A multiphysics simulation design platform and method for three-dimensional braided electromagnetic devices
By using multi-scale proxy models and digital twin calibration technology, the contradiction between computational accuracy and efficiency in the simulation analysis of three-dimensional braided electromagnetic devices was resolved, realizing a high-precision and high-efficiency design and manufacturing closed loop, and improving the success rate of prototype trial production and production efficiency.
Patent Information
- Application Number
- CN202611344884.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-01
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies suffer from a contradiction between computational accuracy and efficiency in the simulation analysis of three-dimensional braided electromagnetic devices, a disconnect between design and manufacturing, and a lack of evolutionary capability due to static models, making it impossible to achieve a high-precision and high-efficiency design and manufacturing closed loop.
Multi-scale surrogate models are used to accelerate simulation, and equivalent homogenization is used to ensure accuracy. A closed loop between design and manufacturing is achieved through process instruction generation and digital twin calibration. This includes technologies such as multi-physics simulation design platform, machine learning multi-scale surrogate models, and online digital twin calibration.
It achieves high-precision and high-efficiency simulation analysis, integrated design and manufacturing, and self-evolving digital twin capabilities, significantly improving the success rate of prototype trial production and production efficiency.
Smart Images

Figure CN123116697A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer-aided engineering (CAE) and intelligent manufacturing technology, and in particular to a multiphysics simulation design platform and method for three-dimensional braided electromagnetic devices. Background Technology
[0002] Three-dimensional braided composite materials are widely used in aerospace, new energy, and rail transportation due to their excellent mechanical properties and design flexibility. With the development of modern electrical equipment towards higher frequencies, higher power densities, and greater integration, the fabrication of electromagnetic devices with complex structures (such as three-dimensional braided inductors, transformer windings, and radomes) using three-dimensional braiding technology has become a research hotspot. However, these devices contain multiple material components, including conductive fibers, magnetically controlled fibers, insulating fibers, and structural fibers, and the braided structure exhibits complex anisotropic characteristics, resulting in extremely complex electromagnetic-thermal-mechanical multiphysics coupling behavior.
[0003] In existing technologies, the analysis of three-dimensional braided devices typically suffers from the following drawbacks:
[0004] The contradiction between computational accuracy and efficiency: If a full-size detailed model is used in traditional finite element simulation (FEM), the computational load is huge and it takes hours or even days, which is difficult to meet the design iteration requirements; if a homogeneous model is used, microscopic topological features such as local braiding angles and fiber bundle contacts are often ignored, resulting in large simulation errors (usually >20%).
[0005] Design and manufacturing are disconnected: Existing simulation software (such as ANSYS Maxwell and COMSOL) only outputs geometric models (STEP / STL format) or performance cloud maps, which cannot be directly converted into process control instructions for three-dimensional weaving equipment (such as yarn carrier movement trajectory, spindle speed, tension control, etc.). Design results often need to be repeatedly modified because the process cannot be realized, resulting in long R&D cycles and high trial production costs.
[0006] The model is static and lacks evolutionary ability: Once the existing simulation model is established, it is fixed and cannot be self-corrected and optimized using the measured data accumulated in the production process (such as actual inductance, temperature rise, weaving defects, etc.), which leads to a decrease in prediction accuracy over time.
[0007] Therefore, there is an urgent need for an intelligent simulation design platform that can balance high precision and high efficiency and connect the entire data flow of "design-simulation-manufacturing". Summary of the Invention
[0008] This invention aims to solve the above-mentioned technical problems by providing a multiphysics simulation design platform and method for three-dimensional braided electromagnetic devices. It accelerates simulation through multi-scale proxy models, ensures accuracy through equivalent homogenization, and achieves a closed loop of design and manufacturing through process instruction generation and digital twin calibration.
[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0010] In a first aspect, the present invention provides a three-dimensional braided electromagnetic device multiphysics simulation design platform, comprising: a multiphysics simulation design platform body;
[0011] The input interface module is used to receive the three-dimensional geometric model, material database, performance constraints, and process constraints of the target three-dimensional braided electromagnetic device.
[0012] The geometry-material parameterization module, connected to the input interface module, is used to discretize the target device into four functional regions: conductive region, magnetocontrollable region, insulating region, and structural region, and generate a parameterized description containing the volume fraction of each region, local braiding angle, and fiber bundle arrangement parameters. The parameterized description is stored in the form of voxels or finite element meshes.
[0013] The equivalent homogenization module is used to construct representative volume elements based on parametric descriptions. By applying periodic boundary conditions, it calculates the equivalent electromagnetic parameters, equivalent thermal parameters, and equivalent mechanical parameters of each functional zone by combining analytical mixing law and periodic finite element method.
[0014] The machine learning multi-scale surrogate model module is used to build a surrogate model from macroscopic design parameters to multi-physics response. The surrogate model is trained with high-fidelity simulation data from the equivalent homogenization module and is used to predict device performance indicators in seconds.
[0015] The multiphysics coupling solution module is used to solve the response of the target device under electromagnetic-thermal-mechanical multiphysics coupling conditions based on the equivalent parameters output by the equivalent homogenization module or the prediction results of the machine learning multiscale surrogate model module.
[0016] The design optimization and process instruction output module is used to adjust the design parameters according to the performance index constraints and solution results, and generate a process instruction file that conforms to the communication protocol of the three-dimensional knitting equipment control system. The process instruction file includes at least the spindle speed instruction, the yarn carrier track coordinate sequence, the knitting tension instruction, and the yarn changing control instruction.
[0017] The database and visualization module is connected to all the aforementioned modules and is used to uniformly store and manage geometric models, material properties, process parameters, simulation results, process instructions and historical measured data in the design process. It also provides a real-time visual interactive interface for 3D geometric model reconstruction, microstructure display, multiphysics distribution cloud map rendering and optimization process tracking.
[0018] The digital twin online calibration module is used to compare the measured data of the physical prototype with the simulation prediction results, calculate the relative deviation, and when the relative deviation exceeds the preset threshold, use the measured data to incrementally train the surrogate model or correct the material database through parameter inversion.
[0019] The external interface module is used to output process instruction files to the three-dimensional knitting equipment control system and impedance matching system.
[0020] Furthermore, the equivalent homogenization module is configured to perform the following steps:
[0021] Construct cuboids with side lengths of 2mm to 10mm as representative volume units, with each side of the representative volume unit containing at least two complete weaving cycles;
[0022] Periodic boundary conditions are applied to representative volume elements, and the continuity of displacement, electric, magnetic and temperature fields on the boundary is ensured by master-slave node coupling or periodic mesh pairing.
[0023] A uniform electric field, magnetic field, temperature gradient, or mechanical strain is applied to a representative volume element, and the equivalent electrical conductivity tensor, equivalent magnetic permeability tensor, equivalent thermal conductivity tensor, and equivalent elastic matrix are obtained through volume averaging calculations.
[0024] Furthermore, the machine learning multi-scale agent model module is implemented using a fully connected neural network or a graph neural network;
[0025] The input feature vector of the surrogate model includes at least the volume fraction of conductive fibers, the volume fraction of magnetic fibers, the braiding angle, and the operating frequency.
[0026] The output response vector of the surrogate model should include at least the inductance, AC resistance, temperature rise, and mechanical stiffness parameters.
[0027] The proxy model is trained with 200 to 1000 training samples generated by high-fidelity finite element analysis, and the single inference time is less than 1 second.
[0028] Furthermore, in the process instruction file generated by the design optimization and process instruction output module:
[0029] The spindle speed command ranges from 0.1 rpm to 3000 rpm;
[0030] The coordinate accuracy of the yarn carrier's path coordinate sequence is ±0.02mm;
[0031] The range of weaving tension instructions is 5cN to 200cN;
[0032] The yarn changing control command includes the switching sequence and spatial coordinates of conductive fiber bundles, magnetic fiber bundles, insulating fiber bundles and structural fiber bundles, and the yarn changing time does not exceed 2 seconds.
[0033] Furthermore, the digital twin online calibration module is configured as follows:
[0034] Calculate the relative deviation δi between the simulation prediction and the measured value;
[0035] when If the variance of the surrogate model's prediction exceeds a preset threshold, it is determined to be a low-confidence region.
[0036] In the low confidence region, incremental learning is used to add the measured samples to the training set and retrain the surrogate model locally until the relative deviation between the updated predicted value and the measured value is reduced to less than 5%.
[0037] Furthermore, the parametric descriptions generated by the geometry-material parametric module also include Litz transposition pitch;
[0038] Litz transposition pitch is used to characterize the transposition period of conductive fiber bundles within the conductive region, and Litz transposition pitch is one of the input features of the machine learning multi-scale surrogate model module.
