Circuit EMI suppression-oriented multi-knowledge collaborative distillation and optimization method and system

Abstract

The invention discloses a circuit EMI suppression-oriented multi-knowledge collaborative distillation and optimization method and system, and the method comprises the steps: constructing a basic large language model based on the circuit design of the electronic industry as a student model, and carrying out the adaptive optimization of the circuit design field. Thirdly, integrating the formalized EMI rule base, the simulation data / model base and the expert experience case base, constructing a multi-source heterogeneous teacher knowledge system, and generating a comprehensive guidance signal G; then, candidate circuit design generated by a student model is evaluated through the system, and a multi-objective loss function including basic distillation, rule conformity, performance fitting and expert experience consistency is constructed; according to the method, the universality is reserved by quantifying differences, compliance is ensured through punishment, the performance difference is reduced, dynamic weight adjustment is introduced, and the model performance reaches a preset target through iterative distillation. The model can understand and process terminologies and design rules in the field of circuit design, and the defect that a general AI model lacks domain knowledge is overcome.

Description

Technical Field

[0001] The present invention belongs to the field of large language model application and knowledge distillation technology, and more specifically, relates to a multi-knowledge collaborative distillation and optimization method and system for circuit EMI suppression. Background Art

[0002] As electronic devices advance towards higher frequencies and smaller sizes, electromagnetic interference (EMI) has become a critical factor affecting product performance and compliance. Traditional EMI design relies primarily on manual experience, design rules, and simulation testing, but these methods have significant limitations. First, professional knowledge transfer is difficult. Experienced engineers' design expertise, such as loop optimization and grounding topology design in PCB layout, often exists in the form of tacit knowledge and lacks systematic organization, making it difficult for new designers to quickly grasp key key points. Second, rule application is inefficient. International standards such as IEC, CISPR, and industry specifications contain numerous EMI design rules, covering routing, filtering, shielding, and other requirements. Manual application can easily lead to design omissions due to misunderstandings or negligence. Third, simulation and testing are costly. Electromagnetic simulation consumes significant computing resources and time, with a single simulation taking hours or even days. Physical testing requires prototyping, which involves hardware costs and repeated iterations, extending development cycles. Fourth, existing tools are limited in capabilities. Traditional CAD tools only provide basic EMI analysis capabilities and lack proactive optimization capabilities. While specialized simulation software such as HFSS and CST offer high accuracy, their high barriers to entry and procurement costs hinder widespread adoption.

[0003] In recent years, large language models have demonstrated great potential in text generation and logical reasoning, but their application in EMI design still faces technical bottlenecks. First, general-purpose LLMs lack domain-specific knowledge. Models struggle to understand specialized constraints in circuit design, such as the impact of power device parasitic parameters on EMI, and the resulting solutions fail to meet actual engineering requirements. Second, knowledge fusion methods are limited. Existing knowledge distillation techniques often rely on a single data source, such as simulation data or rule bases. This makes it difficult to simultaneously integrate heterogeneous knowledge from multiple sources, including rules, simulation data, and expert experience, and thus fails to fully leverage the complementary advantages of these diverse knowledge sources. Finally, model deployment is limited. Directly applying large, general-purpose LLMs consumes significant computing resources, making it difficult to meet the requirements of industrial-grade real-time design and low-cost deployment.

[0004] Currently, the field faces a pressing technical challenge: efficiently integrating heterogeneous EMI knowledge from multiple sources, including industry regulations, simulation data, and expert experience, into a large language model to proactively consider EMI mitigation early in circuit design and achieve lightweight deployment. Existing technologies have yet to offer an effective solution, necessitating an innovative approach to improve EMI design efficiency and accuracy while reducing development costs. Summary of the Invention

[0005] This invention aims to solve the problem of insufficient professional knowledge in circuit EMI design using existing large language models. It integrates multi-source heterogeneous knowledge such as formal EMI rules, simulation experience data, and domain expert knowledge into LLM efficiently and at low cost, so that EMI suppression can be actively considered in the early stages of circuit design, thereby improving design efficiency and product EMI performance, reducing later rectification costs, and achieving efficient circuit EMI suppression and optimization.

[0006] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a multi-knowledge collaborative distillation and optimization method for circuit EMI suppression, comprising:

[0007] S1. Build a basic large language model based on circuit design and development in the electronics industry as the student model to be optimized. Then, perform adaptive optimization on the student model to be optimized in the circuit design field.

[0008] S2. Build a multi-source heterogeneous teacher knowledge system by integrating knowledge sources including a formal EMI rule library, an EMI simulation data / model library, and an EMI expert experience and case library. The teacher knowledge system can perform multi-dimensional evaluations of candidate circuit designs, and the evaluation results can be summarized into a comprehensive guidance signal G.

[0009] S3. Perform a multi-dimensional evaluation on each candidate circuit design generated by the student LLM through each knowledge source in the teacher's knowledge system, and construct a basic distilled LLM based on the evaluation results of each dimension. distill , rule compliance L rule , performance fitting L perf and expert experience consistency L exp The multi-objective loss function includes: retaining general design capabilities by quantifying the difference in output distribution between the student model and the large general language model; imposing penalties on solutions that violate EMI design rules to ensure that the generated design meets industry specifications; minimizing the gap between the EMI performance of the design solution in simulation and the target performance to improve the feasibility of actual engineering; and guiding the model to learn from the expert success case model to avoid known defect patterns; at the same time, introducing a dynamic weight adjustment mechanism to adapt to the needs of different training stages; and adopting an iterative distillation process until the model performance reaches the predetermined target.

