AI-assisted multi-physical field coupling motor optimization design method and system

Abstract

This invention discloses an AI-assisted multiphysics coupled motor optimization design method and system. The method includes: S1 grouping motor design parameters, generating a parameter set using Latin hypercube sampling with geometric constraints, and obtaining training data through finite element simulation; S2 training an AI proxy model based on the training data, combining the AI ​​proxy model with a lumped parameter thermal network (LPTN) to construct an AI-coupled proxy model; the AI ​​proxy model outputs electromagnetic losses as the LPTN heat source, and the LPTN outputs temperature to inversely correct the electromagnetic and stress field parameters, iterating until convergence to form a bidirectional coupling between data-driven and physical mechanisms; S3 replacing finite element simulation with the trained AI-coupled proxy model, using a multi-objective optimization algorithm for global optimization, and outputting the Pareto optimal solution set. This invention significantly shortens the multiphysics simulation time, improves parameter optimization efficiency, ensures prediction errors meet engineering requirements, and realizes full-process digitalization and automation of motor design, applicable to efficient and high-precision optimization design of various motors.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of electrical engineering and artificial intelligence, specifically an AI-assisted multiphysics coupled motor optimization design method and system. Background Technology

[0002] As a core power component in industrial production, transportation, and energy equipment, the performance of electric motors directly determines the operating efficiency and reliability of the equipment. Modern motor design must consider electromagnetic performance, temperature rise performance, and structural stress performance, and these three factors have a strong dynamic coupling relationship: electromagnetic loss is the core heat source of the temperature field, and the temperature rise of key motor components will change the winding resistivity and permanent magnet performance, which in turn affects the electromagnetic output; at the same time, the thermal expansion caused by the temperature rise of the motor rotor and magnets will lead to changes in structural stress, and excessive stress may cause component failure, forming a closed-loop coupling effect of "electromagnetism-thermal-mechanical". This complex coupling relationship across physical fields poses a huge challenge to traditional design methods.

[0003] Chinese invention patent CN113408160B proposes a motor parameter design method based on multi-objective optimization. This method optimizes motor parameters by establishing a motor parameter prediction model and combining it with a stochastic optimization algorithm. Its advantage lies in using finite element simulation results to construct the optimization objective function, achieving a balance between computational accuracy and efficiency to some extent. However, the stochastic optimization algorithm has limited global optimization capability in complex search spaces, making it prone to getting trapped in local optima during the optimization process, affecting the optimization effect of the final design. Furthermore, although this method can shorten computation time, it lacks a dynamic adjustment mechanism for changes in the optimization process and operating conditions, making it difficult to meet the flexibility and adaptability requirements of multi-objective optimization design under different operating conditions.

[0004] Chinese invention patent application CN119578180A discloses a design method for permanent magnet synchronous motors based on a combination of deep learning models and multi-objective optimization algorithms. This method constructs a motor performance prediction model combining LSTM and Transformer, and integrates genetic algorithms and particle swarm optimization algorithms to achieve multi-objective optimization of motor design parameters. The advantage of this method lies in reducing the computational load of finite element simulation by utilizing a neural network surrogate model, thus improving optimization efficiency to some extent. However, this method is sensitive to the quality of training data and model structure parameters, and is prone to decreased prediction accuracy in high-dimensional, highly nonlinear motor design scenarios. Furthermore, its multi-physics coupling mainly relies on the simulation verification stage, lacking effective modeling of the dynamic feedback relationships between electromagnetic, thermal, and mechanical fields during the optimization process, which limits the stability and adaptability of multi-objective optimization under complex operating conditions.

[0005] Although existing technologies have made some progress in multi-objective optimization design and surrogate model application for motors, the following shortcomings still exist in complex working conditions and high-dimensional design scenarios:

[0006] First, traditional sampling methods suffer from numerous invalid combinations and low efficiency in high-dimensional parameter space sampling.

