yield anomaly root cause localization model based on hybrid time-domain and frequency-domain feature fusion; the wafer yield anomaly root cause localization model includes: The time-domain feature extraction submodule is used to calculate the statistical characteristics of real-time data over time series. These statistical characteristics include time-domain statistical features (mean, variance, kurtosis, skewness, autocorrelation coefficient, etc.). The frequency domain feature extraction submodule is used to convert the time series of real-time data to the frequency domain through fast Fourier transform and extract spectral energy distribution features (such as main frequency components, spectral centroid, spectral entropy, etc.). The feature fusion and attention submodule is used to concatenate time-domain features and frequency-domain features to form a fused feature vector, and apply a multi-head self-attention mechanism to calculate the contribution weights of different measurement parameters and different feature domains to the current yield anomaly. The root cause inference submodule is built on a graph neural network, where the nodes of the graph represent equipment, process steps and measurement parameters, and the edges of the graph represent physical connections or process flow relationships. The root cause inference submodule uses fused feature vectors and contribution weights as the initial features of nodes and edge weights, and infers the set of root cause nodes that lead to abnormal yield and the confidence level of each node through graph convolution operations.

[0044] The working principle of the above technical solution is as follows: When the quality yield microservice group issues an abnormal event (e.g., a batch of wafers exceeds the standard for key electrical parameters), the intelligent decision-making unit triggers the root cause localization model. The time-domain feature extraction submodule calculates the statistical characteristics of each measurement parameter (e.g., deposition temperature, etching power, etc.) from historical time-series data uploaded from edge nodes (e.g., data sampled every 5 seconds in the past 24 hours). The frequency-domain feature extraction submodule performs a fast Fourier transform on the same time series to extract frequency-domain features, especially sensitive to periodic fluctuations (e.g., periodic fluctuations in pump speed). The feature fusion and attention submodule converts the time-domain feature vector (dimension N) into a single vector. t ) and frequency domain feature vectors (dimension N) f ) spliced ​​into N t +N f The fused vector is then input into a multi-head attention network. The attention mechanism learns whether the low-frequency drift of the temperature parameter (time-domain feature) or the high-frequency harmonics of the vibration parameter (frequency-domain feature) is more important in the current anomalous event, and assigns a weight (0-1) to each parameter and each feature domain. The root cause inference submodule pre-constructs a graph: nodes are divided into equipment nodes (e.g., PVD chamber, CMP polishing machine), process step nodes (e.g., deposition, lithography, etching), and measurement parameter nodes (e.g., film thickness, linewidth). Edges are established based on the actual process flow and equipment connection relationships. The graph neural network uses the fused feature vector as the initial feature of the nodes and the attention weights as the importance coefficients of the edges, performing multiple rounds of graph convolution. During convolution, the anomalous information propagates along the graph edges, and finally each node outputs a confidence score. The model selects the top K nodes with the highest confidence scores as the root cause output.

[0045] The technical effects of the above solution are as follows: multi-domain feature fusion can capture abnormal fluctuation information hidden in different dimensions, avoiding the failure to detect the root cause of anomalies by a single feature dimension; combined with a multi-head self-attention mechanism, it can automatically focus on parameters and feature domains that are more relevant to anomalies, reducing the interference of invalid features on root cause inference; relying on graph neural networks to model the topological relationship of equipment, processes and parameters in wafer manufacturing, it can make full use of the structured information of the production process, greatly improve the accuracy of root cause localization of yield anomalies, shorten the time for anomaly investigation, reduce yield loss, and adapt to the high requirements of wafer manufacturing for anomaly response efficiency.

[0046] In one embodiment, such as Figure 3 As shown, it also includes a switching migration module for achieving a smooth transition from the existing CIM system to the system described above. The switching migration module includes: The data mirroring and dual-track parallel submodule is used to establish a two-way data mirroring channel with the existing CIM system database while maintaining the operation of the existing CIM system, and to synchronize the full production data in real time; and to simultaneously start the edge access layer, factory collaborative management and control layer and intelligent decision-making unit of the system, so that they receive the same real-time production data stream as the existing CIM system and perform no-load verification. The gradual switching submodule is used to periodically compare the key control instructions output by the system and the existing CIM system for the same batch of wafers, and calculate the instruction matching degree. The instruction matching degree is the percentage of instructions with completely identical content out of the total number of compared instructions. When the instruction matching degree reaches 95% or more for three consecutive batches, the gradual switching is initiated, and a preset proportion of the control flow of the semiconductor manufacturing equipment is gradually switched from the existing CIM system to the system. The preset proportion can be 5%, 10%, 20%, etc., and is gradually increased according to the scale of the production line. The full switchover and rollback submodule is used to switch all control flows of the semiconductor manufacturing equipment to the system within a switching time of no more than 30 seconds in response to a switchover command after a preset time of gradual switching to stable operation; at the same time, the existing CIM system is kept running as a backup, and when an anomaly is detected in the system, it supports rolling back to the existing CIM system within a rollback time of no more than 60 seconds.

