A heat pipe design expert system that integrates performance prediction and process optimization

A heat pipe design expert system combining the PI-KAN model and the NSGA-III/RVEA algorithm solves the problems of performance prediction and process optimization under multi-parameter coupling and physical constraints of heat pipes, achieving efficient and accurate heat pipe design and improving the physical consistency and interpretability of the design.

CN123116841APending Publication Date: 2026-11-17SOUTH CHINA UNIV OF TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611382798.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing heat pipe design methods suffer from insufficient physical consistency in performance prediction under multi-parameter coupling and physical constraints, difficulty in reverse design, and insufficient ability to optimize process parameters under complex multi-objective constraints.

Method used

The PI-KAN model is combined with the NSGA-III and RVEA algorithms, and physical constraints are embedded for performance prediction and process optimization. Through symbolic formula extraction and knowledge graph construction, explicit mapping and correlation analysis between heat pipe design parameters and performance parameters are realized.

Benefits of technology

It improves the accuracy and efficiency of heat pipe design, shortens the design cycle, reduces R&D costs, and enhances the interpretability of the model and the reusability of design knowledge.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116841A_ABST
    Figure CN123116841A_ABST
Patent Text Reader

Abstract

This invention discloses a heat pipe design expert system that integrates performance prediction and process optimization, belonging to the field of heat pipe manufacturing technology. It includes a forward performance prediction module, a reverse process optimization module, a symbolic formula extraction module, a knowledge graph construction module, and a human-computer interaction module. Based on the Physical Information Kolmogorov-Arnold Network (PI-KAN), the system combines data fitting loss with heat pipe physical constraint loss to predict maximum heat transfer power, thermal resistance, and transient start-up response time. According to the feasible region characteristics constrained by design parameters, it adaptively selects NSGA-III or RVEA for multi-objective optimization to obtain the Pareto optimal design scheme. Furthermore, it performs symbolic regression on the B-spline function in PI-KAN to extract explicit mathematical relationships, and constructs a knowledge graph based on physical constraints to achieve correlation analysis and attribution diagnosis of heat pipe design schemes, thereby improving the efficiency and reliability of heat pipe design.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of heat pipe manufacturing technology, and in particular to a heat pipe design expert system that integrates performance prediction and process optimization. Background Technology

[0002] A heat pipe is a passive heat transfer element that utilizes a working fluid phase change and internal circulation to achieve efficient heat transfer. It features high heat transfer efficiency, compact structure, and requires no external power, making it widely used in fields such as computing chip cooling, aerospace, and new energy. As the heat flux density of related equipment continues to increase and application scenarios become increasingly complex, higher demands are being placed on the heat transfer performance, start-up performance, and structural reliability of heat pipes.

[0003] The performance of a heat pipe is influenced by a variety of design parameters. These parameters typically involve multiple dimensions, including outer layer structure parameters, wick structure parameters, and working fluid parameters. These different parameters are coupled with each other and collectively affect the heat pipe's maximum heat transfer power, thermal resistance, and transient start-up response time. Especially for grooved heat pipes, sintered copper powder heat pipes, braided mesh heat pipes, and composite heat pipes with multiple wick structures, there are complex nonlinear mapping relationships between different structural parameters and working fluid parameters, resulting in a high degree of multi-parameter coupling in heat pipe design.

[0004] Current heat pipe designs typically rely on engineering experience, experimental verification, and simplified empirical formulas. For multi-parameter design problems, it is often necessary to repeatedly adjust design parameters and conduct sample manufacturing and performance testing to determine the design solution, resulting in a long design cycle and high experimental costs. At the same time, simplified empirical formulas cannot fully describe the complex nonlinear coupling relationships between multiple design parameters, and their applicability is limited when faced with different heat pipe structures and complex design constraints.

[0005] In recent years, machine learning methods have been increasingly used to establish the mapping relationship between heat pipe design parameters and performance parameters in order to improve the efficiency of performance prediction. However, existing machine learning-based heat pipe performance prediction methods still have certain shortcomings.

[0006] On the one hand, traditional deep neural networks typically predict performance through multiple layers of nonlinear mapping, and their internal mapping relationships are difficult to convert into explicit mathematical relationships that engineers can directly understand and utilize, resulting in limited interpretability of model prediction results.

[0007] On the other hand, existing machine learning methods mostly focus on predicting heat pipe performance based on given design parameters, i.e., establishing a forward mapping from design parameters to performance parameters. When a real-world design task has given target performance requirements and needs to reverse-engineer the combination of design parameters to meet those requirements, additional parameter search or optimization processes are still required. For multiple interrelated performance indicators such as maximum heat transfer power, thermal resistance, and transient start-up response time, trade-offs need to be made among multiple optimization objectives, thus increasing the complexity of reverse design.

[0008] Furthermore, heat pipe design parameters are constrained by physical conditions such as capillary pressure, working fluid storage capacity, heat transfer limit, pipe material stress, and geometric matching. Under complex physical constraints, the feasible region of design parameters may exhibit nonlinear and nonconvex characteristics, making it difficult to balance the convergence and diversity of candidate solutions when using a single optimization strategy. Meanwhile, the predictive power of purely data-driven models largely depends on the quantity and coverage of training samples. When the number of samples is limited or the parameter combinations exceed the sufficient coverage of existing data, the prediction results may not satisfy the basic physical constraints.

[0009] Therefore, how to improve the accuracy and physical consistency of heat pipe performance prediction under the conditions of multi-parameter coupling and physical constraints, and further realize the connection from performance prediction to multi-objective reverse process optimization, while improving the interpretability of model mapping relationship and the reusability of design knowledge, are technical problems to be solved in the process of intelligent heat pipe design. Summary of the Invention

[0010] This invention addresses the technical problems in the prior art, such as insufficient physical consistency in performance prediction under nonlinear coupling of multiple design parameters of heat pipes, difficulties in reverse design, and insufficient ability to optimize process parameters under complex multi-objective constraints. It proposes a heat pipe design expert system that integrates performance prediction and process optimization. This system establishes a forward performance prediction model driven by both physical information and data by embedding heat pipe physical constraints into the PI-KAN model. The trained PI-KAN is used as the performance evaluation model. Combined with the NSGA-III and RVEA multi-objective optimization algorithms based on adaptive selection of constrained feasible region features, it achieves forward prediction of heat pipe performance parameters and reverse optimization of design parameters. Furthermore, through symbolic formula extraction and knowledge graph construction, it realizes the explicit expression of the mapping relationship between heat pipe design parameters and performance parameters, as well as the correlation analysis and attribution diagnosis of design schemes.