[0010] Furthermore, the specific method for the adaptation optimization in the circuit design field in S1 is:

[0011] Collect and organize multi-source data in the field of electronic circuit design, including circuit design specification documents, open source circuit design projects, device datasheets, and design case analysis reports, to build a domain-specific corpus;

[0012] Using an incremental pre-training method, the corpus is used to further train the basic model, enabling the model to deeply learn the professional terminology, design rules, and expression habits in the field of circuit design;

[0013] Through the prompt engineering method, specific prompt templates are designed for circuit design tasks to guide the model to generate design solutions that meet industry standards and actual needs;

[0014] At the same time, in order to collect and organize file samples including Verilog code, KiCad project files, and Jialichuang EDA layout, the model is trained in a targeted manner using code generation and parsing tasks to improve the model's ability to process circuit design file formats, enabling the model to accurately understand and generate design file fragments that comply with the specifications of these tools.

[0015] Furthermore, the specific method for multi-dimensional evaluation of the multi-source heterogeneous teacher knowledge system in S2 includes:

[0016] The rule compliance evaluation of design solution D is performed to obtain the rule compliance score R(D):

[0017] Extract EMI design rules from documents including international standards, industry specifications and corporate design guidelines, and each rule i Represented as a five-tuple:

[0018] r i =(T i ,P i ,C i ,f i ,w i )

[0019] Where, T i is the rule type, P i is a logical predicate, C i is the parameter constraint, f i is the evaluation function, w i is the weight coefficient;

[0020] The entire rule base R can be expressed as a finite set of rules: R = {r1, r2, ..., r n};

[0021] For design solution D, its rule compliance score R(D) is:

[0022]

[0023] Where n is the total number of rules, f i (D) is the solution D for rule r i The evaluation function value of ;

[0024] Design Scheme Xnew The simulation performance evaluation is performed to obtain the simulation performance score S:

[0025] Collect design cases including EMI simulation results of historical projects and other proxy models for training fast EMI prediction for typical circuit modules or topologies, and extract design feature vectors and performance indicator vector

[0026] Train a fast prediction model using machine learning methods:

[0027]

[0028] Minimize the mean squared error (MSE) loss function:

[0029]

[0030] Among them, M is the number of training samples, (X j ,Y j ) is the feature and performance vector of the jth sample;

[0031] For new design X new , its simulation performance score S can be quantified based on the prediction error:

[0032]

[0033] Among them, Y ref is an ideal performance index, and the closer the score is to 1, the better the design performance;

[0034] Evaluate the matching degree between the design solution and the expert experience and obtain the experience matching score E:

[0035] Using natural language processing methods, we extract design patterns and failure patterns from document information, including design notes, technical reports, problem diagnosis records, and analyses of successful and failed EMI design cases by senior EMI engineers in the electronics industry, and build a pattern database.

[0036] Each mode is represented by a feature vector express;

[0037] The matching score E between the design solution and the expert experience is calculated using cosine similarity:

[0038]

[0039] Among them, Z design is the design scheme characteristic vector, Z expert is the expert mode feature vector; the evaluation scores R(D), S, and E of the above three types of knowledge sources are integrated into a comprehensive guidance score G in a weighted manner:

[0040] G=w r ·R+w s ·S+w e ·E

[0041] Among them, w r 、w s 、w e are the weights of rules, simulation, and experience knowledge sources, respectively, and w r +w s +w e =1.

[0042] Furthermore, the specific method of retaining the general design capability by quantifying the difference in output distribution between the student model and the large general language model in S3 is as follows:

[0043] The KL divergence is used to measure the distribution difference between the student model and the large general language model. The specific formula is as follows:

[0044]

[0045] Further:

[0046]

[0047] Where D is the training data set, which contains the circuit design requirement description; x is the input text; y is the output token; P T (y|x) and P S (y|x) is the probability distribution of the teacher model and the student model output y under input x; Y is the set of all possible output tokens; |D| is the number of samples in the dataset.

[0048] Furthermore, the specific method of imposing penalties on solutions that violate EMI design rules in S3 to ensure that the generated design meets industry standards is as follows:

[0049] Evaluate function f based on the rules i , using the negative log-likelihood form:

[0050]

[0051] Where N is the total number of rules; w r,i is the weight of the i-th rule; D S (x) is the design solution generated by the student model based on the input x; f i (D S (x)) is the evaluation score of the design scheme for the i-th rule, ranging from 0 to 1; k i ≥1 is a tuning factor that allows for adjustments to penalty curves for different rules (e.g. k i=2 becomes a square term); introduce severity i ∈(0,1) represents the importance of rule i or the severity of its violation;

[0052] For the key rule, if f i (D S (x))=0, then L rule Set to a maximum value to force the model to avoid violations.

[0053] Furthermore, the specific method of minimizing the gap between the EMI performance of the design solution in simulation and the target performance in S3 and improving the feasibility of actual engineering is as follows:

[0054] Based on the simulation agent model g(X), the mean square error is used:

[0055]

[0056] Where y target is the target EMI performance index vector; g(D S (x)) is the performance prediction value of the simulation agent model for the design solution DS(x);

[0057] When using the reinforcement learning framework, the loss can be converted into a reward:

[0058]

[0059] at this time:

[0060]

[0061] Where λ is the reward scaling factor.

[0062] Furthermore, the specific method of guiding the model to learn expert success case patterns and avoid known defect patterns in S3 is:

[0063] The difference between the design scheme based on cosine similarity calculation and the expert mode:

[0064]

[0065] Where, M is the total number of expert modes; E j Design pattern or defect pattern for the jth expert; w e,j is the mode weight, the successful mode w e,j >0, defect mode w e,j <0.