[0039] Secondly, this invention provides a multiphysics simulation design method for three-dimensional braided electromagnetic devices, comprising the following steps:
[0040] S1: Receive the three-dimensional geometric model, material database, performance constraints, and process constraints of the target three-dimensional braided electromagnetic device;
[0041] S2: Discretize the target device into conductive region, magnetocontrollable region, insulating region and structural region, and generate a parameterized description including local braiding angle and fiber bundle arrangement parameters;
[0042] S3: Construct representative volume elements and calculate the equivalent electromagnetic parameters, equivalent thermal parameters, and equivalent mechanical parameters of each functional zone using analytical mixing law and periodic finite element method;
[0043] S4: Call the machine learning multi-scale proxy model to quickly predict the multiphysics response based on equivalent parameters or macroscopic design parameters;
[0044] S5: Solve the electromagnetic-thermal-mechanical response of the target device based on the output of equivalent parameters or surrogate models;
[0045] S6: Compare the simulation prediction results with the actual measured data of the physical prototype, calculate the relative deviation, and when the relative deviation exceeds the preset threshold, use the actual measured data to perform incremental training on the surrogate model.
[0046] S7: Adjust the design parameters according to the performance index constraints and generate a process instruction file that conforms to the communication protocol of the three-dimensional knitting equipment control system. The process instruction file includes spindle speed instructions, yarn carrier trajectory coordinate sequence, knitting tension instructions and yarn changing control instructions.
[0047] Furthermore, step S5, the multiphysics coupling solution, includes:
[0048] Solve Maxwell's equations in the frequency domain to obtain the current density distribution, magnetic field distribution, and loss of the device at the operating frequency.
[0049] Electromagnetic loss is used as a bulk heat source and input into the steady-state or transient heat conduction equation to calculate the device temperature rise.
[0050] When the change in material properties caused by temperature rise exceeds 5%, update the conductivity and permeability and re-solve the electromagnetic field until the change in performance indicators converges.
[0051] Furthermore, step S7, generating the process instruction file, includes:
[0052] Establish a mapping relationship between design parameters and process control parameters. The mapping relationship will convert the optimized braiding angle, fiber volume fraction and device size into the spindle speed, yarn trajectory coordinate sequence and braiding tension setting value of the three-dimensional braiding equipment.
[0053] Based on the partitioned structure of the device, control commands for yarn changing of the yarn carrier at different spatial positions are generated to achieve sequential switching of conductive fiber bundles, magnetic fiber bundles, insulating fiber bundles and structural fiber bundles.
[0054] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described multiphysics simulation design method for three-dimensional braided electromagnetic devices.
[0055] The beneficial effects of this invention are: it combines high precision and high efficiency: by capturing the influence of microstructure on macro performance through RVE equivalent homogenization, the simulation error is controlled within ±5%; at the same time, the introduction of a machine learning proxy model reduces the single evaluation time from hours to seconds, supporting large-scale design space exploration.
[0056] Design and manufacturing integration: It breaks down the data barriers between traditional simulation and manufacturing, and directly maps the optimized design scheme into the process control instructions (speed, trajectory, tension, yarn change) of the three-dimensional weaving equipment, which significantly improves the success rate of prototype trial production and production efficiency.
[0057] Self-evolving digital twins: A closed-loop feedback mechanism based on measured data has been established. Through incremental learning, model deviations are continuously corrected, making the simulation platform increasingly accurate as production data accumulates, thus solving the problem of model drift during long-term production.
[0058] Data integration throughout the entire lifecycle: Data interaction with braiding equipment, impedance matching systems, and manufacturing methods is achieved through standardized interfaces, enabling intelligent manufacturing. Attached Figure Description
[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0060] Figure 1 The system architecture diagram of the multiphysics simulation design platform for three-dimensional braided electromagnetic devices provided in the embodiments of the present invention is shown.
[0061] Figure 2 This is a flowchart of multi-scale modeling in an embodiment of the present invention.
[0062] Figure 3 This is a schematic diagram of the closed-loop design in an embodiment of the present invention.
[0063] Figure 4 This is a schematic diagram of the equivalent homogenized representative volume element (RVE) in an embodiment of the present invention.
[0064] Figure 5 This is a diagram illustrating the training and inference architecture of a multi-scale proxy model for machine learning, provided in an embodiment of the present invention.
[0065] Figure 6 A flowchart illustrating the multiphysics simulation design method for three-dimensional braided electromagnetic devices provided in this embodiment of the invention.
[0066] In the attached diagram, 1-Multiphysics simulation design platform body; 2-Input interface module; 3-Geometric-material parameterization module; 4-Equivalent homogenization module; 5-Machine learning multi-scale surrogate model module; 6-Multiphysics coupling solution module; 7-Design optimization and process instruction output module; 8-Database and visualization module; 9-Digital twin online calibration module; 10-External interface module; 11-Conductive region; 12-Magnetically controlled region; 13-Insulating region; 14-Structural region. Detailed Implementation
[0067] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0068] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0069] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0070] Reference Figures 1 to 5 A specific embodiment of the present invention provides a multiphysics simulation design platform for three-dimensional braided electromagnetic devices. This platform integrates a complete functional workflow from geometric modeling, material parameterization, multi-scale simulation, intelligent optimization to digital twin calibration, aiming to solve the design challenges of electromagnetic-thermal-mechanical multiphysics coupling in three-dimensional braided electromagnetic devices under complex operating conditions, and to achieve seamless integration of design and manufacturing.
[0071] like Figure 1 As shown, the main body 1 of the three-dimensional braided electromagnetic device multiphysics simulation design platform provided in this embodiment mainly includes: an input interface module 2, a geometry-material parameterization module 3, an equivalent homogenization module 4, a machine learning multi-scale surrogate model module 5, a multiphysics coupling solution module 6, a design optimization and process instruction output module 7, a database and visualization module 8, a digital twin online calibration module 9, and an external interface module 10. The modules communicate with each other via a high-speed data bus to achieve real-time data interaction and transfer.
[0072] Specifically, input interface module 2 serves as the entry point for platform-user interaction, receiving the 3D geometric model, material database, performance constraints, and process constraints of the target 3D braided electromagnetic device. During implementation, the user uploads a 3D geometric model in formats such as STEP, IGES, or STL through a graphical user interface. The system's built-in geometry engine performs topology checks and repairs on the model to ensure its geometric closure. Simultaneously, the user-input material data is stored in a structured manner, including not only basic material constants but also complex conductivity varying with frequency, nonlinear hysteresis loops, and anisotropic thermal conductivity. Furthermore, the user must input performance indicators including the target inductance value, quality factor, maximum operating temperature rise, and stiffness threshold, as well as process constraints such as the physical limits of the 3D braiding equipment, including spindle speed and tension range. This module standardizes the above information and transmits it to the subsequent processing unit.
[0073] The geometry-material parameterization module 3 is connected to the input interface module 2. Its core task is to transform the continuous macroscopic geometric model into a physically meaningful microscopic parameterized field. This module first uses an adaptive voxelization algorithm or tetrahedral mesh generation technique to discretize the input geometric model in space with high precision. Then, based on preset material property mapping rules and electromagnetic function distribution, it identifies and marks the functional attributes of each micro-unit, discretizing the entire device into four functional zones: a conductive zone 11 (composed of conductive fiber bundles), a magnetocontrol zone 12 (composed of magnetic fiber bundles), an insulating zone 13 (composed of insulating fibers), and a structural zone 14 (composed of structural fibers). Based on this, the module performs deep parameterization extraction, accurately calculating the orientation information of the fiber bundles in three-dimensional space by tracing the yarn path vector, and obtaining the local braiding angle. Simultaneously, it analyzes and records the fiber bundle arrangement parameters, such as cross-sectional shape, filling density, and stacking method. The generated parameterized description includes the volume fraction of each zone, the local braiding angle, and the fiber bundle arrangement parameters, and is stored in the form of voxel attribute arrays or finite element mesh node data, providing a foundation for subsequent calculations.
[0074] The equivalent homogenization module 4 is used to construct representative volume elements (RVEs) based on the parametric description generated by the geometry-material parameterization module 3. The RVEs are designed as cuboids with side lengths between 2 mm and 10 mm, and each side contains at least two complete weaving cycles. The module applies periodic boundary conditions to the RVEs, ensuring the continuity of the physical field on the relative boundaries through master-slave node coupling technology. Furthermore, the module employs a hybrid computational strategy combining analytical mixing laws and periodic finite element methods, applying uniform electric fields, magnetic fields, temperature gradients, and mechanical strains to the RVEs. The microscopic response is obtained through a finite element solver, and the macroscopic equivalent performance tensors are calculated using the volume averaging principle. Specifically, these include: an equivalent conductivity tensor reflecting electrical conductivity, an equivalent permeability tensor reflecting magnetic properties, an equivalent thermal conductivity tensor characterizing thermal conductivity, and an equivalent elasticity matrix describing mechanical behavior.