[0011] To solve the technical problem, the technical solution of the present invention is as follows:

[0012] A heat pipe design expert system that integrates performance prediction and process optimization includes a forward performance prediction module, a reverse process optimization module, a symbolic formula extraction module, a knowledge graph construction module, and a human-computer interaction module.

[0013] The positive performance prediction module is configured to acquire heat pipe design parameters and input the heat pipe design parameters into the physical information Kolmogorov-Arnold network PI-KAN, and output heat pipe performance parameters corresponding to the heat pipe design parameters through the PI-KAN; wherein, the training loss of the PI-KAN includes at least the data fitting loss representing the deviation between the prediction result and the sample label, and the physical constraint loss representing the degree to which the prediction result and / or the heat pipe design parameters satisfy the physical constraints of the heat pipe;

[0014] The reverse process optimization module is configured to use the trained PI-KAN as a performance evaluation model for candidate heat pipe design parameters. Based on the constrained feasible region characteristics corresponding to the heat pipe design parameters to be optimized, it selects a multi-objective optimization algorithm between the reference point-based non-dominated sorting genetic algorithm NSGA-III and the reference vector-guided evolutionary algorithm RVEA to iteratively optimize the candidate heat pipe design parameters, and determines the Pareto optimal solution set based on the heat pipe performance parameters output by the PI-KAN.

[0015] The symbolic formula extraction module is configured to obtain the B-spline functions corresponding to at least some of the network edges in the trained PI-KAN, perform network topology sparsification and symbolic function matching on the B-spline functions, and combine them along the forward propagation topology of the PI-KAN according to the symbolic function matching results to obtain an explicit mathematical relationship characterizing the mapping relationship between the heat pipe design parameters and the heat pipe performance parameters.

[0016] The knowledge graph construction module is configured to establish knowledge triples based on the explicit mathematical relations and the physical relations corresponding to the physical constraints. When the heat pipe design scheme in the Pareto optimal solution set does not meet the preset performance requirements and / or meets the preset failure conditions, the module traces the failure association path in reverse along the knowledge triples and feeds back the attribution results to the reverse process optimization module to adjust the search range of candidate heat pipe design parameters.

[0017] The human-computer interaction module is configured to receive heat pipe design requirements and output at least one of the heat pipe performance parameters, the heat pipe design scheme in the Pareto optimal solution set, and the results of the correlation analysis.

[0018] Furthermore, the heat pipe design parameters include outer layer structure parameters, wick structure parameters, and working fluid parameters;

[0019] The outer structure parameters include at least the heat pipe length. , tube outer diameter and the thickness of the tube shell ;

[0020] The structural parameters of the wick are determined according to the type of wick in the heat pipe, wherein:

[0021] For braided wick heat pipes, the wick structural parameters include the number of braids. Woven mesh wire diameter and the number of weaving mesh layers ;

[0022] For a grooved wick heat pipe, the wick structural parameters include the groove depth. trench width and number of trenches ;

[0023] For sintered copper powder wick heat pipes, the wick structural parameters include the mesh size of the copper powder. Powder layer thickness Powder layer distribution and sintering porosity ;

[0024] For a composite heat pipe including at least two wicking structures, the wicking structure parameters include a combination of the respective wicking structure parameters.

[0025] The working fluid parameters include at least the working fluid storage quantity. and working fluid type.

[0026] Furthermore, the heat pipe performance parameters include maximum heat transfer power. Thermal resistance and transient start response time ;

[0027] The PI-KAN includes a learnable univariate function set on the network edge. The learnable univariate function is parameterized using a B-spline function, and the control parameters of the B-spline function are updated through training to establish a nonlinear mapping relationship between the heat pipe design parameters and the heat pipe performance parameters.

[0028] Furthermore, the hyperparameter search space of the PI-KAN includes at least the number of hidden layer neurons, the number of B-spline network intervals, the spline order, and the learning rate;

[0029] The positive performance prediction module searches for the hyperparameters using the RayTune distributed hyperparameter tuning framework and uses the Asynchronous Continuous Halving Algorithm (ASHA) to schedule resources for multiple parallel training experiments.

[0030] The PI-KAN is also configured to adjust the grid node positions corresponding to the B-spline function based on the data distribution of the input training data.

[0031] Furthermore, the physical constraint loss includes at least one of the constraint residuals corresponding to the following physical constraints: mandrel geometric matching constraint, capillary pressure constraint, working fluid storage constraint, heat transfer limit constraint, and tube stress constraint.

[0032] The total loss function of PI-KAN is determined by weighting the data fitting loss and the physical constraint loss corresponding to each physical constraint. By adjusting the weights corresponding to each physical constraint loss, the training of PI-KAN is simultaneously constrained by the sample data and the physical constraints of the heat pipe.

[0033] Furthermore, the reverse process optimization module is configured as follows:

[0034] Obtain the fixed value and / or search range of at least one heat pipe design parameter set by the user;

[0035] With the maximum heat transfer power Maximize the thermal resistance Minimize and the transient startup response time Minimization is used as the multi-objective optimization objective;

[0036] A candidate population consisting of different candidate heat pipe design parameters is generated, and the trained PI-KAN is called to predict the performance of the candidate heat pipe design parameters in the candidate population. The predicted heat pipe performance parameters are used as the fitness evaluation basis for the multi-objective optimization process.

[0037] The candidate population is iteratively evolved to obtain a Pareto optimal solution set containing multiple non-dominant heat pipe design schemes.

[0038] Furthermore, the reverse process optimization module determines the multi-objective optimization algorithm to be used based on the characteristics of the feasible region constrained by the heat pipe design parameters;

[0039] When the constrained feasible region is continuous and the physical constraints are smooth, NSGA-III is adopted, and multi-objective optimization of candidate heat pipe design parameters is performed based on the reference point mechanism.