[0047] The working principle of the above technical solution is as follows: During production line deployment, the existing CIM system is running. The switching migration module first establishes a two-way data mirroring channel: through database triggers or CDC technology, the production data of the existing CIM system is synchronized to the database of this invention in real time. At the same time, the system of this invention starts in "idle" mode, that is, it only receives data, calculates decisions, and outputs instructions, but does not actually send them to the equipment. The sub-modules are gradually switched and periodically (e.g., after each batch of wafers is completed) compare the key control instructions output by the two systems for the same batch.

[0048] Switching tests on 12 different wafer products yielded experimental data regarding the 95% threshold: when the instruction matching rate was below 90%, the average wafer scrap rate after switching was 3.2%; when the matching rate was between 90% and 95%, the scrap rate was 0.8%; and when the matching rate reached above 95%, the scrap rate dropped to below 0.1%. Therefore, 95% was selected as the safe switching threshold. The requirement of three consecutive batches further reduced the risk of occasional deviations.

[0049] The aforementioned 95% threshold, 30-second switching duration, and 60-second rollback duration settings are all based on systematic simulated switching tests conducted by the inventors on a 12-inch wafer mass production line. The test environment was an actual production line at a semiconductor manufacturing plant, including 87 semiconductor manufacturing devices across processes such as etching, deposition, photolithography, and cleaning, involving 12 different wafer products (including logic chips, memory chips, and power devices). The test method was as follows: During the normal operation of the existing CIM system, the system of this invention was simultaneously started for no-load verification. After 72 hours of continuous operation, the sub-modules were gradually switched, and the key control instructions output by the two systems were compared batch by batch. A total of 327 batches of comparative tests were performed, with a cumulative comparison of over 15,000 instructions. The 95% threshold was determined as follows: when the instruction matching rate was below 90%, the average wafer scrap rate after simulated switching was 3.2% (based on statistics of 120 batches, 10 batches of each of 12 products); when the matching rate was between 90% and 95%, the scrap rate dropped to 0.8%; when the matching rate reached above 95%, the scrap rate further dropped to below 0.1% (the actual sample size was 89 batches). The 30-second switching duration was determined as follows: under normal production conditions, by gradually increasing the switching ratio (5%→20%→50%→100%), the time from issuing the full switching instruction to the completion of the transfer of control of all equipment was recorded. The test was repeated 20 times, with a maximum value of 28.7 seconds and a minimum value of 24.2 seconds. 30 seconds was set as the safety upper limit. The 60-second rollback duration was determined based on simulating 10 abnormal scenarios, including master node failure, network partitioning, and microservice instance crashes. The rollback operation included restoring the state from the backup database, reloading the control configuration of the existing CIM system, and verifying the communication link. The measured maximum rollback time was 54 seconds, hence 60 seconds was set as the safety limit. All the above test data can be reproduced in the applicant's laboratory or cooperative wafer fabs.

[0050] The 30-second switchover duration was determined based on the following considerations: In wafer manufacturing, equipment idle time exceeding one minute may lead to chamber temperature drops, gas contamination, etc., requiring re-stabilization. Testing showed that a full switchover completed within 30 seconds allows all equipment to maintain its process state without additional stabilization time; if the switchover duration exceeds 60 seconds, approximately 15% of the equipment requires recalibration, thus setting 30 seconds as the upper limit.

[0051] The 60-second rollback duration is set to quickly restore the original system. Completing the rollback within 60 seconds ensures that the overall production line downtime is controlled within 2 minutes (including the switching decision time), meeting the industry standard for wafer fabs' tolerance for business interruption (usually requiring <5 minutes).

[0052] Once the instruction matching rate for three consecutive batches is ≥95%, the operator can initiate a gradual switchover: for example, starting with 5% of the devices, switching their control traffic to the system of this invention, while the remaining 95% are still controlled by the existing system. Verification is performed again after each switchover cycle; if no anomalies are found, the percentage is gradually increased. The full switchover and rollback submodule switches all devices at once in the final stage (completion time ≤30 seconds), using atomic operations to ensure state consistency. Simultaneously, the existing system remains hot standby. The system of this invention embeds a health monitoring probe; if any microservice instance is detected to have three consecutive heartbeat timeouts or abnormal decision results, a rollback is immediately triggered, switching control back to the existing system within 60 seconds.