[0075] The Machine Learning Multi-Scale Surrogate Model Module 5 is used to build surrogate models from macroscopic design parameters to multiphysics responses. This module deploys a neural network (such as DNN or GNN) built on a deep learning framework and uses high-fidelity simulation data generated by the Equivalent Homogenization Module 4 for offline training. The trained surrogate model can predict device performance indicators within seconds (typically less than 1 second). Its input feature vector integrates macroscopic equivalent material parameters and microstructural details (such as fiber volume fraction, local braiding angle, Litz transposition pitch, and operating frequency); the output covers core indicators such as inductance, AC resistance, maximum temperature rise, and mechanical stiffness. This allows subsequent design optimization to abandon time-consuming high-fidelity finite element solutions, thereby significantly improving iteration efficiency.
[0076] The multiphysics coupling solution module 6 is used to solve the response of the target device under electromagnetic-thermal-mechanical multiphysics coupling conditions based on the equivalent parameters output by the equivalent homogenization module 4 or the prediction results of the machine learning multiscale surrogate model module 5. This module employs a bidirectional sequential coupling strategy: first, it solves Maxwell's equations in the frequency domain to obtain the current density distribution, magnetic field distribution, and losses of the device at its operating frequency; second, it uses electromagnetic losses as a volume heat source to input steady-state or transient heat conduction equations to calculate the device's temperature rise; finally, it uses the temperature field as a load for mechanical analysis. Crucially, this module introduces an iterative mechanism: when the material properties change due to the temperature rise exceed 5%, it automatically updates the conductivity and permeability and re-solves the electromagnetic field until the performance indicators converge, thus ensuring high computational accuracy.
[0077] The design optimization and process instruction output module 7 is crucial for the platform's integrated design and manufacturing capabilities. Based on the performance constraints and the solution results from the multiphysics coupling solution module 6, this module uses a multi-objective optimization algorithm to adjust design parameters and find the optimal solution. Subsequently, the module establishes a mapping relationship between design parameters and process control parameters, converting the optimized weaving angle, fiber volume fraction, and device dimensions into the spindle speed, yarn trajectory coordinate sequence, and weaving tension setpoints for the 3D weaving equipment. Specifically, based on the device's partitioned structure, the module generates yarn-changing control instructions for the yarn carrier at different spatial positions, enabling the sequential switching of conductive fiber bundles, magnetic fiber bundles, insulating fiber bundles, and structural fiber bundles. The final generated process instruction file conforms to the communication protocol of the 3D weaving equipment control system and includes spindle speed instructions, yarn carrier yarn trajectory coordinate sequences, weaving tension instructions, and yarn-changing control instructions.
[0078] The database and visualization module 8 serves as the platform's data hub, responsible for storing and managing all geometric models, material data, parametric descriptions, simulation results, and process documents. Simultaneously, this module provides rich visualization capabilities, capable of displaying 3D geometric models, discretized functional partitions, and multiphysics cloud maps (such as magnetic flux density distribution, temperature field distribution, and stress-strain distribution) to users in a dynamic 3D graphical format, supporting intuitive inspection and analysis of design details.
[0079] The digital twin online calibration module 9 compares the measured data of the physical prototype with the simulation prediction results and calculates the relative deviation. When the relative deviation exceeds a preset threshold (e.g., 5%), the system initiates an incremental learning mechanism. This module adds samples containing measured data to the training set and uses optimization algorithms to locally retrain the machine learning multi-scale surrogate model module 5, correcting the network weights to make the model approximate the physical truth. Furthermore, this module can also correct the initial parameters in the material database through parameter inversion, ensuring the consistency between the digital model and the physical entity.
[0080] The external interface module 10 is used to output process instruction files to the 3D knitting equipment control system and impedance matching system. This module supports multiple industrial communication protocols (such as OPC UA and Modbus TCP), and can transmit the instruction files generated by the design optimization and process instruction output module 7 to the production equipment in real time, driving the 3D knitting equipment to perform automated production, and simultaneously adjusting the parameters of the impedance matching system to achieve closed-loop control from design to manufacturing.
[0081] In summary, through the collaborative work of the above modules, this invention achieves full-process digitalization and intelligentization of three-dimensional braided electromagnetic devices, from microstructure design and multi-physics performance prediction to process parameter generation, greatly improving design efficiency and manufacturing precision.
[0082] Furthermore, the equivalent homogenization module is configured to perform the following steps:
[0083] Construct cuboids with side lengths of 2mm to 10mm as representative volume units, with each side of the representative volume unit containing at least two complete weaving cycles;
[0084] Periodic boundary conditions are applied to representative volume elements, and the continuity of displacement, electric, magnetic and temperature fields on the boundary is ensured by master-slave node coupling or periodic mesh pairing.
[0085] A uniform electric field, magnetic field, temperature gradient, or mechanical strain is applied to a representative volume element, and the equivalent electrical conductivity tensor, equivalent magnetic permeability tensor, equivalent thermal conductivity tensor, and equivalent elastic matrix are obtained through volume averaging calculations.
[0086] In this embodiment, as Figure 2 As shown, the equivalent homogenization module 4 serves as a bridge connecting the microscopic braided structure with the macroscopic device performance. Its main task is to transform the three-dimensional braided composite material, which contains complex anisotropic characteristics, into a homogenized continuous dielectric model. This module first obtains a detailed parametric description of the target device from the geometry-material parameterization module 3. These descriptions include key geometric and topological information such as the spatial distribution, volume fraction, and local braiding angles of the conductive, magnetosensitive, insulating, and structural regions. Based on these inputs, the equivalent homogenization module 4 performs a series of highly automated numerical calculation steps to extract high-precision equivalent parameters for subsequent macroscopic simulations.
[0087] Specifically, the first step performed by the equivalent homogenization module 4 is to construct representative volumetric units. To ensure that the constructed units accurately and statistically reflect the periodic microstructure of the 3D woven material at the macroscopic scale, the system generates a geometry with a regular shape based on the input fiber bundle arrangement parameters and weaving path. For example... Figure 4As shown, in this embodiment, the RVE is constructed as a cuboid with a side length ranging from 2mm to 10mm. This size range is set based on a balance between computational accuracy and efficiency: too small a size cannot contain sufficient microstructural details, while too large a size increases unnecessary computational burden. More importantly, the length of each side of the cuboid RVE must be designed to contain at least two complete weaving cycles. This constraint ensures that the RVE possesses periodic repeating characteristics in all spatial directions, thus satisfying the basic assumptions of homogenization theory and ensuring that the extracted equivalent parameters are volumetrically representative. Inside the RVE, the system accurately reproduces the interweaving morphology of fiber bundles in different functional zones, including the routing paths of conductive fiber bundles in the conductive region, the distribution of magnetic fiber bundles in the magnetocontrol region, and the filling of the insulating and structural regions.
[0088] After constructing the RVE geometric model, the equivalent homogenization module 4 enters the boundary condition application stage, which is the core step to ensure the accuracy of the calculation results. To simulate the actual physical behavior of materials in an infinite medium, the module applies strict periodic boundary conditions to the RVE. In numerical implementation, the system employs master-slave node coupling or periodic mesh pairing. Specifically, the system pairs corresponding nodes or mesh surfaces on opposite faces in the three coordinate directions of the RVE, establishes constraint equations, and forces the degrees of freedom on these paired nodes or surfaces to satisfy specific linear relationships. This approach effectively eliminates boundary effects and ensures the strict continuity of the displacement, electric, magnetic, and temperature fields on the relative boundaries of the RVE. For example, for the displacement field, it ensures that the deformation of one boundary of the RVE is consistent with the deformation of the opposite boundary, thus simulating the deformation compatibility of a continuous medium; for the electromagnetic field, it ensures that the phase and amplitude of the potential and magnetic potential match on the periodic boundaries. Through this periodic constraint, the RVE is virtually treated as a unit in an infinitely repeating structure, allowing calculation results at the microscale to be accurately extrapolated to the macroscale.
[0089] Finally, the equivalent homogenization module 4 performs homogenization calculations and parameter extraction steps. After applying complete periodic boundary conditions, the module uses a finite element solver to apply uniform loads of different physical fields to the RVE. First, a uniform electric field is applied, and the current density distribution inside the RVE is calculated. The equivalent conductivity tensor is obtained through the volume averaging method, which accurately reflects the anisotropic conductivity characteristics of the conductive region. Second, a uniform magnetic field is applied, and the equivalent permeability tensor is calculated by combining the nonlinear hysteresis characteristics of the magnetostrictive region. Simultaneously, a uniform temperature gradient is applied, and the heat conduction equation is solved to calculate the equivalent thermal conductivity tensor to evaluate the heat conduction capacity in the complex fiber network. In addition, the module also applies uniform mechanical strain and calculates the equivalent elastic matrix through structural mechanics. All the above calculations are based on the volume averaging principle, that is, by integrating the microscopic field quantity within the volume of the RVE and dividing by the volume to obtain the macroscopic equivalent quantity. The four types of equivalent parameters (equivalent conductivity tensor, equivalent magnetic permeability tensor, equivalent thermal conductivity tensor, and equivalent elastic matrix) obtained by calculation are finally output to the multiphysics coupling solution module 6 or used to generate training samples for the machine learning multi-scale surrogate model module 5, providing a solid data foundation for high-precision macroscopic simulation of three-dimensional braided electromagnetic devices.