[0007] Second, multiphysics coupling simulation is time-consuming and has low accuracy, and cannot reflect the problem of bidirectional coupling between physical fields;

[0008] Third, traditional optimization algorithms are prone to getting trapped in local optima and lack the ability to balance global performance.

[0009] Fourth, design tools are scattered, and there is no integrated platform to support full-process automation.

[0010] Fifth, the optimization design process relies heavily on finite element simulation, which is time-consuming and resource-intensive, making it difficult to meet the requirements of rapid response and efficient design in engineering applications.

[0011] Sixth, existing parameter sampling and surrogate modeling methods suffer from uneven sample coverage, data redundancy, or insufficient model expressive power in high-dimensional nonlinear design spaces, which affect the prediction accuracy and generalization performance of the surrogate model. Summary of the Invention

[0012] To address the aforementioned technical problems, this invention proposes an AI-assisted multiphysics coupled motor optimization design method and system. This method can significantly shorten the time required for multiphysics simulation, improve parameter optimization efficiency, and control prediction errors within an acceptable range for engineering applications, providing a general technical solution for efficient and high-precision motor optimization design.

[0013] To achieve the above-mentioned technical objectives, the present invention employs the following technical means:

[0014] An AI-assisted multiphysics coupled motor optimization design method includes the following steps:

[0015] S1. Parameter sampling: The motor design parameters are grouped, and a parameter set is generated by Latin hypercube sampling that integrates geometric constraints. Training data is obtained through finite element simulation.

[0016] S2. AI Multiphysics Coupling Modeling: Based on the training data obtained in step S1, train the AI ​​proxy model, combine the AI ​​proxy model with the lumped parameter thermal network LPTN to construct the AI ​​coupled proxy model; the AI ​​proxy model outputs electromagnetic loss as the heat source of LPTN, and the LPTN outputs temperature to correct the electromagnetic field and stress field parameters in reverse, iterating until convergence, forming a two-way coupling between data-driven and physical mechanism.

[0017] S3. AI-assisted multi-objective optimization: The trained AI coupled agent model replaces the finite element simulation, and a multi-objective optimization algorithm is used to perform global optimization and output the Pareto optimal solution set.

[0018] Furthermore, the parameter grouping in step S1 includes a motor setting group, a rotor structure group, and a stator winding group; the minimum parameter interval control is used within the group, and the groups are randomly cross-combined; the geometric constraints include the rotor outer diameter being greater than the inner diameter, the slot fill factor not exceeding the preset upper limit, and the magnet size meeting the minimum processing thickness.

[0019] Furthermore, the AI ​​agent model described in step S2 is an improved Transformer neural network, which sets attention encoders for the stator, rotor, windings and motor for electromagnetic fields, and sets attention encoders for the rotor and motor for stress fields.

[0020] Furthermore, the encoder and decoder of the improved Transformer neural network both include a multi-head attention layer, a feedforward layer, residual connections, and layer normalization.

[0021] Furthermore, the temperature reverse correction in step S2 includes: correcting the resistivity of the winding temperature, correcting the BH curve of the magnet temperature, and correcting the thermal expansion coefficient of the rotor temperature; the iterative convergence condition is that the temperature difference between two adjacent iterations is less than 5°C or the number of iterations reaches 10.

[0022] Furthermore, the multi-objective optimization algorithm in step S3 is the NSGA-II algorithm, and the optimization objectives are to maximize the mean torque, minimize torque fluctuation, and minimize electromagnetic loss.

[0023] Furthermore, the optimized constraints include ensuring that the maximum rotor stress meets the stress safety margin, the winding temperature does not exceed the insulation upper limit, and the magnet temperature does not exceed the demagnetization threshold.

[0024] Furthermore, step S3 includes a testing and verification step, which compares the prediction results of the AI ​​coupled agent model with the finite element simulation to verify the coupling accuracy and optimization effect.

[0025] Furthermore, the AI-coupled agent model enables millisecond-level performance prediction and replaces finite element simulation during the optimization process to reduce computation time.