[0053] The technical benefits of the above solution are as follows: By employing a layered switching mechanism that combines data mirroring in parallel, step-by-step verification and switching, and full-scale rapid switching with heating backup and rollback, the solution addresses the pain point of high production reduction risks associated with replacing old systems with new ones in wafer manufacturing scenarios. System upgrades can be completed without production stoppages. Furthermore, by relying on preset matching thresholds and switching time control, the risks of wafer scrapping and production line interruption during the switching process can be kept within industry-acceptable limits. This ensures continuous production line operation while reducing the implementation costs and safety risks of upgrading and transforming outdated CIM systems.

[0054] In one embodiment, the edge access layer also performs node self-healing and local caching, including: when the edge node detects a network interruption with the factory collaborative management layer, it switches to a degraded operation mode, caches the collected data in local non-volatile storage (such as eMMC or SSD), and continues to perform real-time control of the semiconductor manufacturing equipment based on the rules of the local cache (such as the most recently received process parameter limits); and after the network is restored, it resumes the transmission of the cached, timestamped data to the factory collaborative management layer. The factory collaborative management and control layer is configured with a three-active-node cluster architecture; the three-active-node cluster architecture includes: each microservice deploys at least three instances, and the three instances are distributed in three independent physical racks or three independent availability zones; the data bus uses the Raft consensus algorithm to realize metadata synchronization and master node election, and when any instance or network partition fails, the remaining instances complete the master node re-election within no more than 5 seconds; The intelligent decision-making unit is configured with a stateful hot standby architecture; the stateful hot standby architecture includes: the model parameters and real-time running status of the AI ​​native intelligent decision-making module are synchronized to the standby node every 1 second in the form of memory checkpoints, and when the master node fails, the standby node loads the most recent checkpoint and takes over the service within 5 seconds; the normalized root mean square error of the model prediction result output by the standby node after taking over is less than 0.1% relative to the output of the master node.

[0055] The working principle of the above technical solution is as follows: Each edge node maintains a heartbeat connection with the factory collaborative management and control layer. If three consecutive heartbeats time out, the node determines that the network is interrupted and automatically switches to a degraded mode: local cached data is stored in a circular queue (maximum capacity 32GB), and a preset conservative strategy is used for device control (such as maintaining the current process parameters or adjusting according to safe values). After the network is restored, the edge node uploads the cached data to the data bus through a breakpoint resume protocol (based on HTTP Range or custom fragmentation), and the management and control layer performs backfilling.

[0056] Taking production-level microservices as an example, one instance is deployed in each of the three racks. The data bus maintains the metadata cluster based on Raft.

[0057] The performance metric of "completing master node reselection within 5 seconds" in this application was obtained based on actual testing in the following typical deployment environment: the server configuration is Kunpeng 920 processor (64 cores, 2.6GHz), 256GB memory, operating system is Kylin V10 (kernel version 4.19, with integrated PREEMPT_RT real-time patch), network environment is 10 Gigabit Ethernet (actual throughput not less than 8Gbps), and the single-way network latency between nodes is less than 1ms (99th percentile value of ping test). The three active nodes are deployed in three independent racks in the same data center, and each rack is equipped with an independent access switch and redundant power supply. Under the above configuration, the following failure scenarios were simulated for testing: (1) master node process crash (kill-9); (2) power failure of the physical machine where the master node is located; (3) network partitioning between the master node and the cluster (heartbeat port blocked by iptables). Each scenario was tested 50 times. Test results show that in the process crash scenario, the remaining nodes trigger election through Raft heartbeat timeout detection (timeout threshold set to 300ms). The average time from fault detection to new master node takeover is 2.1 seconds, the 99th percentile is 3.8 seconds, and the maximum is 4.6 seconds. In the physical machine power failure scenario, the average takeover time is 2.8 seconds, the 99th percentile is 4.2 seconds, and the maximum is 4.9 seconds. In the network partition scenario, the average takeover time is 3.2 seconds, the 99th percentile is 4.5 seconds, and the maximum is 4.8 seconds. The maximum takeover time in all test scenarios did not exceed 5 seconds. Therefore, "not exceeding 5 seconds" is a technical indicator that can be stably achieved under the typical configuration described in this application, rather than a theoretical limit. For other hardware configurations (such as using an Intel Xeon processor and CentOS operating system), actual tests show that the takeover time is also within the range of 3-6 seconds. The 5-second requirement can be met by appropriately adjusting the Raft heartbeat timeout parameter (such as adjusting it from 300ms to 200ms). Those skilled in the art, based on the teachings of this embodiment, can implement this reselection performance in similar distributed systems without inventive effort.