[0040] When the constrained feasible region is cut by nonlinear physical constraints and exhibits non-convex and / or fault characteristics, RVEA is adopted, and the convergence and population diversity of candidate solutions are adjusted based on angle-penalized distance APD.

[0041] Furthermore, the symbolic formula extraction module is configured as follows:

[0042] The activity of the B-spline function corresponding to the network edge in the trained PI-KAN is evaluated, and the network edges with activity below a preset threshold are pruned based on the L1 regularization result to obtain a sparse network topology.

[0043] The B-spline functions corresponding to the network edges retained in the sparse network topology are sampled, and univariate symbolic regression is performed using a candidate symbolic function library. The candidate symbolic function library includes at least linear functions, polynomial functions, power functions, exponential functions, logarithmic functions, and trigonometric functions.

[0044] According to the coefficient of determination The fitting results of the candidate sign functions are evaluated, and the candidate sign functions that meet the preset fitting conditions are selected to replace the corresponding B-spline functions.

[0045] According to the forward propagation topology of the replaced PI-KAN, the symbolic functions are nested and combined and algebraically simplified to obtain an explicit mathematical analytical expression that represents the mapping relationship between multiple heat pipe design parameters and at least one heat pipe performance parameter. The explicit mathematical analytical expression is substituted back into the physical constraint equation to perform residual backtesting. When the residual backtesting result does not meet the preset threshold, the univariate symbolic regression is re-executed until the preset threshold is met.

[0046] Furthermore, the knowledge graph construction module is configured as follows:

[0047] Based on the explicit mathematical analytical expression and the physical constraint equation corresponding to the physical constraint, the process parameters, intermediate physical quantities, performance indicators and failure phenomena in heat pipe design are abstracted into entities.

[0048] The entities and relationships between entities in the explicit mathematical expressions and physical constraint equations are analyzed to generate directed knowledge triples that characterize the relationships between the process parameters, intermediate physical quantities, performance indicators, and / or failure phenomena. The directed knowledge triples are stored in a graph database, wherein the failure phenomenon entities correspond one-to-one with the physical constraint residuals.

[0049] When the heat pipe design scheme does not meet the preset performance requirements and / or meets the preset failure conditions, the corresponding performance index entity and / or failure phenomenon entity are used as the graph query node. The reverse tracing is performed along the association path formed by the directional knowledge triplet to determine the failure association path corresponding to the heat pipe design scheme, and the attribution diagnosis result is output.

[0050] A heat pipe design expert method that integrates performance prediction and process optimization, the method comprising:

[0051] Step S1: Obtain the heat pipe design parameters of the heat pipe to be designed and / or the constraint range corresponding to the heat pipe design parameters;

[0052] Step S2: Input the heat pipe design parameters into the trained PI-KAN, and obtain the corresponding heat pipe performance parameters through the PI-KAN. During the training process, the PI-KAN establishes the mapping relationship between the heat pipe design parameters and the heat pipe performance parameters through a training loss function that includes data fitting loss and physical constraint loss.

[0053] Step S3: Based on the constrained feasible region characteristics corresponding to the heat pipe design parameters, determine a multi-objective optimization algorithm from NSGA-III and RVEA for the current heat pipe design task, using the maximum heat transfer power... Maximize the thermal resistance Minimize and the transient startup response time Minimization is used as the multi-objective optimization objective. The PI-KAN is used as the performance evaluation model for candidate heat pipe design parameters. The candidate heat pipe design parameters are iteratively optimized to obtain the Pareto optimal solution set.

[0054] Step S4: Perform network topology sparsification and sign function matching on the B-spline functions corresponding to at least some network edges in the trained PI-KAN. Based on the sign function matching results, combine them along the forward propagation topology of the PI-KAN to obtain an explicit mathematical analytical expression characterizing the mapping relationship between the heat pipe design parameters and the heat pipe performance parameters.

[0055] Step S5: Establish a knowledge triplet based on the explicit mathematical analytical expression and the physical relationship corresponding to the physical constraint, and perform correlation analysis on at least one heat pipe design scheme in the Pareto optimal solution set based on the knowledge triplet. When the design scheme does not meet the preset performance requirements and / or meets the preset failure conditions, trace the failure correlation path in reverse along the knowledge triplet, and feed back the attribution result to step S3 to adjust the search range.

[0056] Step S6: Output at least one heat pipe design scheme and its corresponding heat pipe performance parameters from the Pareto optimal solution set.

[0057] The present invention has the following advantages:

[0058] This invention discloses a heat pipe design expert system that integrates performance prediction and process optimization. It constructs a PI-KAN prediction model by embedding physical law constraints into a Kolmogorov-Arnold network to perform high-precision forward performance predictions of heat pipe process parameters that conform to physical laws. Leveraging the model's efficient reasoning capabilities, it combines NSGA-III and RVEA adaptive multi-objective evolution algorithms to perform reverse process optimization, generating a Pareto optimal solution set for heat pipe design parameters under physical boundary constraints. Simultaneously, it extracts symbolic formulas and constructs a knowledge graph from the trained model to analyze the explicit mathematical relationships between heat pipe process parameters and performance parameters. Through this closed-loop process, heat pipe R&D professionals can efficiently and accurately conduct multi-parameter comprehensive evaluations and design scheme selection, significantly shortening the heat pipe design and verification cycle, reducing R&D trial-and-error costs, and improving the ability to customize heat pipe designs on demand for industrial production. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This invention provides an overall system architecture diagram of a heat pipe design expert system that integrates performance prediction and process optimization.

[0061] Figure 2 A flowchart of a forward performance prediction module for a heat pipe design expert system that integrates performance prediction and process optimization, provided in an embodiment of the present invention;

[0062] Figure 3 A flowchart of the reverse process optimization module of a heat pipe design expert system that integrates performance prediction and process optimization is provided in an embodiment of the present invention.