[0066] Furthermore, the calculation method of the multi-objective loss function in S3 is:

[0067] L=α·L distill +β·Lrule +γ·L perf +δ·L exp

[0068] where α, β, γ, and δ are dynamically adjusted weight coefficients; the initial weight coefficients α0, β0, γ0, and δ0 satisfy α0 + β0 + γ0 + δ0 = 1;

[0069] To meet the requirements of different training stages, a weight dynamic adjustment mechanism is introduced:

[0070] In the initial stage of training, i.e., t ≤ T1: Prioritize learning basic design capabilities and increase α:

[0071]

[0072] In the middle stage of training, i.e., T1 < t ≤ T2: Strengthen rules and performance optimization, and increase β and γ:

[0073]

[0074] In the later stage of training, i.e., t > T2: Focus on experience learning and increase δ:

[0075]

[0076] where t is the number of training steps, T1 and T2 are stage division thresholds, and α0, β0, γ0, and δ0 are initial weight coefficients.

[0077] As the second aspect of the present invention, a multi-knowledge collaborative distillation and optimization system for circuit EMI suppression is provided, which is characterized by including:

[0078] A student model domain adaptation unit, configured to build a basic large language model based on the circuit design and research and development in the electronics industry as the student model to be optimized; and perform adaptation optimization in the circuit design domain for the student model to be optimized;

[0079] A multi-source heterogeneous teacher knowledge system construction unit, configured to build a multi-source heterogeneous teacher knowledge system by integrating knowledge sources including a formal EMI rule library, an EMI simulation data / model library, and an EMI expert experience and case library, and obtain a comprehensive guidance signal G through multi-knowledge source score integration;

[0080] A multi-objective collaborative distillation training unit, configured to perform multi-dimensional evaluation on each candidate circuit design generated by the student LLM through each knowledge source in the teacher knowledge system, and construct a multi-objective collaborative distillation loss function including a basic distillation loss L distill , a rule compliance loss L rule , a performance fitting loss L perf and an expert experience consistency loss L expThe multi-objective loss function includes: retaining general design capabilities by quantifying the difference in output distribution between the student model and the large general language model; imposing penalties on solutions that violate EMI design rules to ensure that the generated design meets industry specifications; minimizing the gap between the EMI performance of the design solution in simulation and the target performance to improve the feasibility of actual engineering; and guiding the model to learn from the expert success case model to avoid known defect patterns; at the same time, introducing a dynamic weight adjustment mechanism to adapt to the needs of different training stages; and adopting an iterative distillation process until the model performance reaches the predetermined target.

[0081] As a third aspect of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored, and the computer program is executed by a processor to perform any step of the above-mentioned multi-knowledge collaborative distillation and optimization method for circuit EMI suppression.

[0082] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0083] 1. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression of the present invention realizes the effective combination of professional knowledge and AI technology by constructing a basic large language model adapted to the field of circuit design. Specifically, a pre-trained language model is used as the basic framework, and professional text data in the field is used for secondary training, covering circuit design specifications, device manuals and other contents. At the same time, prompt engineering is used to guide the model to generate design solutions that meet industry standards, and the model is lightweight. This technical feature enables the model to understand and process professional terms and design rules in the field of circuit design, overcomes the defect of general AI models lacking domain knowledge, helps to improve the professionalism and accuracy of basic design solution generation, and lays the foundation for subsequent EMI optimization.

[0084] 2. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression of the present invention integrates the formalized EMI rule base, simulation data / model base and expert experience case base to build a multi-source heterogeneous teacher knowledge system, thereby realizing the systematic management and quantitative evaluation of EMI knowledge. Specifically, the industry rules are converted into computable logical expressions and evaluation functions, the EMI performance prediction model is trained using historical case data, the degree of expert design pattern matching is quantified through similarity calculation, and finally integrated into a comprehensive guidance signal. This technology realizes the effective fusion of multi-source heterogeneous knowledge, enabling the model to obtain guidance information from multiple dimensions such as rule constraints, performance prediction, and experience reference, solving the problem that the traditional single knowledge source cannot fully guide the design and improving the utilization efficiency of EMI knowledge.

[0085] 3. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression of the present invention realizes the optimization and control of the model training process by constructing a multi-objective loss function including basic distillation, rule compliance, performance fitting and expert experience consistency, and adopting a dynamic weight adjustment mechanism. Specifically, the distribution difference metric is used to retain the general design ability of the teacher model, the penalty mechanism is used to constrain the illegal design, the EMI performance is optimized by minimizing the error, the experience learning is guided based on the similarity calculation, and the weights of each loss are dynamically adjusted according to the training process. This technology ensures that the model balances multiple optimization objectives during training, avoids excessive bias towards a single dimension, effectively improves the model training effect and the overall quality of the generated design solution, and solves the problem of target collaborative optimization during model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a flow chart of a multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to an embodiment of the present invention;

[0087] Figure 2 2 is a diagram of system units according to an embodiment of the present invention. DETAILED DESCRIPTION

[0088] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0089] Example 1

[0090] Please refer to Figure 1 This embodiment 1 provides a multi-knowledge collaborative distillation and optimization method for circuit EMI suppression, including:

[0091] S1. Build a basic large language model based on circuit design and development in the electronics industry as the student model to be optimized. Then, perform adaptive optimization on the student model to be optimized in the circuit design field.

[0092] S2. Construct a multi-source heterogeneous teacher knowledge system by integrating knowledge sources including a formal EMI rule library, an EMI simulation data / model library, and an EMI expert experience and case library. Obtain a comprehensive guidance signal G through the integration of scores from multiple knowledge sources.

