Construction method of twin enhancement process agent for new display electronic component intelligent manufacturing

By constructing a gray box digital twin model and a process agent, the problem of insufficient optimization of process parameters in the manufacturing of new display panels was solved, realizing real-time optimization and anomaly compensation of process parameters, and improving production efficiency and quality stability.

CN123070342APending Publication Date: 2026-11-10BEIHANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611505371.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-24
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies lack systematic methods for optimizing process parameters in the manufacturing of new display panels. This leads to reliance on manual experience for process parameter adjustment, making it difficult to achieve rapid changeover and efficient production. Furthermore, the identification and compensation of abnormal operating conditions are lagging, and there is a lack of real-time dynamic optimization capabilities.

Method used

A gray-box digital twin model integrating a multi-physics mechanism model and a data-driven residual compensation network is constructed. A process agent is designed and trained and deployed in a virtual environment to achieve real-time optimization of process parameters and anomaly compensation, forming a closed-loop control.

Benefits of technology

It enables dynamic optimization of process parameters and real-time perception of quality status, improving the manufacturing efficiency and quality stability of Mini LED, Micro LED and HDI high-density interconnect electronic components, and supporting high-flexibility, high-precision and high-yield production.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070342A_ABST
    Figure CN123070342A_ABST
Patent Text Reader

Abstract

The application provides a twin enhancement process agent construction method for new display electronic component intelligent manufacturing, and relates to the technical field of electronic component intelligent manufacturing. The application runs sample expansion, virtual trial and error iteration and real-time state prediction capability throughout the whole process of agent training, deployment and evolution, changes process parameter optimization from experience driving to mechanism-data fusion driving, and changes quality control from post-detection to pre-prediction and active compensation, thereby effectively solving problems such as narrow process window, slow yield climbing, abnormal compensation lag and OEE improvement difficulty in the manufacturing process of Mini LED, Micro LED and HDI high-density interconnection electronic components, and providing reliable technical support for high-flexibility, high-precision, high-yield and high-efficiency intelligent manufacturing of new display key electronic components.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology for electronic components, and specifically to a method for constructing a twin-enhanced process agent for the intelligent manufacturing of novel display electronic components. Background Technology

[0002] The new display industry is a key development direction of the next-generation information technology industry. High-density interconnect electronic components, such as Mini LED backlight modules, Micro LED display units and driver substrates, and HDI high-density interconnect substrates, are the core basic units supporting high-performance display panels. Their manufacturing quality directly determines the resolution, brightness uniformity, and reliability of the terminal display products. Compared with conventional electronic components, the production process of high-density interconnect electronic components supporting new display panels is lengthy, covering multiple high-precision processing steps such as copper plating, electroplating, etching, laser drilling, and multilayer lamination. The overall manufacturing system is characterized by complex hierarchical structure, a large number of process parameters, and strong cross-linking of parameters between processes. At the same time, there are strong nonlinear coupling effects between the process parameters of each process, and the process parameter window range for stable production in a single process is narrow. The production process is highly sensitive to minor disturbances in the environment, equipment, and materials. The overall manufacturing process exhibits complex production characteristics of multiple variables and strong time-varying characteristics, which places extremely high demands on the precise control and global optimization of process parameters.

[0003] Currently, the industry still uses traditional process control methods based on fixed formulations or statistical process control (SPC) for the production of the aforementioned high-density interconnected electronic components. In actual production scenarios, parameter configuration during product changeover and process debugging phases relies entirely on the on-site practical experience of senior engineers, using a trial-and-error approach of multiple trial productions and repeated debugging and error correction to match process parameters. In routine mass production, the industry still relies on two types of static control methods—fixed process formulations and statistical process control—to monitor processes, with some scenarios supplemented by conventional machine learning models for simple quality data analysis.

[0004] However, existing process control solutions have many inherent shortcomings: parameter tuning during product changeover relies excessively on manual experience and lacks standardized optimization approaches; statistical process control can only perform post-production statistical analysis on single-process production data and cannot decompose the complex parameter coupling relationships between multiple processes; conventional machine learning models rely on large batches of high-quality labeled samples, can only adapt to single fixed production conditions, and have extremely poor generalization ability. In summary, existing process control systems are all fragmented, single-point parameter adjustment methods, and have not built a unified parameter analysis, matching, and iterative optimization system for the entire production process. Currently, there is a general lack of systematic process parameter optimization methods in the manufacturing of new display panels, making it difficult to adapt to the complex and ever-changing panel manufacturing needs. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components, solving the technical problem of the lack of systematic process parameter optimization methods in the current manufacturing process of novel display panels.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for constructing a twin-enhanced process agent for intelligent manufacturing of novel display electronic components, comprising: Acquire and process multi-source heterogeneous data for new display electronic component production lines to obtain structured process datasets; A gray-box digital twin model integrating a multiphysics mechanism model and a data-driven residual compensation network is constructed, and the gray-box digital twin model is dynamically calibrated online using the structured process dataset. Design a process agent for rapid changeover manufacturing, defining its perception layer, intelligent reasoning layer, decision-making layer, and multi-objective reward function; Offline training and verification of the process agent are carried out in the virtual environment constructed by the gray box digital twin model. Through simulation sample generation and abnormal operating condition enhancement training, the process agent learns process parameter optimization and abnormal operating condition compensation strategies. The validated process agent is deployed to the edge node, and receives the future predicted state and compensation signal of the gray box digital twin model in real time through virtual-real synchronization, forming a closed-loop control for the actual production line.

[0007] Preferably, multi-source heterogeneous data for new display electronic component production lines are acquired and processed to obtain a structured process dataset, including: Obtain real-time multi-source production line data corresponding to the current work order, and perform time-series alignment and aggregation on the real-time multi-source production line data; Retrieve the layered structure, process route, material batches, historical production records of similar products, and final quality judgment data related to the current work order from the historical database to form a related dataset; We perform missing data filling, anomaly removal, time synchronization, and feature extraction on real-time production line multi-source data and related datasets after time-series alignment and aggregation to construct a structured process dataset.