[0063] Figure 4 A flowchart of a symbolic formula extraction module for a heat pipe design expert system that integrates performance prediction and process optimization, provided in an embodiment of the present invention;

[0064] Figure 5 The flowchart of the knowledge graph construction module of a heat pipe design expert system that integrates performance prediction and process optimization is provided in an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] Example 1:

[0067] This embodiment provides a heat pipe design expert system that integrates performance prediction and process optimization, including a forward performance prediction module, a reverse process optimization module, a symbolic formula extraction module, a knowledge graph construction module, and a human-computer interaction module.

[0068] In this embodiment, the process parameters in the heat pipe design parameters are divided into three categories according to their physical functions: outer layer structure parameters, wick structure parameters, and working fluid parameters.

[0069] The outer structural parameters are used to describe the overall physical morphology of the heat pipe, including at least the length of the heat pipe. ,diameter Wall thickness This set of parameters applies to all types of heat pipes;

[0070] The specific meaning of the wick parameters depends on the type of wick in the heat pipe:

[0071] For braided wick heat pipes, the wick structural parameters include the number of braids. Woven mesh wire diameter and the number of weaving mesh layers ;

[0072] For a grooved wick heat pipe, the wick structural parameters include the groove depth. trench width and number of trenches ;

[0073] For sintered copper powder wick heat pipes, the wick structural parameters include the mesh size of the copper powder. Powder layer thickness Powder layer distribution and sintering porosity .

[0074] For a composite heat pipe containing two or more wicking structures, the wicking structure parameters are the union of the corresponding wicking parameters.

[0075] The working fluid parameters are used to describe the charging state of the working medium inside the heat pipe, and include at least the working fluid storage volume. And the working fluid type; this set of parameters applies to all types of heat pipes.

[0076] In this embodiment, the heat pipe performance parameters include the maximum heat transfer capacity of the heat pipe. Thermal resistance and transient start response time .

[0077] In this embodiment, the forward performance prediction module is based on the physical information Kolmogorov-Arnold network. According to the Kolmogorov-Arnold theorem, any multivariate continuous function can be represented as a finite superposition of univariate continuous functions. In the KAN network, the input features undergo nonlinear transformation through the univariate functions on the edges, followed by linear summation at the nodes. In this embodiment, these univariate functions on the edges are parameterized as B-spline curves. The network updates the control points of the B-splines through a backpropagation algorithm, thereby learning a high-precision function mapping relationship.

[0078] This embodiment integrates the Ray Tune distributed hyperparameter tuning framework during the model training phase. The PI-KAN model adopts a 3-layer network structure, and the system adjusts the number of hidden layer neurons. B-spline network interval number spline order and learning rate The search space is defined as follows (adjusted according to heat pipe type). The RayTune framework employs the Asynchronous Successive Halving Algorithm (ASHA) as its underlying resource scheduling strategy. By executing multiple trials in parallel, it terminates poorly performing hyperparameter combinations in the early stages of training and concentrates computational power on training combinations with potential. To accelerate model convergence and adapt to the distribution patterns of different heat pipe data, the model automatically adjusts the node positions of the B-spline network based on statistical data distribution during the first training batch input.

[0079] To address the physical inconsistency issue arising from purely data-driven training models, this embodiment transforms the prior physical laws of heat pipe design into residual term embeddings in the loss function. Specifically, the prior physical constraints are as follows:

[0080] (a) Mandrel geometric matching constraints: ;

[0081] (b) Capillary pressure constraint: ;

[0082] (c) Working fluid storage quantity constraints: ;

[0083] (d) Heat transfer limit constraint: ;

[0084] (e) Pipe stress constraint: .

[0085] Introducing physical constraint loss Then, the loss is fitted using the model data. The model's total loss function Defined as:

[0086]

[0087] in, To adapt to dynamic weights, their initial value range is incorporated into the search space of the Ray Tune framework for joint optimization. During training, the model automatically adjusts through a gradient penalty mechanism to balance data fitting and physical constraints, enabling the model to update weights within the physically feasible domain.

[0088] The reverse engineering optimization module described in this embodiment specifically refers to an adaptive multi-objective optimization model built based on the Non-Dominated Sorting Genetic Algorithm (NSGA-III) and the Reference Vector Guided Evolutionary Algorithm (RVEA), using a trained PI-KAN model as the fitness evaluator. When the system determines that the feasible region of the heat pipe design parameters exhibits continuous characteristics and smooth physical constraints, the system calls the NSGA-III algorithm to perform global parallel optimization in the multi-dimensional objective space using a preset uniform reference point mechanism. When the system determines that the heat pipe design parameters are subjected to severe nonlinear physical constraints that result in a highly non-convex feasible region or the existence of discontinuities, the system adaptively calls the RVEA algorithm, using the Angle Penalized Distance (APD) to dynamically balance convergence and population diversity. The adaptive combination of the NSGA-III and RVEA algorithms endows the reverse engineering optimization module with extremely strong industrial robustness, enabling the system to obtain effective engineering solutions for both conventional design requirements and extreme special requirements. Furthermore, the adaptive combination of the two algorithms can solve the problems of "diversity degradation" and "boundary point loss" in the feasible region of heat pipe design parameters caused by the introduction of physical constraints in the KAN model, ensuring that the reverse process optimization module achieves a dynamic balance between convergence and diversity at extreme physical boundaries.

[0089] The symbolic formula extraction module described in this embodiment specifically refers to an analytical module used to explicitly represent the complex nonlinear features implicit in high-dimensional neural networks. According to the Kolmogorov-Arnold theorem, any multivariate continuous function can be represented as a finite superposition of univariate continuous functions. In the KAN architecture, the system deploys learnable nonlinear activation functions (parameterized as B-spline continuous functions in this embodiment) on the network's "edges," and the neuron nodes only perform pure linear addition operations. This topology ensures that the mapping process of the prediction model trained based on PI-KAN to high-dimensional complex multivariate functions can be decoupled into multiple independent univariate one-dimensional mappings, thereby significantly reducing the dimensionality of symbolic regression and providing a mathematical foundation for subsequent lossless replacement of symbolic functions and explicit algebraic formula assembly.