[0090] Furthermore, the machine learning multi-scale agent model module is implemented using a fully connected neural network or a graph neural network;
[0091] The input feature vector of the surrogate model includes at least the volume fraction of conductive fibers, the volume fraction of magnetic fibers, the braiding angle, and the operating frequency.
[0092] The output response vector of the surrogate model should include at least the inductance, AC resistance, temperature rise, and mechanical stiffness parameters.
[0093] The proxy model is trained with 200 to 1000 training samples generated by high-fidelity finite element analysis, and the single inference time is less than 1 second.
[0094] In this embodiment, as Figure 5 As shown, the machine learning multi-scale surrogate model module 5, serving as the platform's core acceleration engine, is configured to use deep learning algorithms to construct high-precision surrogate models, replacing time-consuming traditional finite element iterative calculations, thereby achieving second-level evaluation of design schemes. The core architecture of this module can flexibly employ either a fully connected neural network (DNN) or a graph neural network (GNN). Fully connected neural networks are suitable for processing conventional parameterized vectors, while graph neural networks can more efficiently capture the complex topological connections of fiber bundles in space.
[0095] During the model building and training phase, the machine learning multi-scale surrogate model module 5 first establishes a mapping relationship between input and output. Its input feature vector contains at least the key design parameters extracted from the geometry-material parameterization module 3, specifically including: the volume fraction of conductive fibers in the conductive region, the volume fraction of magnetic fibers in the magnetized region, the local braiding angle that determines the anisotropy of the device, and the actual operating frequency of the device. Correspondingly, the output response vector of the surrogate model is directly defined as the key performance indicators of the target device, including at least inductance, AC resistance, maximum temperature rise, and mechanical stiffness.
[0096] To ensure the model's prediction accuracy, the system invested significant computational resources during the training data acquisition phase. Specifically, the platform systematically sampled within the design parameter space, generating 200 to 1000 representative design samples. For each sample, the system invoked the equivalent homogenization module 4 to calculate the material's equivalent parameters and utilized the multiphysics coupling solution module 6 to perform high-fidelity finite element simulation, obtaining the corresponding real performance data as labels. Using these high-quality "design-performance" data pairs, the neural network underwent supervised training, with the network weights continuously adjusted through backpropagation until the loss function converged to the preset accuracy.
[0097] The trained machine learning multi-scale agent model module 5 possesses extremely fast inference speed. In actual operation, when the user or the design optimization and process instruction output module 7 inputs any new set of design parameters, the model can accurately predict the corresponding inductance, temperature rise, and other multi-physics responses in an extremely short time (less than one second). This characteristic enables the system to traverse thousands of design schemes in a very short time, providing efficient technical support for subsequent automated optimization design.
[0098] Furthermore, in the process instruction file generated by the design optimization and process instruction output module:
[0099] The spindle speed command ranges from 0.1 rpm to 3000 rpm;
[0100] The coordinate accuracy of the yarn carrier's path coordinate sequence is ±0.02mm;
[0101] The range of weaving tension instructions is 5cN to 200cN;
[0102] The yarn changing control command includes the switching sequence and spatial coordinates of conductive fiber bundles, magnetic fiber bundles, insulating fiber bundles and structural fiber bundles, and the yarn changing time does not exceed 2 seconds.
[0103] Reference Figure 1In this embodiment, the design optimization and process instruction output module 7 is the core component that transforms the simulation design scheme into actual manufacturing actions. The process instruction file generated by this module is a set of data that conforms to the communication protocol of the three-dimensional weaving equipment control system, precisely controlling every physical parameter in the weaving process.
[0104] First, to accommodate the needs of different materials and weaving structures, the process instruction file includes detailed spindle speed instructions, with a speed adjustment range set from 0.1 rpm to 3000 rpm. This wide range allows the equipment to stably weave heavy structural fibers at low speeds while rapidly arranging fine conductive fibers at high speeds. Second, to ensure high-precision reproduction of the device's geometry, the yarn carrier trajectory coordinate sequence included in the instruction file has extremely high spatial positioning accuracy, controlled within ±0.02 mm, ensuring that the yarn carrier's movement trajectory in three-dimensional space strictly follows the path designed in the geometry-material parameterization module 3.
[0105] Furthermore, considering the differences in the physical and mechanical properties of different fiber bundles, the module generates precise weaving tension instructions, with a tension setting range covering 5cN to 200cN. This ensures both weaving tightness and structural morphology while preventing fiber damage. Addressing the complex distribution of conductive, magnetic, insulating, and structural regions within the device, the process instruction file also integrates crucial yarn-changing control instructions. These instructions not only define the specific logic for switching fiber bundle types at different spatial coordinates but also strictly specify the timing of sequential switching between conductive, magnetic, insulating, and structural fiber bundles. To ensure the continuity of the weaving process and production efficiency, the system strictly limits the execution time of all yarn-changing actions to no more than 2 seconds, achieving seamless alternating weaving of multiple materials in three-dimensional space.
[0106] Furthermore, the digital twin online calibration module is configured as follows:
[0107] Calculate the relative deviation δi between the simulation prediction and the measured value;
[0108] when If the variance of the surrogate model's prediction exceeds a preset threshold, it is determined to be a low-confidence region.
[0109] In the low confidence region, incremental learning is used to add the measured samples to the training set and retrain the surrogate model locally until the relative deviation between the updated predicted value and the measured value is reduced to less than 5%.
[0110] Reference Figure 1In this embodiment, the digital twin online calibration module is used to establish a real-time feedback loop between virtual simulation and physical manufacturing, which is a key link in realizing intelligent manufacturing and high-precision design. This module receives measured data collected by physical prototype testing instruments in real time through an external interface module, covering key indicators such as inductance, loss, and temperature, and compares it one by one with the simulation prediction results output by the machine learning multi-scale surrogate model module.
[0111] In the specific implementation process, the module first calculates the relative deviation δi between the simulated predicted value and the measured value of each performance index. In order to strictly determine the effectiveness of the model, the system sets a dual calibration standard: if the calculated relative deviation δi is greater than 5%, or if the prediction variance output by the surrogate model indicates that the uncertainty in the region exceeds a preset threshold, the system determines that the current parameter region is a "low confidence region", which means that the existing mathematical model can no longer accurately reflect the real behavior of the physical entity in this region.
[0112] For identified low-confidence regions, the digital twin online calibration module immediately initiates an incremental learning mechanism. The system dynamically adds measured samples containing actual operating parameters and their performance labels as high-priority incremental data to the original training dataset. Subsequently, the module calls an optimization algorithm to locally retrain the surrogate model, using the new data to correct the model's network weights, significantly improving its fitting accuracy in low-confidence regions. This iterative process continues until the relative deviation between the updated model's predicted values and the measured values converges and stabilizes within 5%. Through this self-evolutionary mechanism, the platform ensures that the digital twin always maps the physical entity with high fidelity, guaranteeing extremely high accuracy in the process instructions generated by the subsequent design optimization and process instruction output modules.
[0113] Furthermore, the parametric descriptions generated by the geometry-material parametric module also include Litz transposition pitch;
[0114] Litz transposition pitch is used to characterize the transposition period of conductive fiber bundles within the conductive region, and Litz transposition pitch is one of the input features of the machine learning multi-scale surrogate model module.
[0115] Reference Figure 1 In this embodiment, the geometry-material parameterization module, when constructing the microscopic parameterization description of the target device, specifically introduces the Litz transposition pitch, a key geometric parameter, to meet the performance requirements of high-frequency applications. The Litz transposition pitch is defined as the axial length required for the insulating strands constituting the conductive fiber bundle to complete one full relative position exchange along the bundle axis within the conductive region. This parameter directly quantifies the degree of torsion within the conductive fiber bundle's internal structure and is a core indicator characterizing the braiding properties of Litz wire.
[0116] In practical applications of 3D braided electromagnetic devices, especially at high frequencies, significant skin and proximity effects occur when current flows through the conductive region, leading to uneven current distribution and a sharp increase in AC resistance and eddy current losses. By precisely setting the Litz transposition pitch in the parametric description, the geometry-material parametric module can accurately simulate the physical influence of the internal microstructure of the conductive fiber bundle on the electromagnetic field distribution. This parameter not only guides subsequent equivalent homogenization calculations but also serves as a crucial input feature, transmitted to the machine learning multi-scale surrogate model module.
[0117] Specifically, when constructing the input feature vector of the surrogate model, in addition to macroscopic parameters such as the volume fraction of conductive fibers, the volume fraction of magnetic fibers, and the braiding angle, the Litz transposition pitch is included. This is because the machine learning model needs to learn the nonlinear mapping relationship between this feature and performance indicators to capture the subtle influence of the transposition cycle on the AC loss and inductance of the device. By using the Litz transposition pitch as an input feature, the surrogate model can more accurately predict the high-frequency performance under different transposition processes, thereby guiding designers to find the optimal balance between reducing losses and process complexity, ensuring the high fidelity of the simulation design platform's prediction of the performance of high-frequency electromagnetic devices.
