A Smart Decision-Making System for Material Selection and Stack-up Structure of Multilayer Flexible Circuit Boards
The intelligent decision-making system for material selection and stacked structure of multilayer flexible circuit boards solves the problems of scattered storage of material parameters and reliance on experience in design. It realizes automated decision-making and optimization, improves the efficiency and reliability of flexible circuit board design, and promotes the common progress of design level in the industry.
Patent Information
- Application Number
- CN202511766651.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-11-27
Smart Images

Figure CN121562415B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical technology, specifically a multilayer flexible circuit board material selection and stacked structure intelligent decision-making system. Background Technology
[0002] Multilayer flexible circuit boards are multilayer circuit interconnection structures made of flexible substrates (such as polyimide or polyester film). They are made by laminating conductive copper layers, insulating layers and cover films. Their core feature is that they can be bent, folded or twisted to adapt to complex three-dimensional spatial wiring needs. They are widely used in consumer electronics (such as foldable screen mobile phones and wearable devices), automotive electronics, medical equipment and aerospace to achieve lightweight, high-density circuit connections and dynamic functional integration.
[0003] The core function of multilayer flexible circuit boards is to realize signal transmission, power supply and grounding between different components in electronic systems. They are especially suitable for scenarios that need to work in confined spaces, dynamic bending or harsh environments, such as the folding screen cable of smartphones, the servo control circuit of drones, and the internal interconnection circuit of medical devices.
[0004] Currently, in existing technologies, the material parameters of flexible printed circuit boards are mostly stored in different documents or local databases, lacking a unified full-attribute parameter management system. This makes it impossible to integrate the conventional performance of materials such as substrates and cover films, as well as extended information such as long-term aging reliability and environmental compliance. It also makes it difficult to achieve dynamic docking with material management systems or supplier databases. This fragmented management not only makes it difficult for designers to quickly obtain comprehensive and accurate material data, but the manual update method is also prone to data version lag. When new materials or parameter changes occur, they cannot be synchronized to the design stage in a timely manner, directly affecting the accuracy and timeliness of solution selection. Secondly, design decisions rely heavily on experience, resulting in low efficiency and insufficient optimality in solution generation under multiple constraints. Traditional flexible circuit board design largely depends on engineers' personal experience in material selection and layer configuration. When faced with multiple constraints such as target impedance, total thickness, bending radius, and cost budget, it is difficult to quickly select the globally optimal solution from a vast space of material combinations. While existing design aids can perform simple parameter calculations, they lack intelligent decision-making models trained on massive amounts of data. This prevents automatic analysis and quantitative optimization of constraints, leading to long solution generation cycles, numerous iterations, and a tendency to miss better solutions due to limitations in experience, making it difficult to balance multi-dimensional performance requirements. Third, simulation verification and scheme optimization are disconnected, lacking a closed-loop learning mechanism and making it difficult to guarantee physical feasibility. On the one hand, in existing technologies, simulation and decision-making modules are mostly unidirectional processes. The performance indicators output by simulation are only used as the basis for judging whether a scheme "passes" or "fails," without forming quantitative feedback including deviation analysis, and thus failing to provide precise improvement directions for scheme optimization. On the other hand, the lack of an effective closed-loop learning and optimization mechanism means that it is impossible to extract value from historical design data (especially boundary cases and failure cases) to improve model capabilities, nor can it impose physical rule constraints during the optimization process, which may lead to AI generating absurd schemes that violate the basic principles of materials mechanics and electromagnetism. At the same time, knowledge sharing among users is limited by data privacy protection, making it difficult to achieve the convergence of collective wisdom and rapid improvement of system generalization capabilities. Therefore, to address the above problems, a multilayer flexible circuit board material selection and stacked structure intelligent decision-making system is proposed. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies and solve at least one of the technical problems mentioned in the background art, this invention proposes an intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards.