[0090] Specifically, the symbolic formula extraction module has a built-in candidate symbolic function library, which covers commonly used operators characterizing fundamental physical laws, including at least linear functions (such as...). ), polynomial functions (such as , ), power functions (such as ), exponential function (such as ), logarithmic function (such as ) and trigonometric functions (such as , After training the PI-KAN model, this module performs automated extraction of explicit mathematical formulas and physical consistency loop closure verification through the following steps:

[0091] Step S1: Network topology sparsification based on L1 regularization;

[0092] Since the PI-KAN network initially trained in this embodiment is fully connected and contains redundant or weakly represented mapping relationships, the system first evaluates the weight activity (L1 norm) of all B-spline activation function edges in the network, removes edges with activity evaluation values ​​below a set threshold and free neuron nodes that have lost all input or output connections, and retains the core physical mapping path.

[0093] Step S2: Univariate symbolic regression matching based on the candidate function library;

[0094] After sparsification in step S1, each active edge retained in the grid represents a one-dimensional B-spline curve from a single variable input to a single variable output. The system performs discrete resampling on the retained B-spline curves within their domain and calls the candidate symbol function library to perform fitting attempts sequentially using the symbolic regression algorithm to match one-dimensional curve fitting for the active edges of the grid.

[0095] Step S3: Goodness-of-fit evaluation and formula nesting assembly;

[0096] After completing the one-dimensional curve fitting of the active edge, the system uses the coefficient of determination R0 2 The fitting accuracy of each candidate symbol function is quantitatively evaluated, and the candidate symbol function with the simplest structure and greater than the fitting threshold is selected to replace the original B-spline function on the grid edge. After completing the high-precision symbol replacement of the entire network, after merging and simplifying similar terms, the system finally outputs one or more explicit multivariate algebraic analytical expressions containing minimal mathematical terms. Through these analytical expressions, the deep coupling relationship between multiple process parameters of the heat pipe is explicitly expressed.

[0097] Step S4: Physical residual backtesting and closed-loop iteration;

[0098] Substitute the explicit mathematical analytical expression obtained in step S3 back into the prior physical constraint equation of the heat pipe to calculate the backtest residual value of the explicit formula under the global physical constraint. When the backtest residual exceeds the preset residual threshold, the adaptive refitting mechanism is triggered to adaptively increase the fitting accuracy threshold of the univariate symbolic regression or expand the order of the basis functions of the candidate symbolic function library, and re-execute the univariate symbolic regression and algebraic assembly in steps S2 to S3 until the output explicit mathematical analytical expression meets the preset physical residual threshold to ensure the physical consistency of the analytical expression.

[0099] The knowledge graph construction module described in this embodiment specifically refers to a method that transforms explicit mathematical relationships and physical laws into a machine-readable structured cognitive network. The knowledge graph construction module performs graph construction and advanced reasoning through the following three core steps:

[0100] S1: Abstraction and construction of the ontological model of heat pipe design parameters;

[0101] The system abstracts various elements in heat pipe design into a multi-dimensional set of "entities." These entities are categorized into at least the following: process parameter entities, such as pipe diameter, wall thickness, and working fluid storage capacity; intermediate physical quantity entities, such as effective vapor chamber diameter, equivalent porosity, and capillary suction force; and performance index entities, such as the overall thermal resistance of the heat pipe. Maximum heat transfer power Transient start response time Etc.; physical failure phenomena, such as capillary limit failure, shell stress failure, etc.

[0102] S2: Knowledge triple extraction and storage;

[0103] After obtaining the explicit analytical expression output by the symbolic formula extraction module, the module's built-in analytical algorithm automatically performs semantic segmentation and relation mapping on these two mathematical expressions, combining them with the prior physical constraint equations pre-embedded in the PI-KAN model loss function. This translates the formula into directed knowledge triples in the format <entity>→[relation]→<entity>, and persistently stores them in the graph database, thus completing the knowledge graph construction.

[0104] S3: Attribution diagnosis based on knowledge graph;

[0105] After the graph is built, when a heat pipe design scheme generated by the system is judged by the system to be "not meeting performance requirements" or "failed", the engineer can initiate a diagnosis through this module. The module will take the "failure phenomenon entity" as the starting point, trace the shortest path or the path with the maximum weight in the graph in reverse, and output the source diagnosis to realize the design assistance and knowledge reuse functions of the expert system.

[0106] The human-computer interaction module described in this embodiment specifically refers to a front-end interface used to input heat pipe design requirements, visualize the results, and assist users in understanding and adjusting model training and prediction. Specifically, it includes:

[0107] S1: Interactive interface for model training functions;

[0108] The model training interface provides a dataset import interface. The system automatically parses the data headers on the front end, mapping parameters to "input feature columns" and "target label columns." The model training interface also includes a PI-KAN model hyperparameter configuration interface, allowing users to adjust the model network architecture, activate the required physical constraint equations, and set the weighting coefficients for the physics-data dual-drive loss. ;

[0109] Specifically, after the distributed hyperparameter tuning begins, the system frontend, through the integrated Ray Tune state monitoring interface, tracks the evolution of multiple parallel trials, selects the three optimal candidate network configurations (Top-3 Trials) from the hyperparameter search space, and compares and plots the dual-track loss curves of these three trials in a visualization chart. For each optimal trial, a curve tracks the data fitting loss. The downward trend of one curve, and another curve tracking the physical constraint residuals. The convergence state provides engineers with multi-dimensional comparative decision-making basis for determining the optimal model; the globally optimal PI-KAN model, along with its best hyperparameter combination, is trained and screened and persistently saved to the model library in the form of a file.

[0110] S2: Interactive interface for positive performance prediction function;

[0111] After the model training and deployment are completed, the system provides an interface for forward performance prediction of heat pipe design. After the user loads the deployed PI-KAN proxy model into the interface, the front-end system automatically parses the model's metadata and dynamically generates a process parameter form. The user can input the heat pipe process parameters according to design requirements. When the user initiates a prediction command, the system sends the process parameter form to the PI-KAN model, which predicts the heat pipe performance parameters. Finally, the prediction results and physical constraints are visualized.

[0112] S3: Reverse engineering optimization function interactive interface;

[0113] In this workflow, the user sets the floating-point search range for the target process parameters in the optimization constraint panel. A candidate population is generated based on the NSGA-III / RVEA adaptive optimization engine. In each generation of evolution, the positive performance prediction module is concurrently called to evaluate the population performance. After the calculation is complete, the front end calls the graphics rendering component to map the Pareto front solution set, containing hundreds of dominant solutions, into a visual scatter plot. The graph's axes are bound to... , and Users can generate corresponding heat pipe process parameter work orders by selecting points in the Pareto chart.