[0093] S3. Perform a multi-dimensional evaluation on each candidate circuit design generated by the student LLM through each knowledge source in the teacher's knowledge system, and construct a basic distilled LLM based on the evaluation results of each dimension. distill , rule compliance L rule, performance fitting L perf and expert experience consistency L exp The multi-objective loss function includes: retaining general design capabilities by quantifying the difference in output distribution between the student model and the large general language model; imposing penalties on solutions that violate EMI design rules to ensure that the generated design meets industry specifications; minimizing the gap between the EMI performance of the design solution in simulation and the target performance to improve the feasibility of actual engineering; and guiding the model to learn from the expert success case model to avoid known defect patterns; at the same time, introducing a dynamic weight adjustment mechanism to adapt to the needs of different training stages; and adopting an iterative distillation process until the model performance reaches the predetermined target.

[0094] In the following content, this embodiment further explains the above steps.

[0095] (1) Student model and domain adaptation

[0096] A pre-trained large language model for circuit design (LLM that can generate Verilog descriptions or KiCad or JialiChuang EDA layout file fragments) is selected as the basic student model.

[0097] First, based on the characteristics of the electronic circuit design field, a pre-trained model of appropriate size is selected as the starting point: if higher reasoning efficiency and lower deployment costs are pursued, the DeepSeek1.5B model can be selected. This model has relatively low hardware resource requirements while maintaining a certain language comprehension capability; if stronger complex design task processing capabilities are required, the DeepSeek7B model can be selected, which has more powerful semantic understanding and logical reasoning capabilities.

[0098] Secondly, the selected model is adapted and optimized for the circuit design field: multi-source data in the field of electronic circuit design is collected and organized, including circuit design specification documents, open source circuit design projects, device datasheets, design case analysis reports, etc., to build a domain-specific corpus;

[0099] Using incremental pre-training technology, the corpus is used to further train the basic model, enabling the model to deeply learn professional terminology in the field of circuit design (such as differential routing, power loop, EMI suppression measures, etc.), design rules and expression habits;

[0100] By prompting engineering technology, designing specific prompt templates for circuit design tasks, guiding the model to generate design solutions that meet industry standards and actual needs,

[0101] For example, when generating PCB layout suggestions, the prompt template explicitly requires consideration of constraints such as signal integrity, thermal management, and manufacturability. Furthermore, to improve the model's ability to handle circuit design file formats, a large number of Verilog code, KiCad project files, and Jiali Chuang EDA layout file samples were collected and organized. The model was then trained using code generation and parsing tasks, enabling it to accurately understand and generate design file fragments that meet the specifications of these tools, laying the foundation for subsequent circuit design knowledge distillation and collaborative optimization.

[0102] (2) Construction of a multi-source and heterogeneous teacher knowledge system

[0103] 2.1 Formalized EMI Rule Base

[0104] EMI rules applicable to switch-mode power supplies (SMPS) are systematically reviewed from international standards (such as CISPR22 and CISPR32), industry specifications, and internal company design guidelines. Beyond basic wiring and layout rules, parameter constraints are further refined. For example, it is explicitly stipulated that the loop area formed by the power switch and freewheeling diode must not exceed 50 square millimeters, and the spacing between the input filter capacitor and the input interface must be controlled within 10 millimeters. For high-frequency switching nodes, the trace width must be no less than 0.5 millimeters, and the spacing from sensitive signal lines must be maintained at least 3 millimeters.

[0105] Formal language is used to translate rules into computable evaluation functions. For example, geometric algorithms are used to calculate loop area, distance functions are used to measure component spacing, and Boolean logic is used to determine ground plane integrity. A graded scoring mechanism is implemented for each rule, with full compliance receiving 1 point and partial violations receiving a corresponding deduction based on the degree of deviation. This provides a quantitative basis for subsequent compliance assessments of design solutions.

[0106] In a preferred embodiment, the specific method of converting the rules into a computable evaluation function using a formal language is as follows:

[0107] Extract EMI design rules from documents including international standards, industry specifications and corporate design guidelines, and each rule i Represented as a five-tuple:

[0108] r i =(T i ,P i ,C i ,f i ,w i )

[0109] Where, T i is the rule type, P i is a logical predicate, C i is the parameter constraint, f i is the evaluation function, wi is the weight coefficient;

[0110] The entire rule base R can be expressed as a finite set of rules: R = {r1, r2, ..., r n};

[0111] For design solution D, its rule compliance score R(D) is:

[0112]

[0113] Where n is the total number of rules, f i (D) is the solution D for rule r i The evaluation function value of ;

[0114] 2.2EMI simulation data / model library

[0115] Comprehensively collect case studies of EMI issues in switching power supplies (SMPS), covering various topologies such as flyback, forward, buck, and boost. Each case is deeply analyzed, combining time-domain and frequency-domain analysis data to pinpoint the root cause of the problem. For example, a spectrum analyzer determined that a certain flyback power supply's radiation levels exceeded standards in the 30MHz-200MHz frequency band, attributing this to high-frequency leakage caused by excessive coupling capacitance between the transformer's primary and secondary windings. Simulation software was used to replicate the problem scenario and verify the effectiveness of corrective measures. For example, adding a Faraday shield to the transformer in a simulation model reduced radiation intensity by 15dBμV / m.

[0116] At the same time, we organize excellent switching power supply design cases in the industry and mark the key features in the PCB layout, including the isolation strategy between the power path and the signal path, the topology selection of the filtering circuit, the distribution pattern of the heat dissipation vias, etc., and associate complete EMI test data, covering indicators such as conducted emissions, radiated emissions, and harmonic currents.

[0117] In addition, for typical circuit modules or topologies, a fast EMI prediction agent model is trained based on historical simulation data, and machine learning algorithms such as support vector machines and random forests are used to establish a mapping relationship between design parameters (such as device parameters, trace length, and number of vias) and EMI performance indicators, thereby achieving rapid performance prediction of new design schemes.