[0008] Preferably, a gray-box digital twin model integrating a multiphysics mechanism model and a data-driven residual compensation network is constructed, including: For each key process of the novel display electronic component, a mechanism model adapted to the process characteristics of each process is constructed. The mechanism model covers multi-dimensional physical mechanisms, including reaction kinetics, flow field mass transfer, electric field transport, laser thermal phase transition and thermo-mechanical coupling. A residual compensation network is constructed to compensate for the deviation between the learning mechanism model output and the actual measured values. Combining the learning mechanism model and the residual compensation network, a gray-box digital twin model is obtained. The complete output of the gray-box digital twin model is as follows: in, This represents the quality index of the mechanistic model's predictions. For multiphysics mechanism models, For physical parameters; This represents the estimated value of the prediction residuals for the mechanistic model. For data-driven residual compensation networks, for The hidden state at any given moment. For network parameters, for The input condition vector at time step; For gray box digital twin model The predicted quality value output at any given time.

[0009] Preferably, the intelligent inference layer adopts a hybrid architecture of gray-box digital twin model predictive control and deep reinforcement learning, and outputs process adjustment actions, specifically including: A gray-box digital twin model is used for rolling optimization within a finite prediction time domain to generate baseline actions that conform to physical laws and equipment constraints. A deep reinforcement learning policy network outputs compensation actions based on the current state and the predicted future state to handle model mismatch, unknown disturbances, and multi-objective trade-offs. The expressions for the output process adjustment actions include: in, express The process adjustment actions output in real time; for The baseline action sequence is obtained through rolling optimization based on the gray box digital twin model, providing a nominal control trajectory that conforms to physical laws. For parameters Deep reinforcement learning policy network, enter Moment State Forward prediction with digital twins Implicit state prediction of the step It outputs intelligent compensation quantities for model mismatch, unknown disturbances, and multi-objective trade-offs.

[0010] Preferably, the multi-objective reward function includes: in, for Momentary rewards; This is a quality compliance indicator function. It takes a value of 1 when all online quality indicators are within the quality specification limits, and 0 otherwise. express The vector of online quality metrics at any given time; for Energy or material consumption costs at any given time; The response delay from when an anomaly is detected to when the agent issues an effective compensation action; This is an indication function for whether a quality abnormality or equipment alarm has occurred. express Check for quality abnormalities or equipment alarms at all times; , , , These are the weighting coefficients for each objective.

[0011] Preferably, the offline training and verification of the process agent in the virtual environment constructed by the gray-box digital twin model, through simulation sample generation and abnormal operating condition enhancement training, enables the process agent to learn process parameter optimization and abnormal operating condition compensation strategies, including: Multiple parallel simulation environments were constructed using gray-box digital twin models; The process agent is continuously trained from standard conditions to abnormal operating conditions through multiple parallel simulation environments, specifically including: In the initial training phase, only standard operating conditions are used to enable the process agent to learn basic parameter stability control and quality achievement strategies; complex operating conditions are gradually introduced, and the range of disturbances is gradually expanded. Training, convergence, and security verification of the process agent; The policy network parameters in the validated process Agent are engineered, solidified, and managed for reuse.

[0012] Preferably, the twin-enhanced process agent construction method further includes: Based on online operational data, the parameters of the gray box digital twin model and the process agent are periodically fine-tuned.

[0013] Secondly, the present invention provides a twin-enhanced process agent construction system for intelligent manufacturing of novel display electronic components, comprising: The data acquisition module is used to acquire and process multi-source heterogeneous data for the production line of new display electronic components to obtain a structured process dataset. A digital twin engine is used to construct a gray-box digital twin model that integrates a multiphysics mechanism model and a data-driven residual compensation network, and to perform online dynamic calibration of the gray-box digital twin model using a structured process dataset. The process agent inference core is used to design process agents for rapid changeover manufacturing, defining its perception layer, intelligent inference layer, decision-making layer, and multi-objective reward function. Among them, the intelligent inference layer adopts a hybrid architecture of gray-box digital twin model predictive control and deep reinforcement learning, outputs process adjustment actions, and sets multi-objective reward functions. The virtual-real synchronization middleware is used to conduct offline training and verification of the process agent in the virtual environment constructed by the gray box digital twin model. Through simulation sample generation and abnormal operating condition enhancement training, the process agent learns process parameter optimization and abnormal operating condition compensation strategies. The application service module is used to deploy the validated process agent to edge nodes, and receive the future predicted status and compensation signals of the gray box digital twin model in real time through virtual-real synchronization, forming a closed-loop control for the actual production line.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for constructing a twin-enhanced process agent for intelligent manufacturing of novel display electronic components, wherein the computer program causes a computer to execute the twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described above.

[0015] Fourthly, the present invention provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for executing a twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described above.

[0016] (III) Beneficial Effects This invention provides a method for constructing a twin-enhanced process agent for the intelligent manufacturing of novel display electronic components. Compared with existing technologies, it has the following advantages: This invention integrates the sample expansion, virtual trial-and-error iteration, and real-time state prediction capabilities of digital twins into the entire process of agent training, deployment, and evolution. It transforms process parameter optimization from experience-driven to mechanism-data fusion-driven, and quality control from post-event detection to pre-event prediction and proactive compensation. This effectively solves problems such as narrow process windows, slow yield ramp-up, delayed anomaly compensation, and difficulties in improving OEE in the manufacturing of Mini LED, Micro LED, and HDI high-density interconnect electronic components. It provides reliable technical support for the high-flexibility, high-precision, high-yield, and high-efficiency intelligent manufacturing of key electronic components for new displays. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a block diagram of a twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components according to an embodiment of the present invention; Figure 2 This is an architecture diagram of a twin-enhanced process agent construction system for intelligent manufacturing of novel display electronic components, as described in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This application provides a twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components, which solves the technical problem that there is a general lack of systematic process parameter optimization methods in the current manufacturing process of novel display panels, and realizes dynamic optimization of process parameters, real-time perception of quality status and adaptive compensation for abnormal working conditions.