[0026] This invention also discloses an AI-assisted multiphysics coupled motor optimization design system, comprising:

[0027] The system comprises a parameter sampling module, an AI-coupled agent model construction module, an AI-assisted multi-objective optimization module, and a testing and verification module, and is used to execute the method.

[0028] Beneficial effects: By implementing the technical solution of the present invention, the following beneficial effects can be obtained:

[0029] 1. The grouped Latin hypercube sampling method with integrated geometric constraints can effectively eliminate invalid parameter combinations in motor design, improve the sampling efficiency and sample quality of high-dimensional parameter space, significantly reduce the number of finite element simulations, and reduce simulation resource consumption.

[0030] 2. Construct an AI-coupled proxy model that combines an improved Transformer neural network with a lumped parameter thermal network LPTN to achieve bidirectional closed-loop coupling of electromagnetic-thermal-mechanical multi-physics fields. This model realistically reproduces the dynamic feedback relationship between various physical fields during motor operation, which is closer to the actual physical process than traditional unidirectional coupling and improves the reliability of the simulation.

[0031] 3. By replacing traditional finite element simulation with an AI-coupled proxy model, we can achieve millisecond-level rapid prediction of motor performance, significantly shorten the calculation time of multi-physics coupling, and meet the design requirements of rapid iteration and parallel evaluation of multiple schemes in engineering.

[0032] 4. By combining the AI-coupled proxy model with the NSGA-II multi-objective optimization algorithm, local optima can be effectively avoided in high-dimensional complex design spaces. The model can output a uniformly distributed Pareto optimal solution set, providing an intuitive and reliable engineering solution for motor multi-performance index trade-off design.

[0033] 5. This invention fully covers the entire process of parameter sampling, AI-coupled modeling, AI-assisted multi-objective optimization, and testing and verification, forming an integrated digital and intelligent design solution. It reduces the reliance on manual design experience and high-performance computing hardware, realizes the automation and platformization of motor optimization design, and has strong versatility and scalability.

[0034] 6. During the optimization process, engineering constraints such as stress safety margin, temperature upper limit, and demagnetization threshold are introduced to ensure that the optimization results directly meet the requirements of national standards and actual working conditions, reduce post-processing corrections, and improve the efficiency of design implementation. Attached Figure Description

[0035] Figure 1 Comparison chart of motor design optimization strategies;

[0036] Figure 2 Flowchart of the overall scheme for AI-assisted multiphysics coupled motor optimization design;

[0037] Figure 3 Flowchart of global sampling in parameter space;

[0038] Figure 4 This is a schematic diagram of the structure of a multi-physics coupled neural network surrogate model;

[0039] Figure 5 This is a schematic diagram illustrating the principle of bidirectional coupling of electromagnetic-thermal-mechanical multiphysics fields.

[0040] Figure 6 A flowchart for AI-assisted multiphysics coupling optimization design;

[0041] Figure 7 A two-dimensional electromagnetic field simulation model of a 54-slot, 6-pole motor;

[0042] Figure 8 This is a simulation model diagram of the motor rotor structure's mechanical properties.

[0043] Figure 9 This is a diagram illustrating the training results of an AI agent model.

[0044] Figure 10 A schematic diagram of the Pareto optimal frontier for multi-objective optimization;

[0045] Figure 11 This is a schematic diagram of the optimized motor structure;

[0046] Figure 12 This is a schematic diagram showing the model convergence and performance verification results. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Before implementing the platform process, it is necessary to prepare a GPU server that supports data acquisition and surrogate model training, a computer that supports multi-objective optimization, and install electromagnetic field simulation software ANSYS Maxwell and stress field simulation software ANSYS Mechanical on the server. On the optimization computer, install Python and build a multiphysics coupling model based on the PyTorch framework, and develop a multi-objective optimization algorithm based on the surrogate model. Determine the variation range of relevant motor parameters, such as speed, electrical density, slot fill factor, rotor inner and outer diameters, magnet dimensions, stator and winding dimensions, etc.