[0118] In a preferred embodiment, the specific method for establishing the mapping relationship between the design parameters and the EMI performance indicators is as follows:

[0119] Design Scheme X new The simulation performance evaluation is performed to obtain the simulation performance score S:

[0120] Collect design cases including EMI simulation results of historical projects and other proxy models for training fast EMI prediction for typical circuit modules or topologies, and extract design feature vectors and performance indicator vector

[0121] Train a fast prediction model using machine learning methods:

[0122]

[0123] Minimize the mean squared error (MSE) loss function:

[0124]

[0125] Among them, M is the number of training samples, (X j ,Y j ) is the feature and performance vector of the jth sample;

[0126] For new design X new , its simulation performance score S can be quantified based on the prediction error:

[0127]

[0128] Among them, Y ref is an ideal performance index, and the closer the score is to 1, the better the design performance;

[0129] 2.3EMI Expert Experience and Case Library

[0130] This study collected design notes, technical reports, problem diagnosis records, and typical case studies from senior EMI engineers in the electronics industry, covering various circuit scenarios such as switching power supplies and high-speed PCBs. Natural language processing technology was used to perform multi-level processing on unstructured text. First, named entity recognition was used to extract key terms, such as common-mode choke and Faraday shield. Syntactic analysis was then used to identify design actions, such as shortening SW node traces and adding ground vias, as well as performance impacts, such as a 15dB reduction in radiated noise. A topic model was then used to cluster similar cases, identifying typical design patterns such as power loop optimization and grounding topology design. A causal chain was then constructed for failure cases, such as when poor transformer shielding led to high-frequency leakage and subsequently excessive radiation.

[0131] The case data is further structured into a triple form, namely, design scenario, problem description, and solution. The design scenario includes parameters such as circuit topology and operating frequency, and the solution is refined into device selection recommendations, such as the use of a 680μH common-mode inductor, layout rules, such as the filter capacitor being ≤5mm from the interface, and simulation verification indicators. For successful cases, feature engineering is used to extract the geometric parameters of the PCB layout, such as loop area and trace spacing, and the mapping relationship with EMI test data; for failed cases, a failure mode library is established, marking key defect features such as uneven distribution of thermal vias leading to increased grounding impedance and excessive ground bounce noise. Ultimately, a case data set that combines forward design experience and reverse failure warnings is formed, providing student models with a learnable expert decision-making model and risk avoidance guide.

[0132] In a preferred embodiment, the specific method of guiding the student model through the case data set of forward design experience and reverse failure warning is as follows:

[0133] Using natural language processing methods, we extract design patterns and failure patterns from document information, including design notes, technical reports, problem diagnosis records, and analyses of successful and failed EMI design cases by senior EMI engineers in the electronics industry, and build a pattern database.

[0134] Each mode is represented by a feature vector express;

[0135] The matching score E between the design solution and the expert experience is calculated using cosine similarity:

[0136]

[0137] Among them, Z design is the design scheme characteristic vector, Z expert is the expert mode feature vector; the evaluation scores R(D), S, and E of the above three types of knowledge sources are integrated into a comprehensive guidance score G in a weighted manner.

[0138] 2.4 Integrated Guidance Signal G

[0139] Finally, the scores of the three types of knowledge sources are integrated into the comprehensive guidance signal G:

[0140] G=w r ·R+w s ·S+w e ·E

[0141] Among them, w r 、w s 、w e are the weights of rules, simulation, and experience knowledge sources, respectively, and w r +w s +w e= 1. This comprehensive score G can be used to preliminarily evaluate and rank the overall quality of candidate circuit designs or serve as a reward signal in reinforcement learning. Subsequent multi-objective loss functions are constructed based on more detailed evaluation results from each knowledge source (such as the degree of compliance with each rule, deviations from specific simulation performance indicators, and matching with various expert experience patterns) to guide the model's refined learning.

[0142] (3) Multi-objective collaborative distillation training

[0143] 3.1 Teacher Signal Generation

[0144] For candidate circuit designs generated by students' LLMs, the teacher's knowledge system generates quantitative guidance signals through a multi-dimensional collaborative evaluation mechanism. The rule library performs compliance checks based on geometric constraints and topological structure, calculating the conformance scores of parameters such as power loop area and trace spacing with rule thresholds. A graph neural network is then used to verify that the circuit topology conforms to the pre-set structural pattern. Designs that violate mandatory rules trigger a circuit breaker mechanism, while violations of recommended rules are subject to a point-deducting system that accumulates the cost of violations.

[0145] The simulation data / model extracts key parameters from the design scheme to construct feature vectors, integrates multiple proxy models such as random forests and support vector machines to perform multimodal predictions on indicators such as radiation emissions and conducted noise, outputs confidence intervals and calculates the relative error with the target indicators, and locates key influencing factors through gradient attribution technology.

[0146] Expert experience and case libraries calculate the cosine similarity between the design feature vector and the success mode and failure mode respectively. If the similarity with the failure mode exceeds the threshold, a risk alert is triggered and historical solutions are associated. Based on the successful pattern matching results, improvement suggestions are generated through template filling technology.

[0147] Finally, the rule compliance score, performance prediction error, and experience matching degree are integrated into the final guidance signal G according to dynamic weights, and a SHAP value is generated for each evaluation dimension to quantify the contribution of each factor to the final score, forming an explainable decision basis. G is then converted into a modified gradient in the parameter space to guide the iterative optimization of the student model.

[0148] 3.2 Multi-objective loss function

[0149] Design a comprehensive loss function to guide the training of the student model. The loss function may include:

[0150] Base Distillation Loss: preserves general design capabilities by quantifying the difference in output distribution between the student model and a large general language model.