[0021] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: The current manufacturing process mainly faces the following prominent problems: First, the process of changing models and adjusting parameters relies heavily on manual experience and lacks systematic optimization methods. In the context of rapid iteration of Mini / Micro LED and multiple HDI product models, the adjustment of process parameters mainly relies on engineering experience and trial and error, resulting in long changeover cycles and low parameter adjustment efficiency. This makes it difficult to support the flexible manufacturing needs of multiple varieties and small batches, and the overall equipment efficiency (OEE) has been difficult to improve in the long term.

[0022] Second, the correlation mechanism between process parameters and quality results is unclear, and there is a lack of real-time dynamic optimization capabilities. Due to the complex coupling of multi-process parameters and the existence of strong nonlinear relationships, traditional statistical analysis methods are unable to accurately characterize the mapping relationship between "parameter-structure-defect", resulting in slow first-piece yield ramp-up, large fluctuations in process capability index, and difficulty in achieving stable process window control.

[0023] Third, the identification and compensation of abnormal operating conditions are lagging behind, and there is a lack of predictive control capabilities. In actual production, abnormal phenomena such as chemical aging, electrode passivation, etching deviation, and drilling misalignment are concealed and gradual. Traditional detection methods often only respond passively after the defects have already caused a batch impact, lacking a predictive and adaptive compensation mechanism for future conditions.

[0024] As described above, existing process control methods based on fixed recipes or statistical process control (SPC) are difficult to adapt to the complex manufacturing environment of new display electronic components, which involves multivariate coupling, strong time-varying characteristics, and cross-process propagation. At the same time, ordinary machine learning methods rely on high-quality labeled samples, but the high cost of trial and error and the scarcity of abnormal samples in the electronic component manufacturing process make it difficult for models to be effectively trained and continuously optimized online.

[0025] Therefore, existing technologies generally suffer from problems such as "strong reliance on experience, weak dynamic optimization, insufficient cross-process collaboration, and lack of online adaptive capability" when facing the rapid changeover manufacturing requirements of Mini / Micro LED and high-density interconnected electronic components.

[0026] To address the aforementioned shortcomings, this invention provides a twin-enhanced process agent (intelligent agent) construction method for intelligent manufacturing of novel display electronic components. This method integrates process mechanism knowledge, data-driven learning, and real-time dynamic decision-making. It enables dynamic optimization of process parameters, real-time perception of quality status, and predictive self-healing control of abnormal operating conditions during the manufacturing process of novel display electronic components, thereby supporting the overall improvement of the rapid changeover manufacturing capability of high-density interconnect electronic components.

[0027] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0028] This invention provides a method for constructing a twin-enhanced process agent for intelligent manufacturing of novel display electronic components, such as... Figure 1 As shown, it includes: S1. Acquire and process multi-source heterogeneous data for new display electronic component production lines to obtain structured process datasets; S2. Construct a gray-box digital twin model that integrates a multi-physics mechanism model and a data-driven residual compensation network, and perform online dynamic calibration of the gray-box digital twin model using a structured process dataset. S3. Design a process agent for rapid changeover manufacturing, defining its perception layer, intelligent reasoning layer, decision-making layer, and multi-objective reward function; S4. In the virtual environment constructed by the gray box digital twin model, the process agent is trained and verified offline. Through simulation sample generation and abnormal operating condition enhancement training, the process agent learns the process parameter optimization and abnormal operating condition compensation strategies. S5. Deploy the verified process agent to the edge node, and receive the future prediction status and compensation signal of the gray box digital twin model in real time through virtual-real synchronization to form a closed-loop control for the actual production line.

[0029] This invention integrates the sample expansion, virtual trial-and-error iteration, and real-time state prediction capabilities of digital twins throughout the entire process of agent training, deployment, and evolution. This transforms process parameter optimization from experience-driven to mechanism-data fusion-driven, and quality control from post-event detection to pre-event prediction and proactive compensation. As a result, it effectively solves problems such as narrow process windows, slow yield ramp-up, delayed anomaly compensation, and difficulties in improving OEE in the manufacturing of Mini LED, Micro LED, and HDI high-density interconnect electronic components. This provides reliable technical support for the high-flexibility, high-precision, high-yield, and high-efficiency intelligent manufacturing of key electronic components for new displays.

[0030] It should be noted that, in this embodiment of the invention, "twin enhancement" specifically refers to a triple enhancement mechanism based on a digital twin system for the learning and decision-making capabilities of a process agent, including: (1) Sample augmentation mechanism (i.e., sample expansion): High-fidelity gray-box digital twin models are used to perform multiphysics simulations of Mini / Micro LED and HDI surface treatment processes, generating large-scale virtual process data to expand the limited and costly labeled samples in real production lines, thereby supporting offline supervised learning and reinforcement learning training of process agents.

[0031] (2) Trial and error enhancement mechanism (i.e., virtual trial and error iteration): By deploying process agents in a digital twin virtual environment and conducting large-scale interactive trial-and-error training through reinforcement learning, the optimal control strategies for key processes such as copper plating, electroplating, etching, and laser processing can be safely explored without affecting the real production line.

[0032] (3) Real-time prediction enhancement mechanism (i.e., real-time state prediction): During the online operation phase, the digital twin model maintains high-frequency synchronous evolution with the physical production line, makes short-term time-domain predictions of future quality evolution trends, and feeds back the predicted implicit states and compensation information to the process agent in real time, thereby enhancing the foresight, stability and robustness of its decision-making.

[0033] The following is a detailed explanation of each step in the twin-enhanced agent construction method: In step S1, multi-source heterogeneous data for the new display electronic component production line is acquired and processed to obtain a structured process dataset. The specific implementation process is as follows: It should be noted that the embodiments of the present invention are executed by a twin enhancement process agent construction system for intelligent manufacturing of novel display electronic components (including a computer, intelligent agent, etc., that constructs a twin enhancement process agent for intelligent manufacturing of novel display electronic components). For ease of description, it will be referred to as the system below.