[0049] S1. Parameter Sampling

[0050] like Figure 3 As shown, based on the physical structure and functional modules of the motor, the original high-dimensional geometric parameter set of the motor is reasonably grouped into several subspaces with clear boundaries: motor setting parameter group, rotor structure parameter group, stator and winding structure parameter group. Each parameter group exhibits the characteristics of strong intra-group coupling and weak inter-group coupling.

[0051] S1.1 At the intra-group sampling level, an improved Latin hypercube sampling strategy is adopted, which combines geometric constraints to correct parameter boundaries and eliminate invalid combinations. By introducing a minimum parameter interval control mechanism, the range of each parameter is divided to avoid process infeasibility caused by excessive subdivision; combined with motor geometric constraints, invalid parameter combinations are avoided to ensure data quality and improve sampling effectiveness.

[0052] In terms of the inter-group combination strategy, S1.2 adopts a random cross-combination method to combine the data of each parameter group. While ensuring the sampling quality within each subspace, it enhances the diversity of the overall parameter combination and improves the coverage and exploration breadth of the high-dimensional space.

[0053] S1.3 imports the generated reasonable parameter combinations into the finite element simulation software in batches to carry out electromagnetic field and structural stress field simulations, obtain multi-dimensional performance data including torque, electromagnetic loss, maximum stress, etc., and construct a high-quality and highly representative training dataset.

[0054] S2. Construction of AI Coupled Agent Model

[0055] like Figure 2 , Figure 4 As shown, an AI-coupled agent model is constructed using an improved Transformer neural network as the AI ​​agent model:

[0056] For electromagnetic fields, the encoder module in the improved Transformer neural network consists of a motor-set attention encoder, a stator attention encoder, a rotor attention encoder, and a winding attention encoder. For stress fields, the encoder module in the improved Transformer neural network consists of a motor-set attention encoder and a rotor attention encoder. Each sub-encoder consists of one multi-head attention layer and one feedforward layer, with 4 attention heads. The dimensions of the hidden layer and the feedforward layer are both 44. The activation function is consistent with the original Transformer, and residual connections and layer normalization are added to avoid gradient vanishing during training.

[0057] The multi-head attention calculation formula is MultiHead(Q,K,V)=concat(head1,…,head) h W O The formula for calculating single-head attention is: ;

[0058] S2.2 The stator attention encoder, rotor attention encoder, and winding attention encoder extract features from their respective structural parameters. The motor is equipped with an attention encoder that receives the output features from the three encoders and performs feature vector concatenation and fusion. The concatenation formula is as follows: The fused feature dimension is the sum of the output dimensions of each encoder.

[0059] In the S2.3 improved Transformer neural network, the decoder module consists of one multi-head attention layer, one multi-head cross-attention layer, and one feedforward layer. Residual connections and layer normalization are also added to prevent gradient vanishing. The multi-head cross-attention in the decoder receives the output from the encoder, and finally the decoder outputs the corresponding motor physics results.

[0060] The AI-coupled proxy model of this invention includes an improved Transformer neural network and a lumped-parameter thermal network (LPTN). The encoder of the improved Transformer neural network includes attention encoders for the stator, rotor, windings, and motor. Each encoder consists of one multi-head attention layer and one feedforward layer, with four attention heads. The dimensions of both the hidden and feedforward layers are 44. Residual connections and layer normalization are added after each layer. The decoder includes a multi-head self-attention layer, a multi-head cross-attention layer, a feedforward layer, residual connections, and layer normalization. The output features of the stator, rotor, and winding encoders are concatenated and input into the motor encoder, then into the decoder to output the physical field results. The model is trained using a 7:2:1 dataset partitioning, a learning rate of 0.03, a batch size of 256, and 1000 generations. The Adam optimizer and RMSE loss function are used, with convergence metrics of R² ≥ 0.95 and RMSE < 0.1. Electromagnetic losses are input into the LPTN, and temperature is used to inversely correct the electromagnetic and stress field parameters. Iteration continues until the adjacent temperature difference is < 5℃ or after 10 iterations to converge.