[0114] Example 2:

[0115] This embodiment applies to Embodiment 1, proposing a specific application scenario for a heat pipe design expert system that integrates performance prediction and process optimization. In this embodiment, the target heat pipe type is a copper powder grooved composite heat pipe. The heat pipe design parameters, including the process parameters, are categorized into three types based on their physical function: outer layer structure parameters, wick structure parameters, and working fluid parameters, specifically including:

[0116] (1) Outer layer structural parameters: total length of heat pipe , tube outer diameter Pipe wall thickness ;

[0117] (2) Liquid absorption core structural parameters:

[0118] (i) Trench morphological characteristics: trench depth trench width Number of trenches ;

[0119] (ii) Characteristics of copper powder sintering: mesh size of copper powder particles , sintering powder layer thickness sintering porosity ;

[0120] (3) Working fluid parameters: amount of working fluid stored Working fluid type;

[0121] like Figure 1 As shown, this embodiment provides a heat pipe design expert system that integrates performance prediction and process optimization. This expert system assists users in providing the optimal design solution within the target design range of the heat pipe. In specific implementation, the system includes five modules: a forward performance prediction module, a reverse process optimization module, a symbolic formula extraction module, a knowledge graph construction module, and a human-computer interaction module. The specific steps are as follows:

[0122] Step S1, Forward Performance Prediction Module: such as Figure 2 As shown, the system first initiates the training and deployment process of the PI-KAN model.

[0123] (1) Model initialization and adaptive mesh adjustment: The system reads the historical dataset of the composite heat pipe containing process parameters of three feature dimensions, normalizes the input process parameters and output labels, instantiates the PI-KAN network, and configures the network structure as follows: After the first batch of data is input, the update_grids mechanism is called to adaptively adjust the node distribution of the B-spline network from a uniform distribution to a grid distribution that fits the data range of the composite heat pipe process.

[0124] (2) Multi-objective loss constraints and parameter optimization: The system initiates a RayTune distributed hyperparameter intelligent tuning loop to search for the optimal hyperparameter combination in parallel. After each forward propagation of the network, the system fits the loss based on the data. and physical loss constraints In this embodiment, the physical constraints specifically include: mandrel geometric matching constraints. Capillary pressure constraint Working fluid storage quantity constraints Heat transfer limit constraint Pipe stress constraint Total loss function The Adam optimizer performs backpropagation step optimization until the validation set evaluation meets the early stopping condition, and finally the optimal model weights are saved to the model library.

[0125] (3) Model Inference: After the user inputs the fixed values ​​or constraint ranges of the heat pipe design parameters in the human-computer interaction interface, the system verifies the validity of the parameters, loads the prediction model weights, normalizes the input data, opens a sub-thread for the PI-KAN model to infer the heat pipe performance parameters, and outputs the results on the visualization chart after denormalization. , and Performance metrics.

[0126] Step S2, Adaptive reverse engineering optimization: such as Figure 3 As shown, when a user needs to reverse engineer a design scheme based on performance metrics, the reverse engineering optimization module is invoked.

[0127] (1) Parameter configuration: Configure multi-objective optimization parameters in the human-computer interaction interface, including population size. Iterative Algebra NSGA-III algorithm parameters (number of dimension splits) ) and RVEA algorithm parameters (penalty rate) Reference vector adaptive update frequency .

[0128] (2) Adaptive Algorithm Routing: Start the Optimization Worker sub-thread, and the system automatically evaluates the data characteristics of the current heat pipe process parameter space. When the parameter space is continuous and the physical constraints are smooth, the system selects the NSGA-III algorithm and uses the uniform reference point mechanism to perform global parallel optimization; when the feasible region of heat pipe process parameter optimization is cut by physical constraints, resulting in the parameter feasible region exhibiting non-convex or discontinuous characteristics, the system selects the RVEA algorithm and uses the APD (Angle Penalty Distance) mechanism to perform non-convex penetration optimization.

[0129] (3) Predicting heat pipe process parameters output: The optimizer calls the PI-KAN model to evaluate fitness and finally returns a Pareto optimal solution set containing multiple non-dominant solutions. The Pareto interactive chart is rendered. When the user clicks on a scatter point in the chart, the system automatically extracts the corresponding heat pipe process parameters and adds them to the prediction table to generate a processing work order.

[0130] Step S3, Symbol Formula Extraction Module: (e.g.) Figure 4 As shown, after the PI-KAN model is trained, the system can extract explicit mathematical relationships and perform physical residual backtesting.

[0131] (1) Active edge topology traversal: The system loads the pre-trained PI-KAN model weights and traverses each active edge in the network topology;

[0132] (2) Curve sampling and dimensionality reduction decoupling: B-spline curve sampling is performed on the high-dimensional multivariate function on the active edge, and the function is mapped and decoupled into multiple independent one-dimensional B-spline curves;

[0133] (3) Symbolic Regression: Call the built-in candidate symbol function library, use the symbolic regression algorithm, and use basic mathematical operators (such as exponential function, logarithmic function, polynomial function, etc.) to match the one-dimensional curve;

[0134] (4) Fit evaluation and formula assembly: Perform R 2Goodness-of-fit evaluation is performed to select the explicit sign function with the highest accuracy, which replaces the B-spline curve in the model. Finally, the assembled explicit multivariate sign formula and a visual relationship chart are output through algebraic operations.

[0135] (5) Backtesting of the a priori physical equation residuals: Substitute the assembled explicit multivariate symbolic formula back into the capillary limit equilibrium equation and working fluid charging physical equation of the copper powder trench composite heat pipe to calculate the physical residuals under the extreme conditions; when the physical residuals exceed the preset threshold, automatically adjust the regression fitting accuracy and re-execute steps (2) to (4) until the backtested residuals converge to the threshold range, and finally output an explicit algebraic analytical expression that has both prediction accuracy and physical interpretability.