[0034] S101. Obtain the real-time multi-source production line data corresponding to the current work order, and perform time-series alignment and aggregation on the real-time multi-source production line data. Specifically, this includes: The system accesses real-time databases to retrieve equipment status signals, process parameters, online quality inspection data, and environmental auxiliary data from the new display electronic component production line corresponding to the current work order. This data is then aligned and aggregated according to product model, production batch, work order number, process number, and timestamp. The process is described in detail below: The system accesses real-time data from the production line of new display electronic components through sensors, programmable logic controllers, distributed control systems, manufacturing execution systems, quality inspection equipment, and historical database interfaces. For the copper plating process, it collects data such as solution temperature, pH value, conductivity, copper ion concentration, reducing agent concentration, circulating pump speed, and substrate conveying speed. For the electroplating process, it collects data such as current density, voltage waveform, pulse duty cycle, plating solution flow rate, additive replenishment amount, anode and cathode status, and bath temperature. For the laser drilling process, it collects data such as laser pulse energy, pulse width, repetition frequency, scanning speed, focal point position, hole diameter detection value, and hole position deviation. For the multilayer lamination and HDI interlayer interconnection processes, it collects data such as stack structure, copper thickness distribution, substrate expansion and contraction, lamination temperature curve, pressure curve, target misalignment, and interlayer misalignment detection results.

[0035] The system integrates online AOI (Automated Optical Inspection), AXI (Automatic X-ray Inspection), visual measurement, X-ray fluorescence thickness measurement, hole position detection, film thickness detection, and offline sampling inspection data to generate quality evaluation data synchronized with the process. This quality evaluation data includes, but is not limited to, coating thickness uniformity, copper plating integrity, hole wall roughness, hole diameter consistency, hole position deviation, linewidth / spacing deviation, short circuits, open circuits, voids, delamination, foreign matter contamination, and interlayer alignment errors. For Micro LED and high-end HDI products, the system focuses on recording micron-level circuit defects, microvia misalignment, cumulative interlayer displacement, and local deformation characteristics after multiple laminations.

[0036] S102. Retrieve the layered structure, process route, material batches, historical production records of similar products, and final quality judgment data related to the current work order from the historical database to form a related dataset. Specifically, this includes: The system retrieves the product model, stack-up structure, process route, material batch, process formula, equipment status, historical similar product process parameters, and final quality judgment results corresponding to the current work order from the Manufacturing Execution System and the process knowledge base. This data is linked using "Product Model—Work Order—Batch—Process—Equipment—Timestamp" as the primary index, forming a contextual dataset that describes changes in process parameters, quality ramp-up processes, and anomaly evolution paths in the context of rapid model changeover.

[0037] S103. Perform missing data imputation, anomaly removal, time synchronization, and feature extraction on the time-series aligned and aggregated real-time production line multi-source data and related datasets to construct a structured process dataset organized in the form of "batch—process—time—variable." This structured process dataset includes historical standard condition batch data and multi-batch, multi-model, and multi-condition production data. Specifically, this includes: To address issues such as inconsistent sampling frequencies across different devices, quality inspection delays, sensor noise, and missing local data, the system first performs time synchronization, batch alignment, sliding window aggregation, and outlier removal. Then, it employs a combination of statistical and process rules to identify sensor faults, communication interruptions, and clearly abnormal data. For short-term missing data, it uses neighborhood interpolation, mechanistic trend extrapolation, or historical similar conditions to complete the data. After preprocessing, the system extracts process features, such as fluctuation amplitude, kurtosis, skewness, and settling time from electroplating current waveforms; activity decay indicators from copper plating solution states; energy density and aperture deviation mapping features from laser drilling data; and expansion / contraction direction, local offset, and cumulative deformation trends from lamination data. Finally, a structured process dataset is generated, organized in the form of "batch—process—time—variable."

[0038] Furthermore, based on the data source, operating condition characteristics, and model training purpose, the structured process dataset is divided into two categories: historical standard condition batch data and multi-batch, multi-model, and multi-operating condition production data.

[0039] Among them, historical standard batch data refers to historical production batch data where the equipment is in stable condition, the chemical solution is in normal condition, the raw materials fluctuate little, and the final quality inspection is qualified. It is used to perform benchmark calibration on key physical parameters such as reaction rate constant, mass transfer coefficient, equipment efficiency coefficient, and thermal expansion coefficient in the mechanism models of processes such as copper plating, electroplating, laser drilling, and lamination.

[0040] Multi-batch, multi-model, and multi-condition production data refers to historical production data covering different product models, laminated structures, material batches, equipment status, and process disturbance conditions. This includes conditions such as drug aging, electrode passivation, additive imbalance, laser focus drift, abnormal compression expansion and contraction, and cumulative interlayer offset. This data is used to train a residual compensation network to learn nonlinear disturbances and slow time-varying deviations that are difficult to characterize by the learning mechanism model.

[0041] Step S1 aims to collect and organize data that reflects the process status, quality status, and changeover context, taking into account the characteristics of multi-process, multi-equipment, and multi-parameter coupling in the manufacturing process of Mini LED, Micro LED, and HDI high-density interconnect substrates, so as to provide a highly consistent data foundation for subsequent digital twin modeling and process agent training.

[0042] In step S2, a gray-box digital twin model integrating a multiphysics mechanism model and a data-driven residual compensation network is constructed, and the gray-box digital twin model is dynamically calibrated online using a structured process dataset. The specific implementation process is as follows: S201. For each key process of the novel display electronic component, a mechanism model adapted to the process characteristics of each process is constructed. This mechanism model encompasses multi-dimensional physical mechanisms such as reaction kinetics, flow field mass transfer, electric field transport, laser-induced thermal phase transition, and thermo-mechanical coupling. Specifically, it includes: For the copper plating process, a mechanistic model was established that includes chemical reaction kinetics, bath flow field, mass transfer and diffusion, and deposition rate to describe the deposition behavior of copper ions under the action of complexing agents, reducing agents, and catalytic surfaces.

[0043] For the electroplating process, an electric field-flow field-mass transport coupling model is established to describe the relationship between current density distribution, plating solution circulation state, additive concentration and coating thickness uniformity.

[0044] For laser drilling, a model is established to predict hole diameter, hole wall quality, and hole position deviation based on laser heat input, material absorption, heat conduction, phase transformation ablation, and hole shape evolution.