[0058] During operation, the master node of the AI ​​intelligent decision-making module serializes the model parameters (weights, biases) and runtime state in memory every second and synchronously writes them to the shared memory of the standby node. In the event of a master node failure, the standby node loads the latest checkpoint and warms up the model using input data from the last few seconds. Testing of the wafer yield prediction model (LSTM, 128 input dimensions, single-value output) showed that, under the same input, the NRMSE difference between the master and standby nodes due to differences in floating-point operation order is typically 0.05%-0.12%. Setting a tolerance threshold of 0.1% ensures decision consistency (when the difference exceeds 0.1%, the frequency of scheduling decision changes is <0.5%) without excessively restrictive measures leading to frequent master node switching. In actual testing, the average NRMSE of the model output after the standby node takes over and runs continuously with the master node is 0.07%, meeting the requirements.

[0059] The above performance indicators (synchronization every 1 second, takeover within 5 seconds, and normalized root mean square error less than 0.1%) were measured under the following test environment: both the primary and backup nodes use Kunpeng 920 processors, Kylin V10 operating system, and 10 Gigabit Ethernet interconnection, and the AI ​​model is an 8-layer LSTM (model file about 12MB).

[0060] Verification of a 1-second synchronization interval: During 72 hours of continuous operation, a total of 259,200 synchronization operations were performed, with an average transmission time of 87ms, which is much less than 1 second.

[0061] Verification of takeover time of 5 seconds: Simulating three types of failures, namely master node process crash, physical machine power failure, and network partition, the maximum takeover time was 3.9 seconds, 4.8 seconds, and 4.2 seconds, respectively, none of which exceeded 5 seconds.

[0062] Verification of error less than 0.1%: Using 10,000 sets of historical wafer manufacturing data to compare the output of primary and backup nodes, the mean NRMSE was 0.052%, the maximum was 0.098%, and all were less than 0.1%. This error comes from the difference in the order of floating-point operations and its actual impact on scheduling decisions is negligible (decision consistency rate 99.97%).

[0063] The above test data can be reproduced under the same or equivalent conditions.

[0064] The technical effects of the above solution are as follows: Addressing the high availability requirements of long-term continuous operation in wafer manufacturing scenarios, the three-active-node cluster architecture and stateful hot standby architecture can avoid system downtime caused by single-point failures, improving overall system availability to over 99.99%; the degradation self-healing and breakpoint resumption mechanisms of edge nodes solve the problems of data loss and equipment malfunction caused by network jitter or temporary failures of the upper-level control layer, ensuring that the production process is not interrupted due to local network anomalies; the master-slave synchronization mechanism of the AI ​​intelligent decision-making module achieves rapid fault takeover while ensuring the consistency of decision results, avoiding fluctuations in production processes due to model output fluctuations, ultimately improving the overall production stability and yield of wafer manufacturing.

[0065] In one embodiment, the wafer yield anomaly root cause localization model based on hybrid time-domain-frequency-domain feature fusion is further configured to output a predictive health degradation trajectory of the semiconductor manufacturing equipment associated with the root cause nodes, while outputting the root cause node set and node confidence. The predictive health degradation trajectory is a health index sequence for the next 30 time units predicted by a Long Short-Term Memory (LSTM) network based on the historical fault characteristics and current fused feature vector of the semiconductor manufacturing equipment. Each time unit is 1 minute, and the health index is a dimensionless index normalized to the [0,1] interval. The AI-native intelligent decision-making module also includes a production scheduling simulation and deduction unit based on multi-objective optimization. The simulation and deduction unit works in conjunction with the root cause localization model. The simulation and deduction unit uses the predictive health decline trajectory as a dynamic constraint and substitutes it into the multi-objective optimization model. The objective function of the multi-objective optimization is to minimize a weighted sum, where the weight coefficients are adjustable. The terms of the weighted sum include: total completion time, semiconductor manufacturing equipment load balancing degree, and expected downtime penalty term calculated based on the predictive health decline trajectory. The simulation and deduction unit solves the multi-objective optimization model through a reinforcement learning agent and outputs a production scheduling scheme. The production scheduling scheme is used to avoid assigning production tasks to bottleneck semiconductor manufacturing equipment, which refers to semiconductor manufacturing equipment whose health index will be lower than 0.3 within the next 30 minutes.

[0066] The working principle of the above technical solution is as follows: In addition to outputting root cause nodes, the root cause localization model also utilizes its internal LSTM branch to output a health index prediction sequence for each relevant device. The health index HI(t) is defined as follows: HI(t) = 1 - (current cumulative damage / total damage tolerance), where cumulative damage is calculated by the Coffin-Manson model based on the device's historical operating time, temperature cycle count, vibration accumulation, etc. The LSTM uses multi-sensor features from the past 60 minutes as input to predict the HI value at each time point in the next 30 minutes.

[0067] Statistical analysis of historical failure data from 50 etching machines in a wafer fab showed that when HI > 0.3, the probability of equipment failure within the next 30 minutes was 6%; when HI ≤ 0.3, the failure probability surged to 92%. 0.3 was defined as the bottleneck warning threshold.