[0118] Secondly, referring to Figure 6 This invention provides a multiphysics simulation design method for three-dimensional braided electromagnetic devices, comprising the following steps:
[0119] S1: Receive the three-dimensional geometric model, material database, performance constraints, and process constraints of the target three-dimensional braided electromagnetic device;
[0120] In this embodiment, step S1 is specifically executed by the input interface module, and its implementation process covers the entire preprocessing workflow from data access and geometric cleaning to parameter verification. First, the user uploads the 3D geometric model of the target 3D braided electromagnetic device through a graphical user interface. Supported file formats include common CAD formats such as STEP, IGES, or STL. The system's built-in geometry engine immediately performs a topology check on the model, automatically identifying and repairing non-manifold edges, overlapping surfaces, and minor gaps to ensure the model's watertightness and provide an accurate geometric basis for subsequent mesh generation and parametric modeling.
[0121] Secondly, in the material database processing stage, the module adopts a hierarchical parameterized structure. Users not only input traditional material constants but also need to define advanced parameters such as complex conductivity varying with frequency, nonlinear hysteresis loops, and anisotropic thermal conductivity. The system maps and associates these material properties with the functional space of the device, explicitly specifying which properties apply to the conductive, magnetically controlled, insulating, and structural regions. Simultaneously, users need to input a performance index vector containing the target inductance value, quality factor, maximum operating temperature rise, and mechanical stiffness threshold; the system transforms this into the objective function and constraints for subsequent optimization algorithms.
[0122] Finally, the module performs a "process window" verification to address process constraints. Users need to input the physical limits of the 3D knitting equipment, such as the maximum spindle speed, the acceleration range of the yarn carrier, and the tension controller range. During the receiving phase, the system establishes a correlation between design parameters (such as knitting angle) and equipment physical parameters. Virtual simulation is used to initially verify whether the design parameters exceed the kinematic and dynamic limits of the equipment, ensuring that the input process constraints are self-consistent and executable. After data standardization, the module packages and transmits it to the next-level module.
[0123] S2: Discretize the target device into conductive region, magnetocontrollable region, insulating region and structural region, and generate a parameterized description including local braiding angle and fiber bundle arrangement parameters;
[0124] In this embodiment, step S2 is executed by the platform's geometry-material parameterization module, the core of which lies in transforming the continuous macroscopic geometric model into a physically meaningful microscopic parameterized field. Specifically, this module first employs an adaptive voxelization algorithm or tetrahedral mesh generation technique to perform high-precision spatial discretization of the input target device's three-dimensional geometric model, decomposing it into millions of tiny units. Subsequently, based on preset material property mapping rules and electromagnetic function distribution, the system identifies and marks the functional attributes of each tiny unit, thereby discretizing the entire device into four functional zones: a conductive zone responsible for current transmission, a magnetically controlled zone providing magnetic circuit channels, an insulating zone providing electrical insulation, and a structural zone bearing mechanical loads.
[0125] After partitioning and discretizing, this module further performs deep parameterization extraction to generate a parameterized description. For each discretized cell, the system accurately calculates the orientation information of the fiber bundle in three-dimensional space by tracing the yarn path vector, thereby obtaining the local braid angle. This parameter reflects the anisotropic characteristics of the braided structure at different spatial locations, rather than a simple global average. Simultaneously, the system analyzes and records the fiber bundle arrangement parameters, including the cross-sectional shape of the fiber bundle (e.g., ellipticity), the filling density of the monofilaments within the cross-section, the center-to-center distance between adjacent fiber bundles, and the stacking method (e.g., rectangular or hexagonal stacking). Specifically, for the conductive region, the parameterized description also includes the Litz transposition pitch parameter, used to characterize the twisting period of the internal multi-strand conductors. Finally, this rich geometric-material information, including volume fraction, local braid angle, fiber bundle arrangement parameters, and Litz transposition pitch, is integrated into a structured dataset, stored in the form of voxel attribute arrays or finite element mesh node data, providing a high-fidelity geometric and topological foundation for subsequent multi-scale calculations.
[0126] S3: Construct representative volume elements and calculate the equivalent electromagnetic parameters, equivalent thermal parameters, and equivalent mechanical parameters of each functional zone using analytical mixing law and periodic finite element method;
[0127] In this embodiment, step S3 is specifically executed by the equivalent homogenization module (4), the core of which is to transform the complex heterogeneous micro-woven structure into an equivalent continuous medium that can be macroscopically calculated. First, based on the parameterized description generated in step S2, this module constructs a representative volume element (RVE) that can statistically reflect the micro-periodic characteristics of the material. To ensure a balance between computational accuracy and efficiency, the RVE is designed as a cuboid with a side length between 2 mm and 10 mm, and it is strictly stipulated that each side must contain at least two complete weaving cycles. This size constraint ensures that the RVE can fully capture the interlacing topological characteristics of the fiber bundles and satisfy the periodic assumption of the homogenization theory.
[0128] After the Relative Virtual Environment (RVE) is constructed, the module applies periodic boundary conditions to it. Specifically, through master-slave node coupling or periodic mesh pairing techniques, the continuity of displacement, electric potential, magnetic potential, and temperature fields on the relative boundaries of the RVE is forcibly ensured, thereby simulating the real physical environment of the material in an infinitely large periodic structure. Subsequently, the module employs a hybrid calculation strategy combining analytical hybrid laws (such as the Voigt-Reuss averaging method) and the periodic finite element method: uniform electric field, magnetic field, temperature gradient, and mechanical strain boundary loads are applied to the RVE. After obtaining the response field at the microscale through the finite element solver, the volume averaging principle is used to integrate the microscopic physical quantities, thereby accurately calculating the macroscopic equivalent performance tensors of each functional zone. These parameters specifically include the equivalent electrical conductivity tensor reflecting conductivity, the equivalent magnetic permeability tensor reflecting magnetic properties, the equivalent thermal conductivity tensor characterizing thermal conductivity, and the equivalent elasticity matrix describing mechanical behavior, providing high-fidelity material property inputs for subsequent multiphysics coupling solutions.
[0129] S4: Call the machine learning multi-scale proxy model to quickly predict the multiphysics response based on equivalent parameters or macroscopic design parameters;
[0130] In this embodiment, step S4 is executed by the machine learning multi-scale surrogate model module 5, which is a key step in achieving extremely fast design flow. This module internally deploys a high-precision surrogate model built on a deep learning framework, whose algorithm architecture flexibly adopts either a fully connected neural network (DNN) or a graph neural network (GNN). Among these, the graph neural network (GNN) is particularly suitable for processing 3D weaving data, as it can treat fiber bundles as nodes and their contact relationships as edges, thereby efficiently capturing and learning complex irregular spatial topological features.
[0131] In practical implementation, when the system needs to perform performance prediction, the proxy model receives a comprehensive input feature vector. This vector not only includes the macroscopic equivalent material parameters calculated from the equivalent homogenization module 4, but also deeply integrates the microstructural details extracted from the geometry-material parameterization module 3. Specific input features include at least: the fiber volume fraction in the conductive and magnetostrictive regions, the local braiding angle determining material anisotropy, the Litz transposition pitch characterizing high-frequency loss properties, and the actual operating frequency of the device.
[0132] Through forward propagation calculations using a pre-trained neural network, the surrogate model rapidly outputs the multiphysics response of the target device within an extremely short time of less than one second for a single inference. The output response vector comprehensively covers the core performance indicators of the device, specifically including: inductance and AC resistance reflecting electromagnetic characteristics, maximum temperature rise reflecting thermal management capabilities, and mechanical stiffness reflecting structural reliability. This second-level response capability enables the design optimization and process instruction output module 7 to evaluate and screen a massive number of design schemes in an extremely short time, effectively overcoming the technical bottlenecks of traditional high-fidelity finite element solutions, which are time-consuming and computationally resource-intensive.
[0133] S5: Solve the electromagnetic-thermal-mechanical response of the target device based on the output of equivalent parameters or surrogate models;
[0134] In this embodiment, step S5 is specifically executed by the multiphysics coupling solution module 6. This module first obtains the equivalent electrical conductivity tensor, equivalent magnetic permeability tensor, equivalent thermal conductivity tensor, and equivalent elasticity matrix of each functional zone from the equivalent homogenization module 4, serving as the basic material properties for the macroscopic simulation model. Simultaneously, the module can call the prediction results output by the machine learning multi-scale surrogate model module 5 as a reference for the initial field distribution or mesh refinement.