[0045] For the HDI multilayer lamination process, a thermo-mechanical coupled deformation model is established to describe the relationship between substrate expansion and contraction, copper layer constraint, lamination temperature and pressure curves, and interlayer alignment error. The above-mentioned mechanism model is uniformly represented as follows: in, for The input condition vector at any given time includes adjustable process parameters such as temperature, current density, and laser energy, as well as incoming material status such as substrate thickness and copper thickness distribution. This refers to the set of physical parameters in the mechanistic model, such as reaction rate constant, activation energy, diffusion coefficient, electrical conductivity, absorption coefficient, etc. Quality indicators predicted by the mechanistic model, such as deposition rate, coating thickness distribution, and pore location deviation.

[0046] S202. Construct a residual compensation network to compensate for the deviation between the output of the learning mechanism model and the actual measured values. Combine the learning mechanism model and the residual compensation network to obtain a gray-box digital twin model. Specifically, this includes: Due to factors such as chemical aging, electrode passivation, material batch fluctuations, equipment wear, sensor drift, and local environmental disturbances in the manufacturing process of novel display electronic components, pure mechanistic models cannot fully depict the behavior of the actual production line. Therefore, this embodiment of the invention constructs a residual compensation network to learn the deviation between the output of the mechanistic model and the actual measured values.

[0047] The residual compensation network can be any one of long short-term memory networks, gated recurrent unit networks, temporal convolutional networks, or physical information neural networks, and its input includes... Input condition vector at time step , Hidden state at time (i.e., historical hidden states) and transformation context features, the output is the model residual estimate: in, These are the trainable parameters of the network. The output of the network. This represents an estimate of the prediction residuals from the mechanistic model. Therefore, the complete output of the gray-box digital twin model is: This model can simultaneously reflect the laws of process mechanism and the laws of actual production line data, and is suitable for quality prediction and parameter optimization in the scenario of rapid changeover of new display electronic components.

[0048] S203. Use historical standard batch data from the structured process dataset to perform benchmark calibration on the key physical parameter θ in the mechanism model, and use multi-batch, multi-model, and multi-condition production data from the structured process dataset to train the trainable parameters of the network. This yields the trained gray-box digital twin model. Specifically, it includes: Mechanism parameter calibration and joint training. The system first uses historical standard-condition batch data to perform benchmark calibration of the key physical parameter θ in the mechanism model; then, it uses production data from multiple batches, models, and operating conditions to train the residual compensation network parameters. The training objective is to minimize the error between the predicted values ​​and the actual detected values ​​of the twin model. Mean squared error, weighted quality error, or a joint loss function of multiple indicators can be used. For key indicators such as interlayer alignment error, coating thickness uniformity, and hole position deviation, higher weights can be assigned to improve the model's predictive ability for key quality indicators in high-precision manufacturing of novel displays.

[0049] S204. Deploy the trained gray-box digital twin model, and periodically calculate the deviation between the predicted and measured values ​​based on online quality inspection data. Also, analyze the time-varying sensitive parameters in θ of the gray-box digital twin model. Lightweight updates are performed on local network parameters. Specifically, this includes: After model deployment, the system periodically calculates the deviation between predicted and measured values ​​based on online quality inspection data, and adjusts the time-varying sensitive parameters in θ. The system performs lightweight updates to local network parameters. For example, when a continuous decrease in the copper plating rate is detected, the system can update parameters related to the chemical activity; when the uniformity of the plating thickness decreases, the system can update electrode efficiency and flow field compensation parameters; when the interlayer offset increases after lamination, the system can update thermal expansion and local deformation parameters. The dynamically calibrated gray-box digital twin model can continuously reflect the latest status of the production line, providing a reliable predictive environment for the process agent.

[0050] Step S2 aims to construct a gray-box digital twin model capable of describing the surface treatment and interlayer interconnection processes of novel display electronic components. This model integrates a multiphysics mechanism model and a data-driven residual compensation model, preserving the physical interpretability of the process while possessing the ability to fit complex disturbances, slow time-varying drift, and unknown nonlinear factors.

[0051] In step S3, a process agent for rapid changeover manufacturing is designed, defining its perception layer, intelligent inference layer, decision-making layer, and multi-objective reward function. The intelligent inference layer employs a hybrid architecture combining gray-box digital twin model predictive control and deep reinforcement learning, outputting process adjustment actions and setting the multi-objective reward function. The specific implementation process is as follows: Perception Layer Design: The perception layer obtains fused state information from the actual production line and the gray-box digital twin model, and constructs the agent's state vector at time t. .

[0052] The state vector includes real-time sensor data, online quality inspection data, product context data, and implicit states estimated by the gray-box digital twin model. Directly measurable variables (i.e., real-time sensor data and online quality inspection data) include temperature, current, voltage, flow rate, pressure, laser energy, film thickness, pore size, pore position deviation, and interlayer alignment deviation. Product context data includes product model, stack structure, process route, material batch, and changeover frequency. Implicit states include residual active reagent, anode effective area attenuation rate, plating solution dispersion index, laser heat-affected zone expansion trend, and pressing deformation accumulation trend. The state vector can be represented as: in, The value is the direct measurement of the sensor. This is the online quality inspection value. For contextual features, The hidden states estimated for the gray-box digital twin model.

[0053] Intelligent Inference Layer Design: The intelligent inference layer adopts a hybrid architecture that combines gray-box digital twin model predictive control with deep reinforcement learning.

[0054] Among them, model predictive control utilizes a gray-box digital twin model to perform rolling optimization within a finite prediction time domain, generating baseline actions that conform to physical laws and equipment constraints; the deep reinforcement learning policy network outputs compensation actions based on the current state (i.e., the state at time t) and the future predicted state, used to handle model mismatch, unknown disturbances, and multi-objective trade-offs. The final action is defined as: in, To provide a nominal control trajectory that conforms to physical laws for the baseline motion sequence obtained by rolling optimization based on the gray box digital twin model; For parameters A deep reinforcement learning policy network, inputting the current state. Forward prediction with digital twins Implicit state prediction of the step It outputs intelligent compensation quantities for model mismatch, unknown perturbations, and multi-objective trade-offs. This hybrid decision-making mechanism enables the agent to possess both forward-looking planning capabilities based on mechanistic models and adaptive compensation capabilities based on data learning.

[0055] Decision-making layer and action space design: The decision-making layer maps the actions output by the intelligent reasoning layer into actual equipment control commands.