[0136] Step S4, Knowledge Graph Construction Module: such as Figure 5 As shown, the system stores the isolated knowledge generated by the preceding tasks in a structured manner.

[0137] (1) Entity abstraction: The system predefines entity types, including (e.g., pipe diameter, wall thickness, copper powder mesh count), intermediate physical quantity entities (e.g., pore size, steam chamber size), and performance index entities (e.g., ... , , ) as well as failure phenomena entities (such as capillary limit constraints and working fluid storage constraints).

[0138] (2) Triple generation: The explicit mathematical formulas output by the extraction module and the a priori physical constraint equations built into the system are used to generate directional knowledge triples such as <copper powder grooved composite heat pipe> → [copper powder mesh / sintering porosity] → (determines) → [effective pore radius] → (inversely proportional to) → [capillary suction force] → (cannot be overcome) → [total gas-liquid pressure drop] → (leads to) → <capillary limit failure>.

[0139] (3) Attribution diagnosis: The generated knowledge triples are stored in the graph database. When the heat pipe scheme prediction fails to meet the target, the system can query the shortest path in the graph and output intuitive attribution diagnosis to guide the user to modify the process parameters.

[0140] This embodiment addresses the problems of existing heat pipe design methods, which mainly rely on engineers' experience and trial-and-error and simplified empirical formulas, resulting in long design cycles, high costs, and difficulty in comprehensively considering multiple parameters. It provides a heat pipe design expert system that integrates performance prediction and process optimization. By embedding physical law constraints into a Kolmogorov-Arnold network to construct a PI-KAN prediction model, it performs high-precision forward performance predictions of heat pipe process parameters that conform to physical laws. Leveraging the model's efficient reasoning capabilities, it combines NSGA-III and RVEA adaptive multi-objective evolution algorithms to perform reverse process optimization, generating a Pareto optimal solution set for heat pipe design parameters under physical boundary constraints. Simultaneously, it extracts symbolic formulas and constructs a knowledge graph from the trained model to analyze the explicit mathematical relationships between heat pipe process parameters and performance parameters. Through this closed-loop process, heat pipe R&D professionals can efficiently and accurately conduct multi-parameter comprehensive evaluations and design scheme selection, significantly shortening the heat pipe design and verification cycle, reducing R&D trial-and-error costs, and improving the ability to customize heat pipe designs on demand for industrial production.

[0141] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0142] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A heat pipe design expert system integrating performance prediction and process optimization, characterized in that, It includes a forward performance prediction module, a reverse process optimization module, a symbolic formula extraction module, a knowledge graph construction module, and a human-computer interaction module; The positive performance prediction module is configured to acquire heat pipe design parameters and input the heat pipe design parameters into the physical information Kolmogorov-Arnold network PI-KAN, and output heat pipe performance parameters corresponding to the heat pipe design parameters through the PI-KAN; wherein, the training loss of the PI-KAN includes at least the data fitting loss representing the deviation between the prediction result and the sample label, and the physical constraint loss representing the degree to which the prediction result and / or the heat pipe design parameters satisfy the physical constraints of the heat pipe; The reverse process optimization module is configured to use the trained PI-KAN as a performance evaluation model for candidate heat pipe design parameters. Based on the constrained feasible region characteristics corresponding to the heat pipe design parameters to be optimized, it selects a multi-objective optimization algorithm between the reference point-based non-dominated sorting genetic algorithm NSGA-III and the reference vector-guided evolutionary algorithm RVEA to iteratively optimize the candidate heat pipe design parameters, and determines the Pareto optimal solution set based on the heat pipe performance parameters output by the PI-KAN. The symbolic formula extraction module is configured to obtain the B-spline functions corresponding to at least some of the network edges in the trained PI-KAN, perform network topology sparsification and symbolic function matching on the B-spline functions, and combine them along the forward propagation topology of the PI-KAN according to the symbolic function matching results to obtain an explicit mathematical relationship characterizing the mapping relationship between the heat pipe design parameters and the heat pipe performance parameters. The knowledge graph construction module is configured to establish knowledge triples based on the explicit mathematical relations and the physical relations corresponding to the physical constraints. When the heat pipe design scheme in the Pareto optimal solution set does not meet the preset performance requirements and / or meets the preset failure conditions, the module traces the failure association path in reverse along the knowledge triples and feeds back the attribution results to the reverse process optimization module to adjust the search range of candidate heat pipe design parameters. The human-computer interaction module is configured to receive heat pipe design requirements and output at least one of the heat pipe performance parameters, the heat pipe design scheme in the Pareto optimal solution set, and the results of the correlation analysis.

2. The heat pipe design expert system according to claim 1, characterized in that, The heat pipe design parameters include outer layer structure parameters, wick structure parameters, and working fluid parameters; The outer structure parameters include at least the heat pipe length. , tube outer diameter and the thickness of the tube shell ; The structural parameters of the wick are determined according to the type of wick in the heat pipe, wherein: For braided wick heat pipes, the wick structural parameters include the number of braids. Woven mesh wire diameter and the number of weaving mesh layers ; For a grooved wick heat pipe, the wick structural parameters include the groove depth. trench width and number of trenches ; For sintered copper powder wick heat pipes, the wick structural parameters include the mesh size of the copper powder. Powder layer thickness Powder layer distribution and sintering porosity ; For a composite heat pipe including at least two wicking structures, the wicking structure parameters include a combination of the respective wicking structure parameters. The working fluid parameters include at least the working fluid storage quantity. and working fluid type.

3. The heat pipe design expert system according to claim 1, characterized in that, The heat pipe performance parameters include maximum heat transfer power. Thermal resistance and transient start response time ; The PI-KAN includes a learnable univariate function set on the network edge. The learnable univariate function is parameterized using a B-spline function, and the control parameters of the B-spline function are updated through training to establish a nonlinear mapping relationship between the heat pipe design parameters and the heat pipe performance parameters.

4. The heat pipe design expert system according to claim 1, characterized in that, The hyperparameter search space of the PI-KAN includes at least the number of hidden layer neurons, the number of B-spline network intervals, the spline order, and the learning rate; The positive performance prediction module searches for the hyperparameters using the RayTune distributed hyperparameter tuning framework and uses the Asynchronous Continuous Halving Algorithm (ASHA) to schedule resources for multiple parallel training experiments. The PI-KAN is also configured to adjust the grid node positions corresponding to the B-spline function based on the data distribution of the input training data.