[0061] S3. Multiphysics bidirectional coupling

[0062] like Figure 5 , Figure 12 As shown, this invention employs a true two-way dynamic coupling: the electromagnetic loss output by the electromagnetic field model is fed back to the LPTN temperature field model in real time as an input heat source; the magnet temperature and winding temperature obtained from the temperature field calculation are fed back to the electromagnetic field model to correct the BH curve of the magnetic material and the resistivity of the winding, thereby realizing a two-way electromagnetic-thermal closed loop.

[0063] Simultaneously, the rotor temperature distribution is input into the stress field AI proxy model, ultimately outputting the maximum rotor stress value to determine the motor's structural safety. The entire coupled process is solved iteratively until convergence, with convergence conditions being: the temperature difference between two adjacent iterations is less than 5℃ or the number of iterations reaches 10. The final output includes comprehensive motor performance indicators encompassing electromagnetic performance, temperature rise characteristics, and structural strength.

[0064] like Figure 12As shown, the lumped parameter thermal network model in this invention is an engineering analysis method that simplifies complex thermal systems into equivalent thermal resistance and thermal capacity networks. Its core principle is to ignore the internal temperature gradient of the system, discretize the continuous temperature field into a finite number of "temperature nodes," and quickly solve for the system's temperature distribution through energy conservation and electrothermal analogy. The specific construction process is as follows: First, the system is decomposed into N nodes according to its structure (e.g., motor stator, rotor, windings), materials, and heat dissipation paths; then, the thermal resistance of the heat transfer paths between each node is calculated.

[0065] Thermal resistance:

[0066] , where L is the heat conduction path length, k is the thermal conductivity, and A is the normal heat conduction area.

[0067] Convection heat transfer thermal resistance:

[0068] , where h is the convective heat transfer coefficient and S is the fluid-solid interface area.

[0069] And allocate heat capacity to each node. Where ρ is density, V is volume, and c is density. p Specific heat at constant pressure. Then connect the nodes, thermal resistance, heat capacity and heat source to form a complete thermal network topology. Finally, establish the energy balance equation for each node and solve the system of differential equations to obtain the temperature of each node.

[0070] S4. Model Training

[0071] Before training the surrogate model, the dataset was standardized and divided into training, validation, and test sets in a 7:2:1 ratio. The learning rate was set to 0.03, batch size to 256, and training generations to 1000. The Adam optimizer and RMSE loss function were used for model training, and the trained model parameters were saved.

[0072] S5.AI-assisted multi-objective optimization

[0073] like Figure 6 As shown, the multi-objective optimization algorithm NSGA-II was built based on Python. The parameters such as population size, population evolution generation, motor operation settings variables, motor optimization variables, and constraints were set and fed into the trained AI coupled agent model to quickly obtain a number of key performance indicators, including torque characteristics, electromagnetic loss, temperature of various motor components, and maximum stress.

[0074] These performance outputs serve as objective functions and constraints, feeding back to the NSGA-II optimizer for iterative evolution. In each generation of evolution, the algorithm continuously explores the parameter space, filters out non-dominated solutions, and gradually approaches the global Pareto front. When the preset convergence conditions are met, the optimization process terminates, ultimately outputting a set of Pareto optimal solutions, which is visualized as a Pareto front plot, intuitively showing the trade-offs between the objectives.

[0075] The optimization objectives are: to maximize the average torque, minimize torque fluctuation, and minimize electromagnetic loss.

[0076] The constraints include: the maximum rotor stress meets the stress safety margin, the winding temperature does not exceed the insulation upper limit, and the magnet temperature does not exceed the demagnetization threshold.

[0077] S6 Testing and Verification

[0078] The prediction results of the AI-coupled agent model are compared with those of finite element simulation to verify the model accuracy; a typical coupled scenario is constructed to verify the electromagnetic-thermal-mechanical bidirectional feedback logic; and the optimization results are compared with traditional methods to verify the advantages of global optimization effect and computational efficiency.