[0006] The technical solution adopted by the present invention to solve its technical problem is: the intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards, as described in the present invention, includes a material database module for storing and dynamically updating the full property parameters of flexible circuit board materials; The intelligent decision engine module, coupled to the material database module, is used to generate preliminary material selection and laminated structure configuration schemes based on multi-objective constraints input by the user. A high-fidelity simulation module is coupled to the intelligent decision engine module to perform multi-physics coupling simulation on the preliminary configuration scheme and output simulation performance indicators. The scheme evaluation and feedback module, coupled to the high-fidelity simulation module and the intelligent decision engine module, is used to compare the simulation performance indicators with the multi-objective constraints and generate an evaluation report including deviation analysis. The closed-loop learning and optimization module is connected to the scheme evaluation and feedback module and the intelligent decision engine module, respectively. The closed-loop learning and optimization module includes: The data filtering and value assessment tool is used to assess the value of decision data based on its innovativeness, boundary, and failure significance when offline learning is triggered, and to select high-quality training samples to be stored in the incremental learning sample library. The model update strategy library and strategy selector, wherein the strategy selector dynamically selects one or more strategies from global model fine-tuning, feature attention enhancement, or adversarial sample augmentation training based on the bias type and value evaluation results. Knowledge solidification logs are used to record key samples, strategies, and effects of each model update; The meta-learning controller is used to analyze the similarity between new tasks and historical tasks, and accelerates the optimization process by calling knowledge from sub-models or combined models. The physical rule constraint, integrated into the model update strategy library, is used to apply prior physical rule constraints of mechanics of materials, electromagnetics and thermodynamics during model training. The federated learning coordinator is used to coordinate the federated learning of model parameters across multiple user terminals, enabling cross-user knowledge sharing and privacy protection. The closed-loop learning and optimization module is configured to receive the evaluation report and trigger the following operating mode based on the deviation analysis results: a. Online optimization mode: When the deviation does not exceed the preset threshold, a reinforcement learning algorithm based on the reward function is used to drive the intelligent decision engine to fine-tune and iterate the parameters of the initial scheme; during this process, the meta-learning controller intervenes first to accelerate convergence; b. Offline learning mode: When any of the following conditions are met: deviation exceeding the limit, iteration number exceeding the limit, performance degradation warning, or new material integration, the machine learning model in the intelligent decision engine module is incrementally trained using the incremental learning sample library and the selected update strategy, under the supervision of the physical rule constraint; the federated learning coordinator periodically aggregates model updates from each user terminal, generates and distributes a global enhancement model.
[0007] Preferably, the material database module stores full-property parameters including mechanical performance parameters, electrical performance parameters, thermal performance parameters, chemical stability parameters, and cost parameters; the material database module is dynamically updated by collecting market data, user feedback data, and experimental data in real time.
[0008] Preferably, the intelligent decision engine module uses a reinforcement learning algorithm based on a reward function to generate a preliminary configuration scheme; the reward function is defined as: ; in, As a reward value, For the target quantity, For the first The weight of each objective, For the first One simulation performance index, For the first One target constraint value.
[0009] Preferably, the multiphysics coupling simulation performed by the high-fidelity simulation module includes electromagnetic simulation, thermal simulation, and mechanical stress simulation; the simulation performance indicators include signal integrity indicators, power integrity indicators, thermal distribution indicators, and bending life indicators.
[0010] Preferably, the deviation analysis of the scheme evaluation and feedback module is achieved by calculating the Euclidean distance between the simulation performance index and the multi-objective constraints; the deviation value The calculation is as follows: ; in, For the first One simulation performance index, For the first One target constraint value, This indicates the total number of dimensions in the performance evaluation.
[0011] Preferably, the data filtering and value assessment device of the closed-loop learning and optimization module assesses the value of decision data based on innovativeness, boundary conditions, and failure significance; innovativeness scoring... for: ; in, For current decision-making data, For historical decision-making data, For similarity function, This represents the amount of historical data.
[0012] Preferably, the model update strategy library and strategy selector dynamically select global model fine-tuning, feature attention enhancement, or adversarial sample augmentation training strategies based on the bias type and value evaluation results; wherein, the bias type includes performance bias, constraint bias, and model uncertainty bias.