5. The heat pipe design expert system according to claim 1, characterized in that, The physical constraint loss includes at least one of the constraint residuals corresponding to the following physical constraints: mandrel geometric matching constraint, capillary pressure constraint, working fluid storage constraint, heat transfer limit constraint, and tubular stress constraint. The total loss function of PI-KAN is determined by weighting the data fitting loss and the physical constraint loss corresponding to each physical constraint. By adjusting the weights corresponding to each physical constraint loss, the training of PI-KAN is simultaneously constrained by the sample data and the physical constraints of the heat pipe.

6. The heat pipe design expert system according to claim 1, characterized in that, The reverse process optimization module is configured as follows: Obtain the fixed value and / or search range of at least one heat pipe design parameter set by the user; With the maximum heat transfer power Maximize the thermal resistance Minimize and the transient startup response time Minimization is used as the multi-objective optimization objective; A candidate population consisting of different candidate heat pipe design parameters is generated, and the trained PI-KAN is called to predict the performance of the candidate heat pipe design parameters in the candidate population. The predicted heat pipe performance parameters are used as the fitness evaluation basis for the multi-objective optimization process. The candidate population is iteratively evolved to obtain a Pareto optimal solution set containing multiple non-dominant heat pipe design schemes.

7. The heat pipe design expert system according to claim 6, characterized in that, The reverse process optimization module determines the multi-objective optimization algorithm to be used based on the characteristics of the feasible region constrained by the heat pipe design parameters. When the constrained feasible region is continuous and the physical constraints are smooth, NSGA-III is adopted, and multi-objective optimization of candidate heat pipe design parameters is performed based on the reference point mechanism. When the constrained feasible region is cut by nonlinear physical constraints and exhibits non-convex and / or fault characteristics, RVEA is adopted, and the convergence and population diversity of candidate solutions are adjusted based on angle-penalized distance APD.

8. The heat pipe design expert system according to claim 1, characterized in that, The symbol formula extraction module is configured as follows: The activity of the B-spline function corresponding to the network edge in the trained PI-KAN is evaluated, and the network edges with activity below a preset threshold are pruned based on the L1 regularization result to obtain a sparse network topology. The B-spline functions corresponding to the network edges retained in the sparse network topology are sampled, and univariate symbolic regression is performed using a candidate symbolic function library. The candidate symbolic function library includes at least linear functions, polynomial functions, power functions, exponential functions, logarithmic functions, and trigonometric functions. According to the coefficient of determination The fitting results of the candidate sign functions are evaluated, and the candidate sign functions that meet the preset fitting conditions are selected to replace the corresponding B-spline functions. According to the forward propagation topology of the replaced PI-KAN, the symbolic functions are nested and combined and algebraically simplified to obtain an explicit mathematical analytical expression that represents the mapping relationship between multiple heat pipe design parameters and at least one heat pipe performance parameter. The explicit mathematical analytical expression is substituted back into the physical constraint equation to perform residual backtesting. When the residual backtesting result does not meet the preset threshold, the univariate symbolic regression is re-executed until the preset threshold is met.

9. The heat pipe design expert system according to claim 8, characterized in that, The knowledge graph construction module is configured as follows: Based on the explicit mathematical analytical expression and the physical constraint equation corresponding to the physical constraint, the process parameters, intermediate physical quantities, performance indicators and failure phenomena in heat pipe design are abstracted into entities. The entities and relationships between entities in the explicit mathematical expressions and physical constraint equations are analyzed to generate directed knowledge triples that characterize the relationships between the process parameters, intermediate physical quantities, performance indicators and / or failure phenomena. The failure phenomenon entities correspond one-to-one with the physical constraint residuals, and the directed knowledge triples are stored in a graph database. When the heat pipe design scheme does not meet the preset performance requirements and / or meets the preset failure conditions, the corresponding performance index entity and / or failure phenomenon entity are used as the graph query node. The reverse tracing is performed along the association path formed by the directional knowledge triplet to determine the failure association path corresponding to the heat pipe design scheme, and the attribution diagnosis result is output.

10. A heat pipe design expert method integrating performance prediction and process optimization, characterized in that, The method includes: Step S1: Obtain the heat pipe design parameters of the heat pipe to be designed and / or the constraint range corresponding to the heat pipe design parameters; Step S2: Input the heat pipe design parameters into the trained PI-KAN, and obtain the corresponding heat pipe performance parameters through the PI-KAN. During the training process, the PI-KAN establishes the mapping relationship between the heat pipe design parameters and the heat pipe performance parameters through a training loss function that includes data fitting loss and physical constraint loss. Step S3: Based on the constrained feasible region characteristics corresponding to the heat pipe design parameters, determine a multi-objective optimization algorithm from NSGA-III and RVEA for the current heat pipe design task, using the maximum heat transfer power... Maximize the thermal resistance Minimize and the transient startup response time Minimization is used as the multi-objective optimization objective. The PI-KAN is used as the performance evaluation model for candidate heat pipe design parameters. The candidate heat pipe design parameters are iteratively optimized to obtain the Pareto optimal solution set. Step S4: Perform network topology sparsification and sign function matching on the B-spline functions corresponding to at least some network edges in the trained PI-KAN. Based on the sign function matching results, combine them along the forward propagation topology of the PI-KAN to obtain an explicit mathematical analytical expression characterizing the mapping relationship between the heat pipe design parameters and the heat pipe performance parameters. Step S5: Establish a knowledge triplet based on the explicit mathematical analytical expression and the physical relationship corresponding to the physical constraint, and perform correlation analysis on at least one heat pipe design scheme in the Pareto optimal solution set based on the knowledge triplet. When the design scheme does not meet the preset performance requirements and / or meets the preset failure conditions, trace the failure correlation path in reverse along the knowledge triplet, and feed back the attribution result to step S3 to adjust the search range. Step S6: Output at least one heat pipe design scheme and its corresponding heat pipe performance parameters from the Pareto optimal solution set.