[0079] S6.1 To verify the accuracy of the AI-coupled surrogate model, several design samples under typical working conditions will be selected and input into the surrogate model to obtain its prediction results for key performance indicators such as torque, torque ripple, electromagnetic loss, temperature of key components, and maximum stress. Subsequently, these prediction results will be compared with the high-precision simulation results of finite element analysis software, the relative error and root mean square error of each indicator will be calculated, and a scatter plot of prediction and simulation will be drawn to visually demonstrate the model accuracy.

[0080] To verify the bidirectional coupling mechanism, section 6.2 constructs a typical coupling scenario, such as electromagnetic losses causing winding temperature rise, which in turn affects the material's conductivity and permeability, and in turn affects the electromagnetic field distribution. Through multiple rounds of iterative simulation, the system's ability to accurately implement the closed-loop feedback logic of multi-field coupling is verified, and the evolution trends of each physical field variable during the coupling process are observed to ensure they conform to physical laws. For example, it verifies whether the temperature rise increases with increasing losses, and whether the overall performance tends to stabilize after coupling, ensuring the system can still operate stably under complex conditions.

[0081] S6.3 To verify the optimization effect of the multi-objective algorithm, within the feasible region of all parameters, the optimization objectives include mean torque, torque ripple, and electromagnetic loss. The optimization module generates a Pareto optimal front solution set. The obtained results are compared with the optimal solutions obtained by traditional finite element simulation-based single-point optimization and mesh scanning methods to evaluate whether the NSGA-II algorithm can effectively explore the design space, avoid getting trapped in local optima, and obtain a better global solution set. Simultaneously, the computation time, number of iterations, and convergence speed during the optimization process are recorded to evaluate its efficiency advantage while ensuring accuracy. Furthermore, the design points in the Pareto front solution set are compared with the design parameters and performance of existing prototypes to ensure the effectiveness of the optimization.

[0082] This example provides an AI-assisted multiphysics coupled motor optimization design method and system, and the range of motor parameters involved is shown in the table below.

[0083] Number of phases 3 Rotor type Single V-shape Peak speed 5800 rpm Rotor outer diameter 173 mm Rotor inner diameter 80 mm Iron core stacking 140 mm Magnet length 30~40 mm Magnet width 5~8 mm Minimum spacing of magnets 3~6 mm V-shaped angle of magnet 90~110 deg Edge magnetic bridge length 1.5 mm Length of central magnetic bridge 2.5 mm

[0084] like Figure 7 As shown, a two-dimensional model of a 54-slot, 6-pole motor was constructed using Ansys Maxwell software.

[0085] like Figure 8 As shown, a mechanical model of the motor rotor structure is constructed in Ansys Mechanical by sharing the rotor topology of the motor.

[0086] The target design parameters were selected by professionals in the field of motors. The Latin hypercube sampling algorithm with geometric constraints was used to generate a typical input dataset containing parameters such as stator and rotor geometric parameters and motor simulation parameters. The dataset was then imported into AnsysMaxwell and Ansys Mechanical to obtain performance output. Electromagnetic field and structural field input-output datasets consisting of 10,000 samples were constructed respectively.

[0087] An improved Transformer neural network was constructed. The electromagnetic field and structural field datasets constructed in step 1 were divided into training, validation, and test sets in a 7:2:1 ratio. Electromagnetic field and structural field neural network models were trained separately. The electromagnetic field neural network model consisted of a motor-set attention encoder, a stator attention encoder, a rotor attention encoder, and a winding attention encoder. The stress field neural network model consisted of a motor-set attention encoder and a rotor attention encoder. Each sub-encoder consisted of one multi-head attention layer and one feedforward layer, with 4 attention heads, and both the hidden layer and feedforward layer dimensions were 44. The activation function remained consistent with the original Transformer. The learning rate was set to 0.03, the batch size to 256, and the training generations to 1000. The Adam optimizer and RMSE loss function were used for model training. Training was stopped when the coefficient of determination on the test set reached above 0.95 and the RMSE was below 0.1. The trained model parameters were saved. The training results are shown below. Figure 9 As shown.