[0013] Preferably, the meta-learning controller analyzes the similarity between new and historical tasks using task embedding vectors; task similarity is calculated as follows: ; in, and These are the embedding vectors for the new task and the historical task, respectively.
[0014] Meta-learning controller analyzes the similarity between new and historical tasks. Preferably, the physical rule constraint applies prior physical rule constraints based on mechanics of materials, electromagnetics, and thermodynamics during model training, including stress-strain relationships based on Hooke's law, electromagnetic field constraints based on Maxwell's equations, and temperature distribution constraints based on Fourier's law of heat conduction.
[0015] Preferably, the federated learning coordinator uses a federated averaging algorithm to aggregate model parameters from each user terminal, achieving cross-user knowledge sharing and privacy protection; the aggregation formula is: ; in, These are global model parameters. For the first Model parameters for each user terminal For the first Data volume per user terminal Total data volume For user-side data.
[0016] It is understood that, This invention provides an intelligent decision-making system for material selection and stack-up structure of multilayer flexible circuit boards, realizing design automation and knowledge assetization: transforming the non-standardized design process that relies on personal experience into a reusable automated process driven by data and models, and solidifying the company's core design knowledge into sustainably value-added digital assets; significantly improving R&D efficiency and reducing costs: the online rapid iteration closed loop of "decision-simulation-evaluation" shortens the traditional design-prototyping-testing cycle, which takes several weeks, to hours or even minutes, and significantly reduces the number of physical prototypings through high-precision virtual verification, directly reducing R&D costs and time; and ensuring optimal solution performance. Engineering reliability: The intelligent decision engine can find the global optimal solution in a vast space of material combinations; at the same time, the introduction of physical rule constraints ensures that all AI-recommended solutions conform to basic scientific principles, combining innovation and engineering feasibility; endowing the system with continuous evolution and swarm intelligence capabilities: The closed-loop learning module enables the system to learn from every success and failure, and its decision-making ability continuously improves over time; the introduction of the federated learning mechanism enables the system to gather collective wisdom under the premise of strictly protecting the data privacy of all parties, forming a "super brain" that surpasses the experience of any single team, and promoting the common progress of the entire industry's design level.
[0017] This invention provides an intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards, which ensures the comprehensiveness and timeliness of material data, enabling the system to select materials based on the latest information, improving the accuracy of decision-making and adaptability to market changes, while reducing the selection risk caused by outdated data. Attached Figure Description
[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention.
[0019] In the attached diagram: Figure 1 This is a schematic diagram of the system flow of the present invention; Figure 2 This is a schematic diagram of the material data module flow in this invention; Figure 3 This is a flowchart illustrating the high-fidelity simulation module in this invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.
[0021] Specific implementation examples are given below.