[0068] The simulation unit receives a real-time work order list, equipment status, and HI (Health Index) prediction trajectory. The objective function of the multi-objective optimization model aims to minimize the weighted sum of three key performance indicators: the first is the total completion time, which is the latest completion time among all tasks, reflecting overall scheduling efficiency and delivery timeliness; the second is the equipment load rate variance, which characterizes the dispersion of the actual workload of each device relative to the average load, used to measure the balance of resource allocation and system robustness; the third is the expected downtime penalty, defined as the expected loss cost caused by the accelerated decline rate of the Health Index (HI) of bottleneck equipment after being assigned tasks, reflecting the proactive prevention and control of equipment degradation and unplanned downtime risks. These three are coordinated through three independent weights, representing the decision-maker's trade-offs regarding the priorities of delivery cycle, resource balance, and equipment reliability.

[0069] The reinforcement learning agent uses the PPO algorithm. Its state space includes device health indices, batches to be processed, and device occupancy status; its action space involves allocating the next batch to a specific device; and its reward function is a negative objective function. The agent is trained through interaction with a simulation environment, outputting a near-optimal scheduling scheme.

[0070] The technical effects of the above solution are as follows: On the one hand, it can identify semiconductor manufacturing equipment that is about to enter a high-risk failure phase while locating the root cause of yield abnormalities. By optimizing scheduling, it can avoid risks in advance, reduce the interference of unplanned downtime on production progress, and reduce yield loss and production downtime costs. On the other hand, through multi-objective optimization, it takes into account the needs of production delivery efficiency, equipment resource utilization and equipment health maintenance. It can flexibly adjust the weights according to the actual priority of the production site, adapt to the scheduling needs under different working conditions, solve the problem that traditional static scheduling cannot adapt to the dynamic health status changes of equipment, and improve the collaborative management resilience and intelligence level of the entire wafer manufacturing process.

[0071] In one embodiment, the system includes a switching and migration module and a three-active-node cluster architecture and a stateful hot standby architecture. During the gradual switching process, the three-active-node cluster in the three-active-node cluster architecture remains fully active, and the stateful hot standby architecture continuously performs memory checkpoint synchronization. When an abnormal instance is detected in any microservice group of the system during the gradual switching process, the gradual switching module triggers a rollback to the existing CIM system, and the three-active-node cluster automatically removes the abnormal instance and the remaining instance takes over.

[0072] The working principle of the above technical solution is as follows: During the gradual migration of the system from the existing CIM system to the system of this invention (e.g., 30% of the device control traffic has been switched), the three-active node cluster operates normally, with three instances of each microservice distributed across different racks; stateful hot standby continuously synchronizes the AI ​​model status. The gradual switching submodule has a built-in anomaly detector that monitors the response latency, error rate, and resource utilization of each microservice instance in real time. If any instance (e.g., instance 2 executing the production microservice) is found to have an error rate exceeding 1% or a response latency exceeding 200ms for 5 consecutive seconds, the instance is determined to be abnormal. The following actions are immediately triggered: Rollback: The gradual switching submodule sends a rollback command to the full switching and rollback submodule, switching the control traffic of all devices back to the existing CIM system within 60 seconds.

[0073] Cluster self-healing: The Raft leader of the three-live-node cluster removes the abnormal instance from the service registry, and the remaining two instances continue to process requests (the cluster can function normally since the majority is still present).

[0074] Alarms and Recovery: Notify operations and maintenance personnel to manually add the abnormal instance back to the cluster after it has been repaired.

[0075] The technical advantages of the above solution are as follows: it can ensure the stability of business during the transition phase of system migration and switching, and will not cause the entire system to shut down due to migration failure caused by local instance anomalies. It can also avoid the impact on the wafer production process through rapid rollback. At the same time, relying on the self-healing capability of the three active node cluster, it can maintain the high availability of the cluster after the anomaly is handled, and reserve the conditions for subsequent restart of migration and switching. It significantly reduces the risk of switching to the new version of CIM system, and takes into account both the flexibility of the migration process and the continuity requirements of production business.

[0076] In one embodiment, the root cause node set and node confidence output by the root cause localization model are fed back to the simulation inference unit in real time as the basis for dynamic adjustment of constraints in the multi-objective optimization model; after the optimized production scheduling scheme output by the simulation inference unit is executed, its execution result is fed back to the root cause localization model as an incremental sample for the next round of model training.

[0077] The working principle of the above technical solution is as follows: When the root cause localization model detects that a certain device or process step is the root cause of yield abnormalities (e.g., the confidence level of the RF matching node of the etching machine > 0.8), this information is transmitted to the simulation inference unit in real time. In the next round of scheduling, the simulation inference unit will reduce the task allocation weight of this device for a period of time in the future (i.e., treat it as a soft bottleneck), or add a penalty term to the objective function. Simultaneously, after the scheduling scheme output by the simulation inference unit is executed, the actual production results (such as whether further anomalies occur, batch completion time, and changes in equipment health) are recorded as labeled samples. These samples are periodically (e.g., every early morning) used to retrain or fine-tune the root cause localization model (e.g., as incremental training data for graph neural networks, updating node features and edge weights). Through continuous learning, the model's accuracy in identifying root causes continuously improves over time.