[0056] For the copper plating process, the action space may include the solution temperature setpoint, copper salt replenishment amount, reducing agent replenishment amount, circulation pump speed, and substrate conveying speed; For the electroplating process, the action space may include rectifier current, voltage, pulse duty cycle, additive replenishment amount, plating solution circulation flow rate, and anode compensation parameters; For laser drilling, the motion space can include laser energy, pulse width, repetition frequency, scanning speed, focal position, and hole position compensation. For the lamination and interlayer alignment processes, the action space may include the lamination temperature profile, pressure profile, pre-compensation ratio, local alignment compensation amount, and expansion / contraction correction parameters.

[0057] All actions are constrained by equipment safety boundaries, process specification boundaries, and rate of change to ensure that they will not impact the actual production line.

[0058] Multi-objective reward function design: To guide the agent in learning the optimal policy that satisfies multiple objective trade-offs, the reward function is defined as follows: in, This is a quality compliance indicator function. It takes a value of 1 when all online quality indicators are within the quality specification limits, and 0 otherwise. Indicates quality specifications. express The vector of online quality metrics at any given time; for Energy or material consumption cost at any given moment (i.e., the current time step); The response delay from when an anomaly is detected to when the agent issues an effective compensation action; This is an indication function for whether a quality abnormality or equipment alarm has occurred. express Check for quality abnormalities or equipment alarms at all times; , , , The weighting coefficients for each objective are pre-set by process experts based on production strategies such as prioritizing quality, efficiency, or a balance; these weighting coefficients are non-negative. This step constructs a process agent with sensing, intelligent reasoning, and autonomous decision-making capabilities, enabling it to dynamically output process adjustment actions and anomaly compensation strategies for the rapid changeover production processes of Mini LED, Micro LED, and HDI electronic components.

[0059] In step S4, the process agent is trained and validated offline in the virtual environment constructed by the gray-box digital twin model. Through simulation sample generation and abnormal operating condition enhancement training, the process agent learns process parameter optimization and abnormal operating condition compensation strategies. The specific implementation process is as follows: S401. Construct multiple parallel simulation environments using gray-box digital twin models. Specifically, this includes: The digital twin engine instantiates the gray-box digital twin model into multiple parallel simulation environments, each corresponding to a product model, process route, equipment status, or disturbance condition. The system can simulate the surface treatment processes of Mini LED backlight panels, Micro LED driver substrates, and HDI multilayer boards under different batches and changeover conditions, generating a large number of state-action-quality result interaction trajectories. The virtual samples generated through the twin environments effectively expand upon abnormal samples and boundary condition samples that are difficult to obtain in real production lines.

[0060] S402. Continuously train the process agent from standard conditions to abnormal operating conditions using multiple parallel simulation environments. Specifically, this includes: In the initial training phase, the system was only loaded with standard conditions: fresh chemicals, normal equipment, and stable incoming materials. This allowed the process agent to learn basic parameter stability control and quality compliance strategies. Subsequently, complex operating conditions such as chemical aging, electrode passivation, additive imbalance, laser focus drift, material batch fluctuations, abnormal compression expansion and contraction, sensor drift, and online detection delays were gradually introduced, and the range of disturbances was gradually expanded. For high-precision products such as HDI and Micro LED, the training process focused on strengthening the ability to predict and compensate for interlayer alignment errors, correct micro-hole misalignment, and control coating uniformity.

[0061] S403. Perform training convergence and security verification on the process agent. Specifically, this includes: The process agent learns optimal or near-optimal strategies under different operating conditions by repeatedly interacting with the twin environment to maximize cumulative rewards. When the cumulative reward curve stabilizes and key quality indicators such as compliance rate, anomaly recovery time, and action smoothness meet preset requirements, the strategy is considered to have initially converged. Subsequently, the system validates the strategy using historical batch data not used in training, expert-defined boundary conditions, and safety limit scenarios. If the process agent's output actions violate equipment hard constraints, process specifications, or safety thresholds, the strategy must not proceed to the online deployment stage and must return to the training phase for further optimization.

[0062] S404. The policy network parameters in the validated process Agent are engineered, solidified, and managed for reusability. Specifically, this includes: The policy network parameter ψ in the validated process agent is fixed as a deployable version, and its applicable product family, process route, equipment range, training data version, twin model version, and safety boundary parameters are recorded. The system supports policy version rollback and canary deployment. When introducing new products or changing equipment status, a similar product policy can be selected as the initial policy for hot start, thereby shortening the parameter tuning time after rapid changeover.

[0063] This step involves offline training and validation of the process agent in a digital twin virtual environment, enabling it to achieve stable optimization, rapid changeover, and anomaly handling capabilities before deployment to a real production line.

[0064] In step S5, the validated process agent is deployed to edge nodes. Through real-time virtual-physical synchronization, it receives the future predicted state and compensation signals from the gray-box digital twin model, forming a closed-loop control system for the actual production line. The specific implementation process is as follows: The S501 digital twin engine continuously receives real-time production line data through a virtual-physical synchronization middleware and updates its status based on online measured data as boundary conditions. Specifically, this includes: When sensor data, quality inspection data, or work order context changes, the gray-box digital twin model synchronously updates its internal state and key time-varying parameters, and predicts the quality evolution trend within a preset time window. For the electroplating process, the prediction includes changes in coating thickness distribution and uniformity; for the laser drilling process, the prediction includes hole diameter deviation and hole position offset trends; for the multilayer lamination process, the prediction includes the evolution of local expansion and contraction and interlayer alignment errors.

[0065] S502. Deploy the process agent on an edge computing node close to the production line to receive the current status in a low-latency manner. and digital twin forward prediction Implicit state prediction of the step .

[0066] Within each decision cycle, the process agent first generates baseline actions based on a Siamese model using the MPC (Model Predictive Control) module, and then generates compensating actions using a reinforcement learning policy network. The two actions are then fused to obtain the process adjustment actions. After the process adjustment actions are verified by the security filter layer, they are sent to the PLC (Programmable Logic Controller), DCS (Distributed Control System), or motion control system via OPC-UA (Open Platform Communications Unified Architecture), industrial Ethernet, or equipment control interface to drive the actuators to complete the process parameter adjustment.