[0022] Please see Figures 1-3 This invention provides an intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards, including a material database module for storing and dynamically updating the full property parameters of flexible circuit board materials; The intelligent decision engine module, coupled to the material database module, is used to generate preliminary material selection and laminated structure configuration schemes based on multi-objective constraints input by the user. A high-fidelity simulation module is coupled to the intelligent decision engine module to perform multi-physics coupling simulation on the preliminary configuration scheme and output simulation performance indicators. The scheme evaluation and feedback module, coupled to the high-fidelity simulation module and the intelligent decision engine module, is used to compare the simulation performance indicators with the multi-objective constraints and generate an evaluation report including deviation analysis. The closed-loop learning and optimization module is connected to the scheme evaluation and feedback module and the intelligent decision engine module, respectively. The closed-loop learning and optimization module includes: The data filtering and value assessment tool is used to assess the value of decision data based on its innovativeness, boundary, and failure significance when offline learning is triggered, and to select high-quality training samples to be stored in the incremental learning sample library. The model update strategy library and strategy selector, wherein the strategy selector dynamically selects one or more strategies from global model fine-tuning, feature attention enhancement, or adversarial sample augmentation training based on the bias type and value evaluation results. Knowledge solidification logs are used to record key samples, strategies, and effects of each model update; The meta-learning controller is used to analyze the similarity between new tasks and historical tasks, and accelerates the optimization process by calling knowledge from sub-models or combined models. The physical rule constraint, integrated into the model update strategy library, is used to apply prior physical rule constraints of mechanics of materials, electromagnetics and thermodynamics during model training. The federated learning coordinator is used to coordinate the federated learning of model parameters across multiple user terminals, enabling cross-user knowledge sharing and privacy protection. The closed-loop learning and optimization module is configured to receive the evaluation report and trigger the following operating mode based on the deviation analysis results: a. Online optimization mode: When the deviation does not exceed the preset threshold, a reinforcement learning algorithm based on the reward function is used to drive the intelligent decision engine to fine-tune and iterate the parameters of the initial scheme; during this process, the meta-learning controller intervenes first to accelerate convergence; b. Offline learning mode: When any of the following conditions are met: deviation exceeding the limit, iteration number exceeding the limit, performance degradation warning, or new material integration, the machine learning model in the intelligent decision engine module is incrementally trained using the incremental learning sample library and the selected update strategy, under the supervision of the physical rule constraint; the federated learning coordinator periodically aggregates model updates from each user terminal, generates and distributes a global enhancement model.
[0023] The specific implementation method of the materials database module is as follows: This module serves as the system's data foundation, storing comprehensive attribute parameters that include not only the conventional electrical, thermal, and mechanical properties of materials such as substrates, cover films, adhesives, and copper foils, but also data such as long-term aging reliability, environmental compliance information, and supplier production capacity. Its dynamic update mechanism is manifested in two ways: firstly, system administrators can manually input experimentally verified new material data; secondly, the system can connect to the material management system or supplier database via API to achieve automatic synchronization and version control of key parameters, ensuring that the decision-making basis remains consistent with the latest material system. The specific implementation method of the intelligent decision engine module is as follows: This module is the core decision-making center of the system. User-input multi-objective constraints may include, but are not limited to: target impedance value and its tolerance, maximum allowable total thickness, minimum bending radius, maximum operating temperature, cost budget, and signal loss requirements at specific frequencies. The engine integrates machine learning models (such as deep neural networks, random forests, or reinforcement learning agents) trained on massive amounts of data. Its workflow is as follows: First, it parses and quantifies user constraints; second, it performs rapid filtering, matching, and multi-objective optimization calculations based on model inference across the entire material database; finally, it outputs one or more preliminary configuration schemes that probabilistically best satisfy all constraints. These schemes specifically include the type of selected material, the thickness of each layer, and the stacking order. The specific implementation method of the high-fidelity simulation module is as follows: This module serves as a virtual verification laboratory. Upon receiving a preliminary design, it automatically constructs a corresponding three-dimensional finite element analysis model. Its multiphysics coupling simulation refers to: Electromagnetic simulation: Calculate the characteristic impedance, insertion loss, return loss, and crosstalk of the scheme at high frequencies; Thermal simulation: Analyze the steady-state and transient temperature field distribution of the circuit board under a given power consumption, and identify hot spot areas; Structural mechanics simulation: Simulates the stress and strain distribution of circuit boards under bending, torsion and other conditions to predict their fatigue life; Coupled simulation can reveal potential problems that cannot be found by single-physics