[0135] The solution process employs a two-way sequential coupling strategy. First, Maxwell's equations are solved in the frequency domain to calculate the magnetic field distribution, induced current density, and eddy current loss of the target device under the excitation source, with a focus on obtaining the power loss density distribution in the conductive and magnetostrictive regions. Second, the calculated power loss density is used as a volume heat source and substituted into the three-dimensional heat conduction equations. Using the equivalent thermal conductivity tensor and combined with convective heat dissipation boundary conditions, the transient or steady-state temperature field inside the device is calculated to obtain the specific temperature rise value.
[0136] Subsequently, a mechanical finite element analysis is performed. The calculated temperature field is applied as a thermal load to the structural model, and the thermal deformation and thermal stress caused by thermal expansion are calculated using the equivalent elastic matrix to assess the structural integrity of the device. In the coupled loop, the system monitors the impact of temperature changes on material properties in real time, especially the nonlinear characteristic of conductivity decreasing with increasing temperature. When the relative deviation of material properties caused by temperature changes exceeds a preset threshold (e.g., 5%), the module automatically updates the electromagnetic parameters and re-solves the electromagnetic field. Through this iterative calculation of "electromagnetic-thermal-mechanical" until all physical fields converge, high-precision multi-physics response data such as inductance, loss, temperature rise, and stress distribution are finally output.
[0137] S6: Compare the simulation prediction results with the actual measured data of the physical prototype, calculate the relative deviation, and when the relative deviation exceeds the preset threshold, use the actual measured data to perform incremental training on the surrogate model.
[0138] In this embodiment, step S6 is specifically executed by the digital twin online calibration module 9. Its core purpose is to establish a real-time feedback closed loop between the virtual model and the physical entity, ensuring that the simulation platform can continuously maintain high fidelity. In specific implementation, this module first obtains the actual operating data of the target device measured by physical testing instruments through the external interface module 10. These measured data cover multiple physical field indicators such as inductance, AC resistance, key point temperature, and mechanical deformation.
[0139] Subsequently, the system compares the measured data point by point with the simulation predictions output by the machine learning multi-scale surrogate model module 5 under the same working conditions, and strictly calculates the relative deviation δi between the two according to the formula δi=|Pi-Mi| / |Mi|×100%. The module sets strict judgment logic: if the calculated relative deviation δi exceeds 5%, or the prediction variance output by the surrogate model is higher than a preset threshold, the system determines that the current parameter region is in the "low confidence region" of the model, indicating that the existing mathematical model can no longer accurately reflect the physical characteristics of the region.
[0140] At this point, the system automatically initiates an incremental learning mechanism for correction. This module dynamically appends the measured samples, which contain the input design parameters and actual performance labels, to the original training dataset as high-priority data. Next, the module calls an optimization algorithm to perform targeted local retraining on the machine learning multi-scale surrogate model module 5, using the newly acquired physical ground truth to correct the network weights, forcing the model to approximate the measured data in low-confidence regions. This iterative process continues until the relative deviation between the updated model's predicted values and the measured values converges and stabilizes within 5%, thereby achieving self-evolution and accuracy improvement of the digital twin.
[0141] S7: Adjust the design parameters according to the performance index constraints and generate a process instruction file that conforms to the communication protocol of the three-dimensional knitting equipment control system. The process instruction file includes spindle speed instructions, yarn carrier trajectory coordinate sequence, knitting tension instructions and yarn changing control instructions.
[0142] In this embodiment, step S7 is specifically executed by the design optimization and process instruction output module 7. This module first uses a multi-objective optimization algorithm to iteratively adjust the design parameters based on the solution results of the previous steps and preset performance constraints (such as inductance, loss, temperature rise, etc.) until the optimal design scheme that satisfies all constraints is obtained. Subsequently, the module transforms the abstract geometric and material parameters into specific executable machine control code, generating a process instruction file that conforms to the communication protocol of the three-dimensional knitting equipment control system.
[0143] This instruction file contains four core process control instructions to ensure the accurate implementation of the design: First, it generates spindle speed instructions, precisely setting the speed range between 0.1 rpm and 3000 rpm according to the process characteristics of different materials and weaving rhythm requirements, enabling flexible switching between low-speed stable start-up and high-speed efficient weaving; Second, it calculates and outputs the coordinate sequence of the yarn carrier's path, generating the motion coordinate points of the yarn carrier in three-dimensional space through high-precision discretization of the design path, ensuring that the coordinate accuracy is controlled within ±0.02 mm, thereby accurately constructing the complex space of the device. The geometric structure is as follows: Third, the weaving tension command is set, and the weaving tension is dynamically adjusted within the range of 5cN to 200cN based on the physical and mechanical properties of the fiber bundle. This ensures both the tightness and morphological stability of the weaving structure, while preventing excessive tension from damaging the fibers. Fourth, an intelligent yarn-changing control command is generated. This command defines in detail the material switching logic between the conductive area, magnetic control area, insulation area, and structural area. It specifies the switching sequence of different types of fiber bundles at specific spatial coordinates and strictly constrains the execution time of all yarn-changing actions to no more than 2 seconds, thereby achieving seamless and efficient weaving of multi-material composite structures.
[0144] Furthermore, step S5, the multiphysics coupling solution, includes:
[0145] Solve Maxwell's equations in the frequency domain to obtain the current density distribution, magnetic field distribution, and loss of the device at the operating frequency.
[0146] Electromagnetic loss is used as a bulk heat source and input into the steady-state or transient heat conduction equation to calculate the device temperature rise.
[0147] When the change in material properties caused by temperature rise exceeds 5%, update the conductivity and permeability and re-solve the electromagnetic field until the change in performance indicators converges.
[0148] In this embodiment, step S5 utilizes the multiphysics coupling solution module 6 to achieve electromagnetic-thermal bidirectional strong coupling calculation, accurately capturing the actual physical behavior of the device under high-power conditions. Specifically, the module first numerically solves Maxwell's equations in the frequency domain based on the initial equivalent conductivity and permeability provided by the equivalent homogenization module 4. This process not only quickly obtains the current density vector distribution and magnetic flux density distribution of the target device at its operating frequency, but also further calculates the Joule heat loss in the conductive region and the eddy current loss in the magnetized region.
[0149] Subsequently, the module interpolates and maps the calculated total electromagnetic loss data to each node of the finite element mesh, substituting it as a volume heat source term into the steady-state or transient heat conduction equations. Combined with preset convective heat dissipation boundary conditions, the solver calculates the temperature field distribution inside the device, obtaining the temperature rise values in key regions. Specifically, to address the significant impact of temperature on material properties, the system introduces an iterative mechanism based on physical property feedback: the system evaluates in real-time the nonlinear changes in material properties caused by temperature rise, focusing on detecting the deviation of conductivity and permeability from their current values. When the detected change in these material properties due to temperature rise exceeds 5%, the system determines that the thermal effect has significantly altered the electromagnetic environment. At this point, it automatically updates the conductivity and permeability parameters using the latest temperature data and triggers a cyclic process of resolving the electromagnetic field. This iterative process continues until the performance index changes calculated in the two iterations converge to within a preset error range, thereby ensuring high fidelity in the final multiphysics response output.
[0150] Furthermore, step S7, generating the process instruction file, includes:
[0151] Establish a mapping relationship between design parameters and process control parameters. The mapping relationship will convert the optimized braiding angle, fiber volume fraction and device size into the spindle speed, yarn trajectory coordinate sequence and braiding tension setting value of the three-dimensional braiding equipment.
[0152] Based on the partitioned structure of the device, control commands for yarn changing of the yarn carrier at different spatial positions are generated to achieve sequential switching of conductive fiber bundles, magnetic fiber bundles, insulating fiber bundles and structural fiber bundles.
[0153] In this embodiment, the refinement process of generating the process instruction file in step S7 is executed by the design optimization and process instruction output module 7. Its core lies in establishing a high-precision inverse kinematic mapping mechanism to transform abstract microscopic design parameters into physical actions that can be executed by the three-dimensional knitting equipment. Specifically, this module first embeds the kinematic model of the equipment and constructs a nonlinear mapping relationship from design parameters to process control parameters.
[0154] During the mapping process, the module converts the optimized local weaving angle into spindle speed commands for each axis of the machine. By adjusting the speed ratio of different groups of spindles, the convergence angle of the fiber bundles at the interlacing point is precisely controlled. The fiber volume fraction is mapped into weaving tension commands, and the tightness of the fiber bundles is dynamically adjusted through the tension servo system to achieve the target filling density. At the same time, the overall geometric dimensions of the device are discretized into a time-dimensional tracking trajectory coordinate sequence. This sequence defines the precise position of the yarn carrier in the machine's Cartesian coordinate system at microsecond intervals, ensuring that the macroscopic dimensional error of the formed parts is controlled within the millimeter level.