[0067] It should be noted that, in the specific implementation process, to ensure the safe online operation of the system, this embodiment of the invention sets a safety filtering layer between the output actions of the process agent and the execution by the equipment. This layer sets upper and lower limits for actions, rate of change constraints, and anomaly rollback logic based on equipment capabilities, process specifications, and quality risks. When the output actions of the process agent exceed the safe range, the system automatically truncates them to the safe boundary or switches to conservative MPC actions; when continuous anomalies are detected or the model confidence is insufficient, the system issues a manual confirmation request. The application service module displays the current quality status, twin prediction trends, agent decision logs, reasons for action adjustments, and safety filtering results in real time, facilitating supervision and takeover by process engineers.

[0068] Meanwhile, after the action is executed, the system continues to collect actual quality feedback and compare it with the twin prediction results. When trends such as thinner coating thickness, increased hole position deviation, excessive interlayer offset, or increased defect probability are observed, the process agent can adjust the current density, chemical replenishment amount, laser energy, scanning path, or pressing compensation amount in advance, realizing the transformation from "post-event detection" to "pre-event prediction and proactive compensation," reducing the risk of batch defect diffusion.

[0069] This step involves deploying the trained process agent into the actual production line of new display electronic components, and achieving prediction-based closed-loop control through the synchronous operation of the gray box digital twin model with the real production line.

[0070] In its implementation, the method further includes: step S6, periodically fine-tuning the parameters of the gray-box digital twin model and the process agent based on online operational data to achieve continuous evolution of the model and process strategy. Specifically, this includes: S601, Construction of an online experience replay library. The system records the status in each decision-making cycle. ,action Twin predictions, actual quality results, and rewards Anomaly identification and manual intervention records are used to form online experience data. At the same time, the final test results of the complete batch are correlated with process parameter curves, defect distribution, and process agent action logs to form a "condition characteristics - control strategy - quality result" triplet case.

[0071] S602. Model parameters and policy network are updated periodically. The system triggers updates based on time period, batch size, prediction error, or quality fluctuations. When the twin model prediction error continues to increase, the system uses the latest calibration data to adjust θ and... Incremental updates are performed; when the performance of the process agent deteriorates under certain new products or disturbances, the system makes small-step adjustments to the policy network parameters ψ in the updated twin environment. The updated model and policy must undergo offline security verification before replacing the online version.

[0072] S603, Case Knowledge Accumulation and Rapid New Product Migration. The continuous learning module accumulates historical cases into a case knowledge base. When a new Mini LED, Micro LED, or HDI product model is introduced to the production line, the system searches for similar cases based on the stack-up structure, linewidth and spacing, aperture requirements, material type, and process route. It then calls upon similar strategies as a warm-up solution, which is pre-verified by the digital twin environment. This approach reduces reliance on manual experience, shortens parameter adaptation time after new product changeovers, and improves first-article yield and process stability.

[0073] This step enables the system to continuously adapt to changes such as equipment aging, material batch fluctuations, drug solution decay, and the introduction of new products during long-term operation, avoiding twin model distortion and agent strategy aging.

[0074] This invention also provides a twin-enhanced process agent construction system for intelligent manufacturing of novel display electronic components, such as... Figure 2 As shown, it includes: The data acquisition module is used to acquire and process multi-source heterogeneous data for the production line of new display electronic components to obtain a structured process dataset. A digital twin engine is used to construct a gray-box digital twin model that integrates a multiphysics mechanism model and a data-driven residual compensation network, and to perform online dynamic calibration of the gray-box digital twin model using a structured process dataset. The process agent reasoning core is used to design process agents for rapid changeover manufacturing, defining its perception layer, intelligent reasoning layer, decision-making layer, and multi-objective reward function; The virtual-real synchronization middleware is used to conduct offline training and verification of the process agent in the virtual environment constructed by the gray box digital twin model. Through simulation sample generation and abnormal operating condition enhancement training, the process agent learns process parameter optimization and abnormal operating condition compensation strategies. The virtual-real synchronization middleware realizes data interaction, state synchronization, model update and prediction state push between the digital twin model and physical production line equipment. The application service module is used to deploy the validated process agent to edge nodes, and receive the future predicted status and compensation signals of the gray box digital twin model in real time through virtual-real synchronization, forming a closed-loop control for the actual production line.

[0075] In the specific implementation process, the twin-enhanced process agent construction system also includes a continuous learning module. This module is used to periodically fine-tune the parameters of the gray box digital twin model and the process agent based on online operating data, so as to realize the continuous evolution of the model and process strategy.

[0076] It is understood that the twin enhancement process agent construction system for intelligent manufacturing of novel display electronic components provided in this embodiment of the invention corresponds to the twin enhancement process agent construction method for intelligent manufacturing of novel display electronic components described above. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the twin enhancement process agent construction method for intelligent manufacturing of novel display electronic components, and will not be repeated here.

[0077] This invention also provides a computer-readable storage medium storing a computer program for constructing a twin-enhanced process agent for intelligent manufacturing of novel display electronic components, wherein the computer program causes a computer to execute the twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described above.

[0078] This invention also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for executing the twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described above.

[0079] In summary, compared with existing technologies, it has the following beneficial effects: This invention integrates the sample expansion, virtual trial-and-error iteration, and real-time state prediction capabilities of digital twins throughout the entire process of agent training, deployment, and evolution. This transforms process parameter optimization from experience-driven to mechanism-data fusion-driven, and quality control from post-event detection to pre-event prediction and proactive compensation. As a result, it effectively solves problems such as narrow process windows, slow yield ramp-up, delayed anomaly compensation, and difficulties in improving OEE in the manufacturing of Mini LED, Micro LED, and HDI high-density interconnect electronic components. This provides reliable technical support for the high-flexibility, high-precision, high-yield, and high-efficiency intelligent manufacturing of key electronic components for new displays.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a twin-enhanced process agent for intelligent manufacturing of novel display electronic components, characterized in that, include: Acquire and process multi-source heterogeneous data for new display electronic component production lines to obtain structured process datasets; A gray-box digital twin model integrating a multiphysics mechanism model and a data-driven residual compensation network is constructed, and the gray-box digital twin model is dynamically calibrated online using the structured process dataset. Design a process agent for rapid changeover manufacturing, defining its perception layer, intelligent reasoning layer, decision-making layer, and multi-objective reward function; Offline training and verification of the process agent are carried out in the virtual environment constructed by the gray box digital twin model. Through simulation sample generation and abnormal operating condition enhancement training, the process agent learns process parameter optimization and abnormal operating condition compensation strategies. The validated process agent is deployed to the edge node, and receives the future predicted state and compensation signal of the gray box digital twin model in real time through virtual-real synchronization, forming a closed-loop control for the actual production line.