field analysis, such as how deformation caused by heating affects impedance consistency. The specific implementation method of the scheme evaluation and feedback module is as follows: This module serves as a bridge between simulation and optimization. It compares the simulation performance metrics output by the simulation module with the user-input multi-objective constraints item by item and quantitatively. Its evaluation report not only includes simple "pass / fail" judgments, but more importantly, it performs deviation analysis, such as: "At 10GHz frequency, the impedance value deviates from the target value by -5Ω (exceeding the ±3Ω tolerance)" and "At maximum power consumption, the highest temperature exceeds the limit by 8°C," etc. This report provides precise and quantitative improvement directions for subsequent optimization. The specific implementation method of the closed-loop learning and optimization module is as follows: This module is key to the system's "intelligence" and "evolution" capabilities, and its internal components work together as follows: Data filtering and value evaluator: When offline learning is triggered, it does not simply use all historical data, but filters out the most valuable data based on three dimensions: innovativeness (whether the solution is novel), boundary (whether the solution is on the edge of performance meeting or not meeting the standards) and failure significance (the degree and reasons for the failure of the solution), and stores it in the incremental learning sample library, which greatly improves the quality and efficiency of training data. Model update strategy library and strategy selector: This component dynamically selects the most suitable model update strategy from the strategy library based on the type of bias (such as systematic bias, random bias) and value assessment results in the evaluation report. For example, for systematic bias caused by ignoring a key material parameter, a feature attention enhancement strategy is used; for models with unstable performance under certain extreme conditions, adversarial sample augmentation training is used to improve their robustness. Physical rule constraint: This component is crucial. It integrates the fundamental laws and empirical formulas of mechanics of materials, electromagnetism and thermodynamics as inviolable hard constraints into the model training process. For example, when optimizing heat conduction paths, the model must obey Fourier's law, thereby ensuring that all optimization schemes are physically feasible and avoiding absurd results that "defy common sense" from the AI. Meta-learning controller: Its role is similar to that of an experienced engineer. When encountering a new design task, it analyzes the similarity between the task and historical tasks, and quickly calls on existing sub-model knowledge or model combinations for initialization. This allows reinforcement learning in online optimization mode to start from a higher starting point, accelerates convergence, and avoids the long trial and error of "starting from scratch" every time. Federated Learning Coordinator: This mechanism allows multiple user terminals (such as different departments or different partner companies) to train the model locally using their own private data, and then upload the encrypted model parameters to the central server for aggregation, generating a more powerful global augmentation model and redistributing it to each terminal; this achieves cross-user knowledge sharing and privacy protection, enabling the system to learn from a wider range of application scenarios and quickly improve its generalization ability; This system achieves design automation and knowledge assetization through the precise coupling and closed-loop connection of the above five modules: transforming the non-standardized design process that relies on personal experience into a reusable automated process driven by data and models, and solidifying the company's core design knowledge into digital assets that can be sustainably valued. Significantly improves R&D efficiency and reduces costs: The online rapid iteration closed loop of decision-making, simulation, and evaluation shortens the traditional design-prototyping-testing cycle, which takes several weeks, to hours or even minutes. High-precision virtual verification significantly reduces the number of physical prototyping attempts, directly reducing R&D costs and time. Ensures optimal solution and engineering reliability: The intelligent decision engine can find the globally optimal solution in the vast space of material combinations. At the same time, the introduction of physical rule constraints ensures that all AI-recommended solutions conform to basic scientific principles and are both innovative and engineering feasible. Endowing the system with continuous evolution and collective intelligence capabilities: The closed-loop learning module enables the system to learn from every success and failure, and its decision-making ability continuously improves over time; The introduction of the federated learning mechanism enables the system to gather collective wisdom and form a "super brain" that surpasses the experience of any single team, while strictly protecting the data privacy of all parties, thus promoting the common progress of the entire industry's design level.
[0024] Furthermore, such as Figures 1-3 As shown, the material database module stores all property parameters, including mechanical property parameters, electrical property parameters, thermal property parameters, chemical stability parameters, and cost parameters; the material database module is dynamically updated by collecting market data, user feedback data, and experimental data in real time.
[0025] The materials database module stores all attribute parameters and dynamically updates them: The material data module stores all property parameters covering mechanical properties, electrical properties, thermal properties, chemical stability, and cost parameters, and achieves dynamic updates and maintenance of data by collecting market data, user feedback data, and experimental data in real time.
[0026] Understandably, this ensures the comprehensiveness and timeliness of material data, enabling the system to select materials based on the latest information, improving the accuracy of decision-making and adaptability to market changes, while reducing the selection risk caused by outdated data.