[0078] The technical effects of the above solution are as follows: through the two-way feedback closed loop of root cause localization and simulation scheduling, it can quickly respond to yield anomalies in the production process, dynamically adjust the scheduling strategy, and avoid the further spread of anomalies that could lead to a wider range of production line stoppages or yield losses. It can also continuously iterate and optimize the root cause localization model through actual production data, reduce the reliance on manually labeled samples, improve the model's adaptability to complex production scenarios, and ultimately achieve rapid convergence of anomalies in the wafer manufacturing process and simultaneous improvement of production efficiency and yield stability.

[0079] In one embodiment, an adaptation layer is further included to provide hardware and basic software adaptation for the edge access layer, factory collaborative management and control layer, and intelligent decision-making unit; the adaptation layer includes: The hardware adaptation submodule is used to adapt to CPU servers based on proprietary instruction sets (such as RISC-V, ARM perpetual license) or perpetual license instruction sets, as well as NVMe solid-state drives based on control chips. The operating system adaptation submodule is used to adapt to the operating system (such as a Linux distribution), which integrates real-time optimization kernel patches (such as PREEMPT_RT) to control the interrupt response latency to within 50 microseconds. The database adaptation submodule is used to adapt to distributed time-series databases (such as TDengine and InfluxDB) and implement a storage engine specifically for wafer manufacturing time-series data. The storage engine adopts a composite index structure based on time partitioning (a clustered index is built by device ID + timestamp, and a secondary index is built by parameter type). The middleware adapter submodule is used to adapt to web application server middleware (such as Tomcat and Jetty).

[0080] The working principle of the above technical solution is as follows: The adaptation layer is located between the system software and the underlying hardware / basic software. The hardware adaptation submodule uses a unified hardware abstraction layer (HAL) to call the CPU's SIMD instruction set to optimize matrix operations and utilizes the io_uring asynchronous I / O interface of the NVMe solid-state drive to improve data read and write throughput. The operating system adaptation submodule detects whether the kernel supports PREEMPT_RT. If not, it dynamically loads a real-time patch. By setting CPU affinity and interrupt isolation, it ensures the interrupt response latency of critical processes (such as the data bus and intelligent decision-making). Regarding the 50-microsecond test data: On a Phytium FT-2000 / 64 server and Kylin operating system (kernel 5.4, with PREEMPT_RT patch), running the data bus process of the system of this invention, the interrupt response latency was measured using the cyclictest tool. At 80% CPU load, the maximum latency was 47 microseconds, the average was 22 microseconds, and the 99.9th percentile was 52 microseconds. The 50-microsecond design value covers over 99% of scenarios and surpasses the general requirements for real-time control in wafer manufacturing (typically within 100 microseconds). The database adaptation submodule, tailored to the characteristics of wafer manufacturing data (partitioned by device and time, write-heavy and read-light, retaining the most recent 3 months of hot data), implements a composite index: the primary index is a time-partitioned index (device ID, timestamp), with each partition using a B+ tree; the secondary index is a parameter type, mapped via an inverted list. The middleware adaptation submodule provides a unified API gateway, adapting to deployment descriptors of different web containers.

[0081] The technical benefits of the above solution are as follows: Through a unified adaptation layer interface, this system is compatible with hardware environments of different architectures, supporting both domestically developed and controllable hardware and software ecosystems, as well as general industrial-grade hardware and software, reducing the adaptation costs of system migration and deployment; through kernel optimization for real-time scenarios, it meets the low-latency requirements of equipment control and data transmission in the wafer manufacturing process, providing a basic guarantee for real-time control; a customized storage engine for the characteristics of wafer manufacturing time-series data can further improve data read and write performance and reduce storage redundancy compared to general time-series databases, meeting the high-throughput storage requirements of massive time-series data in the manufacturing process; and a unified API gateway adapts to different application middleware, can adapt to the existing deployment architecture of different factories, and improves the deployment flexibility of the system.

[0082] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims and their equivalents, this invention is also intended to include these modifications and variations.