[0151] The KL divergence is used to measure the distribution difference between the student model and the large general language model. The specific formula is as follows:

[0152]

[0153] Further:

[0154]

[0155] Where D is the training data set, which contains the circuit design requirement description; x is the input text; y is the output token; P T (y|x) and P S (y|x) is the probability distribution of the teacher model and the student model output y under input x; Y is the set of all possible output tokens; |D| is the number of samples in the dataset.

[0156] EMI rule compliance penalty: Penalizes solutions that violate EMI design rules to ensure that the generated design meets industry specifications.

[0157] Evaluate function f based on the rules i , using the negative log-likelihood form:

[0158]

[0159] Where N is the total number of rules; w r,i is the weight of the i-th rule; D S (x) is the design solution generated by the student model based on the input x; f i (D S (x)) is the evaluation score of the design scheme for the i-th rule, ranging from 0 to 1; k i ≥1 is a tuning factor that allows for adjustments to penalty curves for different rules (e.g. k i =2 becomes a square term); introduce severity i ∈(0,1) represents the importance of rule i or the severity of its violation.

[0160] For the key rule, if f i (D S (x))=0, then L rule Set to a maximum value to force the model to avoid violations.

[0161] EMI performance fitting loss: Minimize the gap between the EMI performance of the design solution in simulation and the target performance, improving the feasibility of actual engineering.

[0162] Based on the simulation agent model g(X), the mean square error is used:

[0163]

[0164] Where y targetis the target EMI performance metric vector; g(D S (x)) is the performance prediction value of the simulation surrogate model for the design scheme DS(x);

[0165] When using the reinforcement learning framework, the loss can be converted into a reward:

[0166]

[0167] At this time:

[0168]

[0169] where λ is the reward scaling factor.

[0170] Consistency loss / reward of expert experience: Guide the model to learn the successful case patterns of experts and avoid known defect patterns.

[0171] Calculate the difference between the design scheme and the expert pattern based on cosine similarity:

[0172]

[0173] In the formula, M is the total number of expert patterns; E j is the j-th expert design pattern or defect pattern; w e,j is the pattern weight, the successful pattern w e,j >0, the defect pattern w e,j <0.

[0174] Based on the above loss function, the calculation method of the multi-objective loss function is:

[0175] L = α·L distill + β·L rule + γ·L perf + δ·L exp

[0176] ]》where α, β, γ, δ are dynamically adjusted weight coefficients; the initial weight coefficients α0, β0, γ0, δ0 satisfy α0 + β0 + γ0 + δ0 = 1;

[0177] In a preferred embodiment, to meet the requirements of different training stages, a weight dynamic adjustment mechanism is introduced:

[0178] In the initial stage of training, i.e., t ≤ T1: Prioritize learning the basic design ability and increase α:

[0179]

[0180] In the middle stage of training, i.e., T1 < t ≤ T2: Strengthen the rules and performance optimization, and increase β and γ:

[0181]

[0182] Later in training, i.e., t>T2: Focus on experience learning and increase δ:

[0183]

[0184] Among them, t is the number of training steps, T1 and T2 are the stage division thresholds, and α0, β0, γ0, and δ0 are the initial weight coefficients.

[0185] 3.3 Iterative Collaborative Distillation

[0186] An iterative collaborative distillation mechanism is used to optimize the student LLM, forming a closed-loop learning system. In the first iteration, the student model generates an initial design solution based on the basic distillation loss. The teacher's knowledge system then conducts a multimodal evaluation based on three dimensions: rule compliance, simulation performance, and expert experience. The rule library verifies the geometric constraints and topology of the design solution, outputting a compliance score. The simulation model predicts EMI performance and calculates deviations from the target value. The expert experience library uses pattern matching to identify potential risks and improvement areas.

[0187] The evaluation results of each knowledge source are integrated into a gradient signal to guide the student model's parameter updates. Rule violations are converted into penalty gradients, performance prediction errors are used to construct optimization gradients, and empirical matching generates pattern learning gradients. In the second iteration, the student model adjusts its design strategy based on the feedback from the previous iteration, such as reducing the power loop area or optimizing the grounding topology. The teacher system then re-evaluates the model. This improves rule compliance, but may reveal new performance bottlenecks, such as excessive radiation in a certain frequency band.

[0188] In subsequent iterations, the weights of the multi-objective loss function are dynamically adjusted to strengthen optimization of substandard performance indicators. As iterations progress, the student model gradually grasps the implicit rules of EMI design, such as the impact of high-frequency device layout on parasitic parameters. Training terminates when the fluctuation of the integrated guidance signal G remains below a preset threshold over three consecutive iterations, or when the compliance rate of the design solution on the validation set exceeds a set standard. The resulting model is capable of generating circuit design solutions that simultaneously meet rule constraints, exhibit excellent simulation performance, and align with expert experience.

[0189] (4) Model optimization and application

[0190] The student model trained through the collaborative distillation process generates designs. After receiving user input for circuit requirements, such as topology, power level, and operating frequency, the model first uses a built-in rule engine to screen for compliant component selection options. It then uses an EMI simulation prediction model to evaluate the impact of different layout and routing methods on electromagnetic performance. The model ultimately outputs a complete design solution, including a schematic, PCB layout recommendations, and a parts list. For example, for switching power supply designs, the model can recommend appropriate filter capacitor parameters and plan the optimal routing for the power loop.

[0191] In a design evaluation scenario, after users upload existing design files, the model analyzes them from multiple dimensions. It uses graph neural networks to analyze circuit topology and checks against a rule base for violations such as excessive loop area and split grounding. It uses simulation proxy models to predict conducted and radiated emission metrics and quantify performance deviations. It also compares design features with failure modes in an expert experience library to identify potential risks. Finally, it generates a detailed evaluation report that not only identifies specific problem locations but also provides graded optimization recommendations. For example, it recommends adjusting trace lengths for minor violations and recommends changing the filter circuit topology for serious issues.