[0027] Furthermore, such as Figures 1-3 As shown, the intelligent decision engine module uses a reinforcement learning algorithm based on a reward function to generate a preliminary configuration scheme; the reward function is defined as: ; in, As a reward value, For the target quantity, For the first The weight of each objective, For the first One simulation performance index, For the first One target constraint value.
[0028] The intelligent decision engine module employs a reinforcement learning algorithm based on a reward function: Specifically, the intelligent decision engine module uses reinforcement learning algorithms to balance the weights of multiple objectives through a reward function, and generates a preliminary configuration scheme based on the closeness of the simulation performance indicators to the target values, thereby optimizing the decision-making process.
[0029] Understandably, this enables the system to automatically generate configuration schemes that meet multiple objectives, improve decision-making efficiency and optimization capabilities, reduce manual intervention, and ensure that the schemes are more in line with actual application requirements.
[0030] Furthermore, such as Figures 1-3 As shown, the multiphysics coupling simulation performed by the high-fidelity simulation module includes electromagnetic simulation, thermal simulation, and mechanical stress simulation; the simulation performance indicators include signal integrity indicators, power integrity indicators, thermal distribution indicators, and bending life indicators.
[0031] Multiphysics coupling simulation and simulation performance indicators performed by the high-fidelity simulation module: Specifically, the high-fidelity simulation module performs multi-physics coupled simulations such as electromagnetic simulation, thermal simulation, and mechanical stress simulation, and outputs performance indicators such as signal integrity, power integrity, thermal distribution, and bending life to comprehensively evaluate the performance of the solution.
[0032] Understandably, this provides a comprehensive performance assessment, helps identify potential problems, improves the reliability and durability of the solution, and shortens the design verification cycle while reducing actual testing costs.
[0033] Furthermore, such as Figures 1-3 As shown, the deviation analysis of the scheme evaluation and feedback module is achieved by calculating the Euclidean distance between the simulation performance index and the multi-objective constraints; the deviation value The calculation is as follows: ; in, For the first One simulation performance index, For the first One target constraint value, This indicates the total number of dimensions in the performance evaluation.
[0034] Deviation analysis of the scheme evaluation and feedback module: The scheme evaluation and feedback module performs deviation analysis by calculating the Euclidean distance between the simulation performance index and the multi-objective constraints, thus quantifying the gap between the scheme and the objective.
[0035] Understandably, this provides an intuitive deviation assessment, facilitating the rapid identification and correction of shortcomings in the proposed solution, optimizing the decision-making process, and ensuring that the final solution more closely aligns with the design objectives. Furthermore, such as Figures 1-3 As shown, the data filtering and value assessment device of the closed-loop learning and optimization module evaluates the value of decision data based on innovativeness, boundary conditions, and failure significance; innovativeness score for: ; in, For current decision-making data, For historical decision-making data, For similarity function, This represents the amount of historical data.
[0036] Data filtering and value assessment for the closed-loop learning and optimization module: The data screening and value assessment tool in the closed-loop learning and optimization module evaluates the value of decision data based on its innovativeness, boundary conditions, and failure significance. It scores data by comparing the similarity between current data and historical data, and selects high-value data for model optimization.
[0037] Understandably, this improves the model's learning efficiency and adaptability, ensuring that the system can learn from novel and critical data, avoid overfitting, and enhance the accuracy and robustness of decision-making.
[0038] Furthermore, such as Figures 1-3 As shown, the model update strategy library and strategy selector dynamically select global model fine-tuning, feature attention enhancement, or adversarial sample augmentation training strategies based on the bias type and value evaluation results; wherein, the bias type includes performance bias, constraint bias, and model uncertainty bias.
[0039] Model update strategy library and strategy selector: The model update strategy library and strategy selector dynamically select global model fine-tuning, feature attention enhancement, or adversarial sample augmentation training strategies to adjust the model based on the type of bias (such as performance bias, constraint bias, and model uncertainty bias) and the value evaluation results.