[0155] Furthermore, to address the complex multi-material partitioned structure of the device, the module implements a space-triggered yarn-changing control strategy. Based on the discretized mesh model, the system identifies the geometric boundaries and spatial topological relationships of the conductive, magnetically controlled, insulating, and structural regions. When the yarn carrier's trajectory crosses the virtual boundaries of these functional partitions, the module automatically generates a yarn-changing control command. This command precisely schedules the mechanical actuators by identifying the correspondence between the yarn carrier ID and the target material's yarn storage location, achieving sequential switching and seamless connection of conductive, magnetic, insulating, and structural fiber bundles at specific spatial coordinates. This ensures that the physical partitioning of the final manufactured product strictly matches the simulation design model.
[0156] This embodiment also provides a computer device applicable to the above-described method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the above-described method as proposed in the above embodiments.
[0157] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0158] This embodiment also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0159] Example 1: Simulation Design of Three-Dimensional Braided Power Inductor
[0160] like Figures 1 to 5 As shown, the target device is a hollow solenoid-shaped three-dimensional braided power inductor with the following performance constraints: at a working frequency of 1MHz, the inductance L = 10μH ± 10%, the DC resistance does not exceed 30mΩ, the AC resistance does not exceed 50mΩ, the steady-state temperature rise does not exceed 20K, the outer diameter does not exceed 10mm, and the length is 25mm.
[0161] In step S1, the input interface module 2 receives the STL geometric model, material database, and performance constraints of the power inductor. In the material database, the conductive fiber bundles are silver-plated copper fiber bundles with a single filament diameter of 30 μm, 19 filaments per bundle, and a 3 μm thick polyimide insulating layer on the bundle surface; the magnetic fiber bundles are amorphous soft magnetic composite fiber bundles with an outer diameter of 80 μm and a relative permeability of 200; the insulating fibers are polyimide fibers with a single filament diameter of 20 μm; and the structural fibers are aramid fibers with a single filament diameter of 25 μm.
[0162] In step S2, the geometry-material parameterization module 3 discretizes the inductor model into a conductive region 11, a magnetically controlled region 12, an insulating region 13, and a structural region 14. The volume fractions are set as follows: conductive region 45%, magnetically controlled region 25%, insulating region 20%, and structural region 10%; local braiding angle... The Litz transposition pitch is p=15mm, the outer diameter of the device is 8mm, the inner diameter is 4mm, and the length is 25mm.
[0163] In step S3, the equivalent homogenization module 4 constructs a representative volume element of 2mm × 2mm × 2mm (e.g., Figure 4 As shown in the figure, the equivalent conductivity of the conductive region is calculated using periodic boundary conditions and the finite element method. The equivalent relative permeability of the magnetized region is 45, the equivalent thermal conductivity of the insulating region is 0.25 W / (m·K), and the equivalent elastic modulus of the structural region is 65 GPa.
[0164] In step S4, the machine learning multi-scale surrogate model module 5 takes the volume fraction of conductive fibers, the volume fraction of magnetic fibers, the braiding angle θ, the Litz transposition pitch p, and the operating frequency f as inputs, and the inductance L, AC resistance, and temperature rise ΔT as outputs to generate 300 sets of high-fidelity training samples; it then trains a fully connected neural network with a hidden layer structure of [128, 256, 256, 128] for 500 rounds, and the validation set determination coefficient is used. The mean absolute percentage error (MAPE) is 4.2%.
[0165] In steps S5 to S7, the multiphysics coupling solution module 6 and the design optimization and process instruction output module 7 work together, employing the NSGA-II multi-objective optimization algorithm (population size 50, generation number 100, crossover probability 0.8, mutation probability 0.1) for multi-objective optimization. The optimized design scheme is: conductive fiber volume fraction 47%, magnetic fiber volume fraction 23%, and braiding angle... Litz transposition pitch p = 14 mm. Surrogate model predictions: inductance L = 10.3 μH, AC resistance 44 mΩ, temperature rise 18 K.
[0166] Finally, the optimized design scheme was output as a JSON format process file through the external interface module 10 and imported into the 3D weaving equipment for manufacturing. The measured inductance of the prototype was L=10.1μH, AC resistance was 46mΩ, and temperature rise was 19K, with an average deviation of no more than 5% from the simulation results. The measured data was fed back to the digital twin online calibration module 9. After one incremental training, the prediction error of the surrogate model for this type of power inductor was reduced to within ±3%.
[0167] Example 2: Simulation Design of 3D Braided Motor Stator
[0168] The target device is a C-shaped hollow cup motor stator, with the following performance constraints: output torque not less than 0.23 N·m at a rated current of 5A, stator outer diameter 60 mm, length 50 mm, and bending angle [not specified]. After bending 100,000 times, the change in winding resistance is less than 3%.
[0169] In steps S1 to S2, the input interface module 2 receives the parameterized model of the C-shaped stator, and the geometry-material parameterization module 3 sets the conductive region to consist of 7 bundles of silver-plated copper fibers arranged according to the Litz principle, with each bundle containing 21 monofilaments of 40 μm in diameter and a volume fraction of 40%; the magnetron region consists of amorphous / nanocrystalline soft magnetic composite fiber bundles with a volume fraction of 20%; the insulating region consists of aramid / polyimide composite fibers with a volume fraction of 25%; the structural region consists of aramid fibers with a volume fraction of 15%; and the local braiding angle... The stator wall thickness is 1.5mm.
[0170] In step S3, the equivalent homogenization module 4 calculates the following: the equivalent relative permeability of the magnetized region is 55, the equivalent elastic modulus of the structural region is 58 GPa, and the equivalent conductivity of the conductive region is 3.2 × 10⁻⁶. 7 S / m. In step S4, the machine learning multi-scale surrogate model module 5 takes the braiding angle θ, magnetic fiber volume fraction, stator wall thickness, and current as inputs, and outputs the output torque, winding temperature rise, and bending stiffness as outputs, generating 400 sets of training samples; the neural network structure is [128, 256, 256, 128], trained for 500 rounds, and the validation set determination coefficient is... The mean absolute percentage error (MAPE) is 4.8%.
[0171] In steps S5 to S7, the multiphysics coupling solution module 6 predicts through electromagnetic-force coupling analysis: at a current of 5A and a braiding angle... Under conditions of 22% magnetic fiber volume fraction and 1.6mm stator wall thickness, the output torque is 0.25 N·m, the winding temperature rise is 22K, and the bending stiffness is 12 N·m / rad. The optimized output torque deviation is less than ±5%. Finally, the optimized design scheme is output to a 3D weaving machine for 3D weaving manufacturing. The measured prototype has an output torque of 0.24 N·m at 5A current, and the winding resistance change after 100,000 bends is 2.1%, meeting the design specifications. After the measured data is fed back by the digital twin online calibration module 9, the equivalent elastic modulus of aramid fiber in the structural region of the material database is corrected from 58 GPa to 55 GPa, and the prediction error MAPE of the surrogate model for bending stiffness is reduced from 4.8% to 3.5%.
[0172] Example 3: Data Interface and Online Calibration Application
[0173] like Figure 1 and Figure 3As shown, the external interface module 10 achieves collaboration with the 3D weaving equipment control system and impedance matching system through standardized data exchange. After the design optimization is completed, a JSON format process control instruction file is output to the 3D weaving equipment control system, including spindle speed instructions (e.g., conductive fiber bundle spindle 150rpm, magnetic fiber bundle spindle 120rpm, insulating fiber bundle spindle 180rpm, structural fiber bundle spindle 160rpm), yarn trajectory coordinate sequence (sampling period 1ms, coordinate accuracy ±0.02mm), weaving tension instructions (conductive fiber bundle 30cN, magnetic fiber bundle 25cN, insulating fiber bundle 20cN, structural fiber bundle 35cN), and yarn changing sequence (e.g., switching from the conductive area to the magnetic control area at a length of 12mm, yarn changing time 1.5s); the 3D weaving equipment executes the actual 3D weaving process according to the above instructions.
[0174] Output the target impedance characteristic curve to the impedance matching system (e.g., in the operating frequency range of 100kHz to 10MHz, the port impedance increases inductively with increasing frequency, and the phase approaches...). The platform also imports the spatial arrangement parameters of the conductive region 11, magnetically controlled region 12, insulating region 13, and structural region 14 from the four-functional partition structure description of the 3D braided electromagnetic device as the initial input of the geometry-material parameterization module 3. The simulation optimization results do not change the physical structural boundaries of the four functional partitions, but only generate specific dimensions and process parameters that meet the performance indicators.
[0175] Digital twin online calibration example: After continuously producing 50 power inductors, the digital twin online calibration module 9 statistically obtained the average deviation of the inductance prediction as follows: The average deviation of AC resistance prediction was +6.5%. The module automatically triggered incremental training, adding 50 sets of measured samples to the training set. After local retraining of the surrogate model, the inductance prediction deviation was reduced to -1.8%, and the AC resistance prediction deviation was reduced to +2.3%, achieving the online calibration target.