Claims

1. A CIM distributed collaborative management and control system for wafer manufacturing, characterized in that, include: The edge access layer, deployed at the wafer manufacturing equipment, includes multiple distributed edge nodes. Each edge node is used to receive and process real-time data from at least one semiconductor manufacturing equipment that is connected to the edge node, and dynamically adjusts the size of its data acquisition window and sampling step size, as well as the sliding step size of its feature extraction window, according to the load status from the factory collaborative management layer. The load status includes at least one of CPU utilization or task queue depth. The edge node also has a built-in protocol adaptive parsing module based on device fingerprints, which is used to: extract the device fingerprint of the access semiconductor manufacturing equipment; match the standard communication protocol template according to the device fingerprint; if the match fails, start the dynamic learning mode, listen to the communication messages between the device and the host computer, and reverse-engineer a private protocol parser for the semiconductor manufacturing equipment by analyzing the semantic features of the message frame structure, start character, end character and data length field; the private protocol parser is stored in association with the device fingerprint and uploaded to the unified protocol repository of the factory collaborative management layer; The factory collaborative management and control layer communicates with the edge access layer via a data bus. The factory collaborative management and control layer adopts a microservice architecture, deploying production execution microservice groups, equipment management microservice groups, and quality yield microservice groups. It is configured to receive real-time data from the edge access layer. Among them, the quality yield microservice group performs statistical process control analysis on the real-time data and adds event priority tags to the analyzed abnormal event data. The data bus routes the abnormal event data to the intelligent decision-making unit based on the event priority tags. The intelligent decision-making unit communicates with the factory collaborative management and control layer and has a built-in AI native intelligent decision-making module. The AI ​​native intelligent decision-making module is used to receive abnormal event data, locate the root cause of wafer yield abnormalities, and perform multi-objective optimization simulation and deduction of production scheduling schemes. The simulation and deduction results are fed back to the factory collaborative management and control layer, which generates control commands and sends them to the corresponding semiconductor manufacturing equipment through the edge access layer.

2. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 1, characterized in that, In dynamic learning mode, the protocol adaptive parsing module reverse-engineers a private protocol parser by: statistically analyzing the occurrence patterns of start and end characters in communication messages to determine message boundaries; analyzing the range of value changes of bytes at fixed offset positions to infer length fields, command codes, or data identifiers; and using entropy analysis to assist in identifying integer, floating-point, or string data types.

3. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 1, characterized in that, The data bus includes a semantic translation layer and a dynamic routing layer; The semantic transformation layer is pre-set with a core object data model for the wafer manufacturing field. It is used to semantically annotate and standardize heterogeneous data from the production execution microservice group, equipment control microservice group, quality yield microservice group and edge nodes according to the core object data model to obtain standardized data. Among them, the core objects include work orders, batches, wafers, equipment, process steps, measurement parameters and defects. The dynamic routing layer receives standardized data and distributes it to the production execution microservice group, equipment management microservice group, and quality yield microservice group that subscribe to the standardized data, according to the semantically labeled data type, data source, and preset priority rules, in descending order of priority. The dynamic routing layer also prioritizes routing abnormal event data to the intelligent decision-making unit based on the event priority tags.

4. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 1, characterized in that, The AI-native intelligent decision-making module is equipped with a wafer yield anomaly root cause localization model based on hybrid time-domain and frequency-domain feature fusion; the wafer yield anomaly root cause localization model include: The time-domain feature extraction submodule is used to calculate the statistical features of real-time data over time, including time-domain statistical features. The frequency domain feature extraction submodule is used to transform the time series of real-time data into the frequency domain through fast Fourier transform and extract the spectral energy distribution features. The feature fusion and attention submodule is used to concatenate time-domain features and frequency-domain features to form a fused feature vector, and apply a multi-head self-attention mechanism to calculate the contribution weights of different measurement parameters and different feature domains to the current yield anomaly. The root cause inference submodule is built on a graph neural network, where the nodes of the graph represent equipment, process steps and measurement parameters, and the edges of the graph represent physical connections or process flow relationships. The root cause inference submodule uses fused feature vectors and contribution weights as the initial features of nodes and edge weights, and infers the set of root cause nodes that lead to abnormal yield and the confidence level of each node through graph convolution operations.

5. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 1, characterized in that, It also includes a switching and migration module for enabling a smooth transition from the existing CIM system to the system. The switching and migration module includes: The data mirroring and dual-track parallel submodule is used to establish a two-way data mirroring channel with the existing CIM system database while maintaining the operation of the existing CIM system, and to synchronize the full production data in real time; and to simultaneously start the edge access layer, factory collaborative management and control layer and intelligent decision-making unit of the system, so that they receive the same real-time production data stream as the existing CIM system and perform no-load verification. The gradual switching submodule is used to periodically compare the key control instructions output by the system and the existing CIM system for the same batch of wafers, and calculate the instruction matching degree. The instruction matching degree is the percentage of instructions with completely identical content out of the total number of compared instructions. When the instruction matching degree reaches more than 95% for three consecutive batches, the gradual switching is initiated, and a preset proportion of the control flow of semiconductor manufacturing equipment is gradually switched from the existing CIM system to the system. The full switchover and rollback submodule is used to switch all control flows of the semiconductor manufacturing equipment to the system within a switching time of no more than 30 seconds in response to a switchover command after a preset time of gradual switching to stable operation; at the same time, the existing CIM system is kept running as a backup, and when an anomaly is detected in the system, it supports rolling back to the existing CIM system within a rollback time of no more than 60 seconds.

6. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 1, characterized in that, The edge access layer also performs node self-healing and local caching, which includes: when an edge node detects a network interruption with the factory collaborative management layer, it switches to a degraded operation mode, caches the collected data in local non-volatile storage, and continues to perform real-time control of the semiconductor manufacturing equipment based on the rules of the local cache; and after the network is restored, it resumes the transmission of the cached data with added timestamps to the factory collaborative management layer. The factory collaborative management and control layer is configured with a three-active-node cluster architecture. The three-active-node cluster architecture includes: each microservice deploys at least three instances, and the three instances are distributed in three independent physical racks or three independent availability zones; the data bus uses the Raft consensus algorithm to realize metadata synchronization and master node election, and when any instance or network partition fails, the remaining instances complete the master node re-election within no more than 5 seconds. The intelligent decision-making unit is configured with a stateful hot standby architecture. The stateful hot standby architecture includes: the model parameters and real-time running status of the AI ​​native intelligent decision-making module are synchronized to the standby node every 1 second in the form of memory checkpoints, and the standby node loads the most recent checkpoint and takes over the service within 5 seconds when the master node fails; the normalized root mean square error of the model prediction results output by the standby node after taking over is less than 0.1% relative to the output of the master node.

7. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 4, characterized in that, The wafer yield anomaly root cause localization model based on hybrid time-domain and frequency-domain feature fusion is also configured to output a predictive health degradation trajectory of the semiconductor manufacturing equipment associated with the root cause nodes, while outputting the root cause node set and node confidence. The predictive health degradation trajectory is a health index sequence for the next 30 time units predicted by a Long Short-Term Memory (LSTM) network based on the historical fault characteristics and current fused feature vector of the semiconductor manufacturing equipment. Each time unit is 1 minute, and the health index is a dimensionless index normalized to the [0,1] interval. The AI-native intelligent decision-making module also includes a production scheduling simulation and deduction unit based on multi-objective optimization. The simulation and deduction unit works in collaboration with the root cause localization model. The simulation and deduction unit uses the predictive health decline trajectory as a dynamic constraint and substitutes it into the multi-objective optimization model. The objective function of the multi-objective optimization is to minimize the weighted sum, where the weight coefficients are adjustable. The terms of the weighted sum include: total completion time, semiconductor manufacturing equipment load balancing degree, and expected downtime penalty term calculated based on the predictive health decline trajectory. The simulation and deduction unit solves the multi-objective optimization model through reinforcement learning agents and outputs a production scheduling scheme. The production scheduling scheme is used to avoid assigning production tasks to bottleneck semiconductor manufacturing equipment, which refers to semiconductor manufacturing equipment whose health index will be lower than 0.3 within the next 30 minutes.

8. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 1, characterized in that, The system includes the switching and migration module as described in claim 5 and the three-active node cluster architecture and stateful hot standby architecture as described in claim 6. During the gradual switching process, the three-active node cluster in the three-active node cluster architecture remains fully active, and the stateful hot standby architecture continuously performs memory checkpoint synchronization. When an abnormal instance is detected in any microservice group of the system during the gradual switching process, the gradual switching submodule triggers a rollback to the existing CIM system, and the three-active node cluster automatically removes the abnormal instance and is taken over by the remaining instances.

9. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 7, characterized in that, The root cause node set and node confidence output by the root cause localization model are fed back to the simulation inference unit in real time, serving as the basis for dynamic adjustment of constraints in the multi-objective optimization model; after the optimized production scheduling scheme output by the simulation inference unit is executed, its execution result is fed back to the root cause localization model, serving as incremental samples for the next round of model training.

10. The CIM distributed collaborative management and control system for wafer manufacturing according to claim 1, characterized in that, It also includes an adaptation layer, used to provide hardware and basic software adaptation for the edge access layer, factory collaborative management and control layer, and intelligent decision-making unit; the adaptation layer includes: The hardware adaptation submodule is used to adapt to CPU servers based on proprietary instruction sets or permanently licensed instruction sets, as well as NVMe solid-state drives based on control chips. The operating system adaptation submodule is used to adapt to the operating system, which integrates real-time optimization kernel patches to control interrupt response latency to within 50 microseconds. The database adaptation submodule is used to adapt to distributed time-series databases and implement a storage engine specifically for wafer manufacturing time-series data. The storage engine adopts a composite index structure based on time partitioning. The middleware adapter submodule is used to adapt to web application server middleware.

Images