[0040] Understandably, this enables dynamic optimization and adaptive adjustment of the model, improves the model's ability to handle complex scenarios, reduces the impact of biases, and enhances the system's stability and generalization performance.
[0041] Furthermore, such as Figures 1-3 As shown, the meta-learning controller analyzes the similarity between new and historical tasks using task embedding vectors; task similarity is calculated as follows: ; in, and These are the embedding vectors for the new task and the historical task, respectively.
[0042] The meta-learning controller analyzes the similarity between new and historical tasks: The meta-learning controller uses task embedding vectors to analyze the similarity between new and historical tasks, and evaluates task relevance by calculating the cosine similarity between vectors, thereby quickly adapting to new tasks.
[0043] Understandably, this accelerates the decision-making process for new tasks, improves the system's transfer learning capabilities, reduces repetitive training time, and ensures the consistency and efficiency of the solution across different tasks.
[0044] Furthermore, such as Figures 1-3 As shown, the physical rule constraint applies prior physical rule constraints based on mechanics of materials, electromagnetics, and thermodynamics during model training, including stress-strain relationships based on Hooke's law, electromagnetic field constraints based on Maxwell's equations, and temperature distribution constraints based on Fourier's law of heat conduction.
[0045] The physical rule constraint applies prior physical rule constraints: The physical rule constraint applies prior physical rules of mechanics of materials, electromagnetics, and thermodynamics during model training, such as stress-strain relationship based on Hooke's law, electromagnetic field constraint based on Maxwell's equations, and temperature distribution constraint based on Fourier's law of heat conduction, to ensure that the scheme conforms to physical laws.
[0046] Understandably, this enhances the scientific validity and feasibility of the solution, prevents decisions that violate the laws of physics, improves the reliability and practical application value of the system, and reduces simulation errors.
[0047] Furthermore, such as Figures 1-3 As shown, the federated learning coordinator uses a federated averaging algorithm to aggregate model parameters from each user terminal, achieving cross-user knowledge sharing and privacy protection; the aggregation formula is: ; in, These are global model parameters. For the first Model parameters for each user terminal For the first Data volume per user terminal Total data volume This refers to the number of users.
[0048] The federated learning coordinator uses a federated averaging algorithm: The federated learning coordinator uses a federated averaging algorithm to aggregate model parameters from each user terminal, and performs a weighted average based on the amount of data from each user terminal to achieve cross-user knowledge sharing while protecting user privacy.
[0049] Understandably, this promotes collaborative learning and knowledge accumulation, improves the model's generalization ability and accuracy, while ensuring data privacy and security, making it suitable for distributed decision-making in multi-user environments.
[0050] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.
Claims
1. A multilayer flexible circuit board material selection and stacking structure intelligent decision-making system, characterized in that: Includes a materials database module for storing and dynamically updating all property parameters of flexible circuit board materials; The intelligent decision engine module, coupled to the material database module, is used to generate preliminary material selection and laminated structure configuration schemes based on multi-objective constraints input by the user. The high-fidelity simulation module is coupled to the intelligent decision engine module and is used to perform multi-physics coupling simulation on the preliminary configuration scheme and output simulation performance indicators. The scheme evaluation and feedback module, coupled to the high-fidelity simulation module and the intelligent decision engine module, is used to compare the simulation performance indicators with the multi-objective constraints and generate an evaluation report including deviation analysis. The closed-loop learning and optimization module is connected to the scheme evaluation and feedback module and the intelligent decision engine module, respectively. The closed-loop learning and optimization module includes: The data filtering and value assessment tool is used to assess the value of decision data based on its innovativeness, boundary, and failure significance when offline learning is triggered, and to select high-quality training samples to be stored in the incremental learning sample library. The model update strategy library and strategy selector, wherein the strategy selector dynamically selects one or more strategies from global model fine-tuning, feature attention enhancement, or adversarial sample augmentation training based on the bias type and value evaluation results. Knowledge solidification logs are used to record key samples, strategies, and effects of each model update; The meta-learning controller is used to analyze the similarity between new tasks and historical tasks, and accelerates the optimization process by calling knowledge from sub-models or combined models. The physical