[0192] In an interactive design environment, the model operates as a real-time auxiliary tool. While designers are drawing circuit diagrams or performing layout operations in EDA software, the model accesses design data in real time through an API, instantly triggering rule checks and risk predictions. For example, if a high-frequency signal line is detected near a power device, a prompt will pop up, explaining the potential crosstalk risk and providing routing replanning suggestions. During layout and routing, loop area and signal integrity metrics are dynamically calculated, visually demonstrating the impact of the design on EMI performance. This helps designers promptly revise their designs during the design phase and reduces later verification costs.

[0193] Example 2

[0194] Please refer to Figure 2 This embodiment 2 provides a multi-knowledge collaborative distillation and optimization system for circuit EMI suppression, including:

[0195] The student model domain adaptation unit is used to build a basic large language model based on circuit design and development in the electronics industry as the student model to be optimized; and to perform adaptive optimization in the circuit design field for the student model to be optimized;

[0196] The multi-source heterogeneous teacher knowledge system construction unit is used to construct a multi-source heterogeneous teacher knowledge system by integrating knowledge sources including the formal EMI rule library, EMI simulation data / model library, and EMI expert experience and case library, and obtain a comprehensive guidance signal G through the integration of multiple knowledge source scores;

[0197] The multi-objective collaborative distillation training unit is used to perform multi-dimensional evaluation on each candidate circuit design generated by the student LLM through each knowledge source in the teacher's knowledge system, and construct a basic distillation LLM based on the evaluation results of each dimension. distill , rule compliance L rule , performance fitting L perf and expert experience consistency L exp The multi-objective loss function includes: retaining general design capabilities by quantifying the difference in output distribution between the student model and the large general language model; imposing penalties on solutions that violate EMI design rules to ensure that the generated design meets industry specifications; minimizing the gap between the EMI performance of the design solution in simulation and the target performance to improve the feasibility of actual engineering; and guiding the model to learn from the expert success case model to avoid known defect patterns; at the same time, introducing a dynamic weight adjustment mechanism to adapt to the needs of different training stages; and adopting an iterative distillation process until the model performance reaches the predetermined target.

[0198] Example 3

[0199] This embodiment 3 also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement any step of a multi-knowledge collaborative distillation and optimization method for circuit EMI suppression.

[0200] The computer-readable storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc., which can store program codes.

[0201] For an introduction to the computer-readable storage medium provided in this application, please refer to the above method embodiment, and this application will not go into details here.

[0202] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-knowledge collaborative distillation and optimization method for circuit EMI suppression, characterized by: include: S1. Build a basic large language model based on circuit design and development in the electronics industry as the student model to be optimized. Then, perform adaptive optimization on the student model to be optimized in the circuit design field. S2. Build a multi-source heterogeneous teacher knowledge system by integrating knowledge sources including a formal EMI rule library, an EMI simulation data / model library, and an EMI expert experience and case library. The teacher knowledge system can perform multi-dimensional evaluations of candidate circuit designs, and the evaluation results can be summarized into a comprehensive guidance signal G. S3. Perform a multi-dimensional evaluation on each candidate circuit design generated by the student LLM through each knowledge source in the teacher's knowledge system, and construct a basic distilled LLM based on the evaluation results of each dimension. distill , rule compliance L rule , performance fitting L perf and expert experience consistency L exp A multi-objective loss function including ; retains general design capabilities by quantifying the difference in output distribution between the student model and a large general language model; imposes penalties on solutions that violate EMI design rules to ensure that the generated design meets industry specifications; Minimize the gap between the EMI performance of the design solution in simulation and the target performance, and improve the feasibility of actual engineering; The model is guided to learn from expert success case patterns and avoid known defect patterns. At the same time, a dynamic weight adjustment mechanism is introduced to adapt to the needs of different training stages. An iterative distillation process is adopted until the model performance reaches the predetermined target.

2. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to claim 1 is characterized in that: The specific method for adaptive optimization in the circuit design field in S1 is: Collect and organize multi-source data in the field of electronic circuit design, including circuit design specification documents, open source circuit design projects, device datasheets, and design case analysis reports, to build a domain-specific corpus; Using an incremental pre-training method, the corpus is used to further train the basic model, enabling the model to deeply learn the professional terminology, design rules, and expression habits in the field of circuit design; Through the prompt engineering method, specific prompt templates are designed for circuit design tasks to guide the model to generate design solutions that meet industry standards and actual needs; At the same time, in order to collect and organize file samples including Verilog code, KiCad project files, and Jialichuang EDA layout, the model is trained in a targeted manner using code generation and parsing tasks to improve the model's ability to process circuit design file formats, enabling the model to accurately understand and generate design file fragments that comply with the specifications of these tools.

3. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to claim 1 is characterized in that ,The specific methods for multi-dimensional evaluation of the multi-source ,heterogeneous teacher knowledge system in S2 include: The rule compliance evaluation of design solution D is performed to obtain the rule compliance score R(D): Extract EMI design rules from documents including international standards, industry specifications and corporate design guidelines, and each rule i Represented as a quintuple: r i =(T i ,P i ,C i ,f i ,w i ) Where, T i is the rule type, P i is a logical predicate, C i is the parameter constraint, f i is the evaluation function, w i is the weight coefficient; The entire rule base R can be expressed as a finite set of rules: R = {r1, r2, ..., r n }; For design solution D, its rule compliance score R(D) is: Where n is the total number of rules, f i (D) is the solution D for rule r i The evaluation function value of ; Design X new The simulation performance evaluation is performed to obtain the simulation performance score S: Collect design cases including EMI simulation results of historical projects and other proxy models for training fast EMI prediction for typical circuit modules or topologies, and extract design feature vectors and performance indicator vector Train a fast prediction model using machine learning methods: Minimize the mean squared error (MSE) loss function: Among them, M is the number of training samples, (X j ,Y j ) is the feature and performance vector of the jth sample; For new design X new , its simulation performance score S can be quantified based on the prediction error: Among them, Y ref is an ideal performance index, and the closer the score is to 1, the better the design performance; Evaluate the matching degree between the design solution and the expert experience and obtain the experience matching score E: Using natural language processing methods, we extract design patterns and failure patterns from document information, including design notes, technical reports, problem diagnosis records, and analyses of successful and failed EMI design cases by senior EMI engineers in the electronics industry, and build a pattern database. Each mode is represented by a feature vector express; The matching score E between the design solution and the expert experience is calculated using cosine similarity: Among them, Z design is the design scheme characteristic vector, Z expert is the expert mode feature vector; the evaluation scores R(D), S, and E of the above three types of knowledge sources are integrated into a comprehensive guidance score G in a weighted manner: G=w r ·R+w s ·S+w e ·E Among them, w r 、w s 、w e are the weights of rules, simulation, and experience knowledge sources, respectively, and w r +w s +w e =1.

4. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to claim 1 is characterized in that: The specific method of retaining the general design capability in S3 by quantifying the difference in output distribution between the student model and the large general language model is as follows: The KL divergence is used to measure the distribution difference between the student model and the large general language model. The specific formula is as follows: Further: Where D is the training data set, which contains the circuit design requirement description; x is the input text; y is the output token; P T (y|x) and P S (y|x) is the probability distribution of the teacher model and the student model output y under input x; Y is the set of all possible output tokens; |D| is the number of samples in the dataset.

5. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to claim 1 is characterized in that: The specific method of imposing penalties on solutions that violate EMI design rules in S3 to ensure that the generated design meets industry standards is as follows: Evaluate function f based on the rules i , using the negative log-likelihood form: Where N is the total number of rules; w r,i is the weight of the i-th rule; D S (x) is the design solution generated by the student model based on the input x; f i (D S (x)) is the evaluation score of the design scheme for the i-th rule, ranging from 0 to 1; k i ≥1 is a tuning factor that allows for adjustments to penalty curves for different rules (e.g. k i =2 becomes a square term); introduce severity i ∈(0,1) represents the importance of rule i or the severity of its violation; For the key rule, if f i (D S (x))=0, then L rule Set to a maximum value to force the model to avoid violations.

6. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to claim 1 is characterized in that: The specific method for minimizing the gap between the EMI performance of the minimized design solution in simulation and the target performance in S3 and improving the practical engineering feasibility is as follows: Based on the simulation surrogate model g(X), the mean square error is adopted: Where y target is the target EMI performance index vector; g(D S (x)) is the performance prediction value of the simulation agent model for the design solution DS(x); When using the reinforcement learning framework, the loss can be converted into a reward: At this time: where λ is the reward scaling factor.

7. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to claim 1 is characterized in that: The specific method for guiding the model in S3 to learn the successful case patterns of experts and avoid known defect patterns is as follows: Calculate the difference between the design solution and the expert pattern based on cosine similarity: Where, M is the total number of expert modes; E j Design pattern or defect pattern for the jth expert; w e,j is the mode weight, the successful mode w e,j >0, defect mode w e,j <0.

8. The multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to any one of claims 4 to 7, characterized in that: The calculation method of the multi-objective loss function in S3 is as follows: L=α·L distill +β·L rule +γ·L perf +δ·L exp where α, β, γ, δ are dynamically adjusted weight coefficients; the initial weight coefficients α0, β0, γ0, δ0 satisfy α0 + β0 + γ0 + δ0 = 1; To meet the requirements of different training stages, a weight dynamic adjustment mechanism is introduced: In the initial stage of training, i.e., t ≤ T1: Prioritize learning the basic design ability and increase α: In the middle stage of training, i.e., T1 < t ≤ T2: Strengthen rules and performance optimization and increase β and γ: In the later stage of training, i.e., t > T2: Focus on experience learning and increase δ: where t is the training step, T1, T2 are the stage division thresholds, and α0, β0, γ0, δ0 are the initial weight coefficients.

9. A multi-knowledge collaborative distillation and optimization system for circuit EMI suppression, characterized by: It includes: A student model domain adaptation unit for constructing a basic large language model based on the research and development of electronic industry circuit design as the student model to be optimized; And perform adaptation optimization for the student model to be optimized in the circuit design field; A multi-source heterogeneous teacher knowledge system construction unit for constructing a multi-source heterogeneous teacher knowledge system by integrating knowledge sources including a formal EMI rule library, an EMI simulation data / model library, and an EMI expert experience and case library, and obtaining a comprehensive guidance signal G through the integration of multi-knowledge source scores; The multi-objective collaborative distillation training unit is used to perform multi-dimensional evaluation on each candidate circuit design generated by the student LLM through each knowledge source in the teacher's knowledge system, and construct a basic distillation LLM based on the evaluation results of each dimension. distill , rule compliance L rule , performance fitting L perf and expert experience consistency L exp A multi-objective loss function including ; retains general design capabilities by quantifying the difference in output distribution between the student model and a large general language model; imposes penalties on solutions that violate EMI design rules to ensure that the generated design meets industry specifications; Minimize the gap between the EMI performance of the design solution in simulation and the target performance and improve the practical engineering feasibility; And guide the model to learn the successful case patterns of experts and avoid known defect patterns; at the same time, to meet the requirements of different training stages, introduce a weight dynamic adjustment mechanism; and adopt an iterative distillation process until the model performance reaches the predetermined target.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to perform a multi-knowledge collaborative distillation and optimization method for circuit EMI suppression according to any one of claims 1-8.

Images