[0176] In summary, this invention brings several significant benefits by constructing an integrated digital design and manufacturing platform for three-dimensional braided electromagnetic devices. Firstly, it achieves a precise mapping between microstructure and macroscopic performance. By constructing representative volumetric units and using electromagnetic-thermal-mechanical multiphysics coupling solutions, it can accurately capture the complex anisotropic characteristics of braided composite materials and the temperature rise effect caused by electromagnetic losses. This ensures the electromagnetic performance and structural reliability of the device under high-frequency, high-power conditions, effectively solving the performance prediction distortion problem caused by neglecting microstructural anisotropy in traditional design methods.
[0177] Secondly, this invention significantly improves design optimization efficiency. By introducing a machine learning multi-scale surrogate model, the complex finite element calculation process that traditionally takes several hours is transformed into a rapid prediction at the millisecond level. This makes it possible to perform large-scale multi-objective global optimization while meeting process constraints, greatly shortening the product development cycle and reducing trial and error costs.
[0178] Furthermore, this invention possesses a high degree of manufacturability. The system establishes a direct mapping relationship between design parameters and process parameters of three-dimensional weaving equipment. In particular, the intelligent yarn changing control and trajectory planning for multi-material partitioning ensure the precise molding of complex functional structures and completely eliminate the manufacturing gap between the design model and the physical prototype.
[0179] Finally, this invention possesses self-evolution capabilities. Through a digital twin closed-loop feedback mechanism, the system can incrementally train and correct the proxy model using measured data from physical prototypes. As production data accumulates, the model's prediction accuracy continuously improves, ensuring a high degree of consistency between design theory and engineering practice, and providing data-driven core technology support for intelligent manufacturing.
[0180] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A multiphysics simulation design platform for three-dimensional braided electromagnetic devices, characterized in that, include: Multiphysics simulation design platform body (1); The input interface module (2) is used to receive the three-dimensional geometric model, material database, performance index constraints and process constraints of the target three-dimensional braided electromagnetic device; the geometry-material parameterization module (3) is connected to the input interface module (2) and is used to discretize the target device into four functional regions: conductive region (11), magnetic control region (12), insulating region (13) and structural region (14), and generate a parameterized description containing the volume fraction of each region, local braiding angle and fiber bundle arrangement parameters. The parameterized description is stored in the form of voxels or finite element meshes; the equivalent homogenization module (4) is used to construct representative volume elements (RVE) according to the parameterized description, and calculate the equivalent electromagnetic parameters, equivalent thermal parameters and equivalent mechanical parameters of each functional region by applying periodic boundary conditions and combining analytical mixing law and periodic finite element method. The machine learning multi-scale surrogate model module (5) is used to establish a surrogate model from macroscopic design parameters to multi-physics response. The surrogate model is trained with high-fidelity simulation data from the equivalent homogenization module (4) and is used to predict device performance indicators within seconds. The multi-physics coupling solution module (6) is used to solve the response of the target device under electromagnetic-thermal-mechanical multi-physics coupling conditions based on the equivalent parameters output by the equivalent homogenization module (4) or the prediction results of the machine learning multi-scale surrogate model module (5). The design optimization and process instruction output module (7) is used to adjust the design parameters according to the performance indicator constraints and solution results, and generate a process instruction file that conforms to the communication protocol of the three-dimensional weaving equipment control system. The process instruction file includes at least the spindle speed instruction, the yarn carrier track coordinate sequence, the weaving tension instruction, and the yarn changing control instruction. The database and visualization module (8) is connected to all the aforementioned modules and is used to uniformly store and manage geometric models, material properties, process parameters, simulation results, process instructions and historical measured data in the design process. It also provides a real-time visual interactive interface for three-dimensional geometric model reconstruction, microstructure display, multi-physics field distribution cloud map rendering and optimization process tracking. The digital twin online calibration module (9) is used to compare the measured data of the physical prototype with the simulation prediction results, calculate the relative deviation, and when the relative deviation exceeds the preset threshold, use the measured data to incrementally train the proxy model or correct the material database through parameter inversion. The external interface module (10) is used to output the process instruction file to the three-dimensional knitting equipment control system and impedance matching system.
2. The platform according to claim 1, characterized in that, The equivalent homogenization module (4) is configured to perform the following steps: constructing a cuboid with a side length of 2 mm to 10 mm as a representative volume element, wherein each side of the representative volume element contains at least two complete weaving cycles; applying periodic boundary conditions to the representative volume element, and ensuring the continuity of the displacement field, electric field, magnetic field and temperature field on the boundary through master-slave node coupling or periodic mesh pairing; applying a uniform electric field, magnetic field, temperature gradient or mechanical strain to the representative volume element, and obtaining the equivalent electrical conductivity tensor, equivalent magnetic permeability tensor, equivalent thermal conductivity tensor and equivalent elastic matrix through volume averaging calculation.
3. The platform according to claim 1, characterized in that, The machine learning multi-scale surrogate model module (5) is implemented using a fully connected neural network or a graph neural network; the input feature vector of the surrogate model includes at least the volume fraction of conductive fibers, the volume fraction of magnetic fibers, the braiding angle, and the operating frequency; the output response vector of the surrogate model includes at least the inductance, AC resistance, temperature rise, and mechanical stiffness index; the surrogate model is trained with 200 to 1000 sets of training samples generated by high-fidelity finite element analysis, and the single inference time is less than 1 second.
4. The platform according to claim 1, characterized in that, In the process instruction file generated by the design optimization and process instruction output module (7): the spindle speed instruction ranges from 0.1 rpm to 3000 rpm; the coordinate accuracy of the yarn carrier's trajectory coordinate sequence is ±0.02 mm; the weaving tension instruction ranges from 5 cN to 200 cN; the yarn changing control instruction includes the switching sequence and spatial position coordinates of conductive fiber bundles, magnetic fiber bundles, insulating fiber bundles and structural fiber bundles, and the yarn changing time does not exceed 2 seconds.
5. The platform according to claim 1, characterized in that, The digital twin online calibration module (9) is configured to: calculate the relative deviation between the simulation prediction value and the measured value. ; when If the variance of the surrogate model's prediction exceeds a preset threshold, it is identified as a low-confidence region. Within the low-confidence region, incremental learning is used to add the actual samples to the training set and locally retrain the surrogate model until the relative deviation between the updated prediction value and the actual value is reduced to within 5%.
6. The platform according to claim 1, characterized in that, The parameterized description generated by the geometry-material parameterization module (3) also includes the Litz transposition pitch; the Litz transposition pitch is used to characterize the transposition period of the conductive fiber bundle in the conductive region (11), and the Litz transposition pitch is one of the input features of the machine learning multi-scale surrogate model module (5).
7. A multiphysics simulation design method for three-dimensional braided electromagnetic devices, characterized in that, Includes the following steps: S1: Receive the three-dimensional geometric model, material database, performance constraints, and process constraints of the target three-dimensional braided electromagnetic device; S2: Discretize the target device into a conductive region (11), a magnetized region (12), an insulating region (13), and a structural region (14), and generate a parameterized description including local braiding angles and fiber bundle arrangement parameters; S3: Construct a representative volume element (RVE), and calculate the equivalent electromagnetic parameters, equivalent thermal parameters, and equivalent mechanical parameters of each functional zone using analytical mixing law and periodic finite element method; S4: Call a machine learning multi-scale surrogate model to quickly predict the multiphysics response based on the equivalent parameters or macroscopic design parameters; S5: Solve the electromagnetic-thermal-mechanical response of the target device based on the equivalent parameters or surrogate model output; S6: Compare the simulation prediction results with the physical prototype measured data, calculate the relative deviation, and when the relative deviation exceeds the preset threshold, use the measured data to perform incremental training on the surrogate model; S7: Adjust the design parameters according to the performance index constraints and generate a process instruction file that conforms to the communication protocol of the three-dimensional knitting equipment control system. The process instruction file includes spindle speed instructions, yarn carrier trajectory coordinate sequence, knitting tension instructions, and yarn changing control instructions.
8. The method according to claim 7, characterized in that, Step S5, the multiphysics coupling solution, includes: solving Maxwell's equations in the frequency domain to obtain the current density distribution, magnetic field distribution, and losses of the device at the operating frequency; using electromagnetic losses as a bulk heat source and inputting them into the steady-state or transient heat conduction equations to calculate the device temperature rise; when the material performance change caused by the temperature rise exceeds 5%, updating the conductivity and permeability and resolving the electromagnetic field until the performance index changes converge.
9. The method according to claim 7, characterized in that, The process instruction file generation in step S7 includes: establishing a mapping relationship between design parameters and process control parameters, wherein the mapping relationship converts the optimized braiding angle, fiber volume fraction and device size into the spindle speed, yarn trajectory coordinate sequence and braiding tension setting value of the three-dimensional braiding equipment; and generating yarn changing control instructions for the yarn carrier at different spatial positions according to the partition structure of the device, so as to realize the sequential switching of conductive fiber bundles, magnetic fiber bundles, insulating fiber bundles and structural fiber bundles.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multiphysics simulation design method for three-dimensional braided electromagnetic devices as described in any one of claims 7 to 9.