rule constraint, integrated into the model update strategy library, is used to apply prior physical rule constraints of mechanics of materials, electromagnetics and thermodynamics during model training. The federated learning coordinator is used to coordinate the federated learning of model parameters across multiple user terminals, enabling cross-user knowledge sharing and privacy protection. The closed-loop learning and optimization module is configured to receive the evaluation report and trigger the following operating mode based on the deviation analysis results: a. Online optimization mode: When the deviation does not exceed the preset threshold, a reinforcement learning algorithm based on the reward function is used to drive the intelligent decision engine to fine-tune and iterate the parameters of the initial scheme; during this process, the meta-learning controller intervenes first to accelerate convergence; b. Offline learning mode: When any of the following conditions are met: deviation exceeding the limit, iteration number exceeding the limit, performance degradation warning, or new material integration, the machine learning model in the intelligent decision engine module is incrementally trained using the incremental learning sample library and the selected update strategy, under the supervision of the physical rule constraint; the federated learning coordinator periodically aggregates model updates from each user terminal, generates and distributes a global enhancement model.
2. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The material database module stores all property parameters, including mechanical performance parameters, electrical performance parameters, thermal performance parameters, chemical stability parameters, and cost parameters; the material database module is dynamically updated by collecting market data, user feedback data, and experimental data in real time.
3. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The intelligent decision engine module uses a reinforcement learning algorithm based on a reward function to generate a preliminary configuration scheme; the reward function is defined as: ; in, As a reward value, For the target quantity, For the first The weight of each objective, For the first One simulation performance index, For the first One target constraint value.
4. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The high-fidelity simulation module performs multiphysics coupling simulations, including electromagnetic simulation, thermal simulation, and mechanical stress simulation; the simulation performance indicators include signal integrity indicators, power integrity indicators, thermal distribution indicators, and bending life indicators.
5. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The deviation analysis of the scheme evaluation and feedback module is achieved by calculating the Euclidean distance between the simulation performance index and the multi-objective constraints; the deviation value The calculation is as follows: ; in, For the first One simulation performance index, For the first One target constraint value, This indicates the total number of dimensions in the performance evaluation.
6. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The data filtering and value assessment device of the closed-loop learning and optimization module evaluates the value of decision data based on innovativeness, boundary conditions, and failure significance; innovativeness score. for: ; in, For current decision-making data, For historical decision-making data, For similarity function, This represents the amount of historical data.
7. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The model update strategy library and strategy selector dynamically select global model fine-tuning, feature attention enhancement, or adversarial sample augmentation training strategies based on the bias type and value evaluation results; the bias types include performance bias, constraint bias, and model uncertainty bias.
8. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The meta-learning controller analyzes the similarity between new and historical tasks using task embedding vectors; task similarity is calculated as follows: ; in, and These are the embedding vectors for the new task and the historical task, respectively.
9. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The physical rule constraint applies prior physical rule constraints based on mechanics of materials, electromagnetics, and thermodynamics during model training, including stress-strain relationships based on Hooke's law, electromagnetic field constraints based on Maxwell's equations, and temperature distribution constraints based on Fourier's law of heat conduction.
10. The intelligent decision-making system for material selection and stacking structure of multilayer flexible circuit boards as described in claim 1, characterized in that: The federated learning coordinator uses a federated averaging algorithm to aggregate model parameters from each user terminal, achieving cross-user knowledge sharing and privacy protection; the aggregation formula is: ; in, These are global model parameters. For the first Model parameters for each user terminal For the first Data volume per user terminal Total data volume This refers to the number of users.
Citation Information
Patent Citations
Simulation system intelligent decision-making method and system based on knowledge graph and federated learning
CN120087794A
Laminating method and system of multi-layer flexible circuit board
CN120786822A