2. The twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described in claim 1, characterized in that, Acquire and process multi-source heterogeneous data for new display electronic component production lines to obtain structured process datasets, including: Obtain real-time multi-source production line data corresponding to the current work order, and perform time-series alignment and aggregation on the real-time multi-source production line data; Retrieve the layered structure, process route, material batches, historical production records of similar products, and final quality judgment data related to the current work order from the historical database to form a related dataset; We perform missing data filling, anomaly removal, time synchronization, and feature extraction on real-time production line multi-source data and related datasets after time-series alignment and aggregation to construct a structured process dataset.

3. The twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described in claim 1, characterized in that, Constructing a gray-box digital twin model that integrates a multiphysics mechanism model and a data-driven residual compensation network, including: For each key process of the novel display electronic component, a mechanism model adapted to the process characteristics of each process is constructed. The mechanism model covers multi-dimensional physical mechanisms, including reaction kinetics, flow field mass transfer, electric field transport, laser thermal phase transition and thermo-mechanical coupling. A residual compensation network is constructed to compensate for the deviation between the learning mechanism model output and the actual measured values. Combining the learning mechanism model and the residual compensation network, a gray-box digital twin model is obtained. The complete output of the gray-box digital twin model is as follows: in, This represents the quality index of the mechanistic model's predictions. For multiphysics mechanism models, For physical parameters; This represents the estimated value of the prediction residuals for the mechanistic model. For data-driven residual compensation networks, for The hidden state at any given moment. For network parameters, for The input condition vector at time step; For gray box digital twin model The predicted quality value output at any given time.

4. The twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described in claim 1, characterized in that, The intelligent inference layer adopts a hybrid architecture of gray-box digital twin model predictive control and deep reinforcement learning, and outputs process adjustment actions, specifically including: A gray-box digital twin model is used for rolling optimization within a finite prediction time domain to generate baseline actions that conform to physical laws and equipment constraints. A deep reinforcement learning policy network outputs compensation actions based on the current state and the predicted future state to handle model mismatch, unknown disturbances, and multi-objective trade-offs. The expressions for the output process adjustment actions include: in, express The process adjustment actions output in real time; for The baseline action sequence is obtained through rolling optimization based on the gray box digital twin model, providing a nominal control trajectory that conforms to physical laws. For parameters Deep reinforcement learning policy network, enter Moment State Forward prediction with digital twins Implicit state prediction of the step It outputs intelligent compensation quantities for model mismatch, unknown disturbances, and multi-objective trade-offs.

5. The twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described in any one of claims 1 to 4, characterized in that, The multi-objective reward function includes: in, for Momentary rewards; This is a quality compliance indicator function. It takes a value of 1 when all online quality indicators are within the quality specification limits, and 0 otherwise. express The vector of online quality metrics at any given time; for Energy or material consumption costs at any given time; The response delay from when an anomaly is detected to when the agent issues an effective compensation action; This is an indication function for whether a quality abnormality or equipment alarm has occurred. express Check for quality abnormalities or equipment alarms at all times; , , , These are the weighting coefficients for each objective.

6. The twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described in any one of claims 1 to 4, characterized in that, The offline training and validation of the process agent in the virtual environment constructed by the gray-box digital twin model, through simulation sample generation and abnormal operating condition enhancement training, enables the process agent to learn process parameter optimization and abnormal operating condition compensation strategies, including: Multiple parallel simulation environments were constructed using gray-box digital twin models; The process agent is continuously trained from standard conditions to abnormal operating conditions through multiple parallel simulation environments, specifically including: In the initial training phase, only standard operating conditions are used to enable the process agent to learn basic parameter stability control and quality achievement strategies; complex operating conditions are gradually introduced, and the range of disturbances is gradually expanded. Training, convergence, and security verification of the process agent; The policy network parameters in the validated process Agent are engineered, solidified, and managed for reuse.

7. The twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described in any one of claims 1 to 4, characterized in that, The twin-enhanced agent construction method further includes: Based on online operational data, the parameters of the gray box digital twin model and the process agent are periodically fine-tuned.

8. A twin-enhanced process agent construction system for intelligent manufacturing of novel display electronic components, characterized in that, include: The data acquisition module is used to acquire and process multi-source heterogeneous data for the production line of new display electronic components to obtain a structured process dataset. A digital twin engine is used to construct a gray-box digital twin model that integrates a multiphysics mechanism model and a data-driven residual compensation network, and to perform online dynamic calibration of the gray-box digital twin model using a structured process dataset. The process agent reasoning core is used to design process agents for rapid changeover manufacturing, defining its perception layer, intelligent reasoning layer, decision-making layer, and multi-objective reward function; The virtual-real synchronization middleware is used to conduct offline training and verification of the process agent in the virtual environment constructed by the gray box digital twin model. Through simulation sample generation and abnormal operating condition enhancement training, the process agent learns process parameter optimization and abnormal operating condition compensation strategies. The application service module is used to deploy the validated process agent to edge nodes, and receive the future predicted status and compensation signals of the gray box digital twin model in real time through virtual-real synchronization, forming a closed-loop control for the actual production line.

9. A computer-readable storage medium, characterized in that, It stores a computer program for constructing a twin-enhanced process agent for the intelligent manufacturing of novel display electronic components, wherein the computer program causes a computer to execute the twin-enhanced process agent construction method for the intelligent manufacturing of novel display electronic components as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing a twin-enhanced process agent construction method for intelligent manufacturing of novel display electronic components as described in any one of claims 1 to 7.