[0088] An AI-coupled agent model was constructed and combined with the NSGA-II multi-objective optimization algorithm to achieve fast multi-objective and multi-physics coupled optimization design of a 54-slot 6-pole motor. The population size was set to 60, the number of generations to 40, and the optimal solution set was visualized using the Pareto front. Figure 10 As shown.

[0089] The optimized motor can be obtained by selecting a design point from the Pareto solution set, such as... Figure 11 As shown. The model convergence and performance verification results are as follows. Figure 12 As shown.

[0090] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent modifications can be made to the technical solutions of the present invention, and all such equivalent modifications fall within the protection scope of the present invention. Furthermore, it should be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction.

Claims

1. An AI-assisted multi-physical field coupling motor optimization design method, characterized in that, Includes the following steps: S1. Parameter sampling: The motor design parameters are grouped, and a parameter set is generated by Latin hypercube sampling that integrates geometric constraints. Training data is obtained through finite element simulation. S2. AI Multiphysics Coupling Modeling: Based on the training data obtained in step S1, train the AI ​​proxy model, combine the AI ​​proxy model with the lumped parameter thermal network LPTN, and construct the AI ​​coupled proxy model; the electromagnetic loss output by the AI ​​proxy model is used as the heat source of LPTN, and the temperature output by LPTN is used to correct the electromagnetic field and stress field parameters in reverse. Iterate until convergence, forming a two-way coupling between data-driven and physical mechanism. S3. AI-assisted multi-objective optimization: The trained AI coupled agent model replaces the finite element simulation, and a multi-objective optimization algorithm is used to perform global optimization and output the Pareto optimal solution set.

2. The AI-assisted multi-physical field coupled motor optimization design method according to claim 1, characterized in that, The parameter grouping in step S1 includes motor setting group, rotor structure group, and stator winding group; the minimum parameter interval control is used within the group, and random cross combination is used between groups; the geometric constraints include the rotor outer diameter being greater than the inner diameter, the slot fill factor not exceeding the preset upper limit, and the magnet size meeting the minimum processing thickness. 3.The AI-assisted multi-physical field coupled motor optimization design method of claim 1, wherein, The AI ​​agent model described in step S2 is an improved Transformer neural network, which sets attention encoders for the stator, rotor, windings and motor for electromagnetic fields, and sets attention encoders for the rotor and motor for stress fields.

4. The AI-assisted multiphysics coupled motor optimization design method according to claim 3, characterized in that, The encoder and decoder of the improved Transformer neural network both include a multi-head attention layer, a feedforward layer, residual connections, and layer normalization.

5. The AI-assisted multiphysics coupled motor optimization design method according to claim 1, characterized in that, The temperature reverse correction in step S2 includes: correcting the resistivity of the winding temperature, correcting the BH curve of the magnet temperature, and correcting the thermal expansion coefficient of the rotor temperature; the iterative convergence condition is that the temperature difference between two adjacent iterations is less than 5℃ or the number of iterations reaches 10.

6. The AI-assisted multiphysics coupled motor optimization design method according to claim 1, characterized in that, The multi-objective optimization algorithm described in step S3 is the NSGA-II algorithm, with the optimization objectives being to maximize the mean torque, minimize torque fluctuation, and minimize electromagnetic loss.

7. The AI-assisted multiphysics coupled motor optimization design method according to claim 6, characterized in that, The optimized constraints include ensuring that the maximum rotor stress meets the stress safety margin, the winding temperature does not exceed the insulation limit, and the magnet temperature does not exceed the demagnetization threshold.

8. The AI-assisted multiphysics coupled motor optimization design method according to claim 1, characterized in that, Step S3 is followed by a testing and verification step, which compares the prediction results of the AI ​​coupled agent model with the finite element simulation to verify the coupling accuracy and optimization effect.

9. The AI-assisted multiphysics coupled motor optimization design method according to claim 1, characterized in that, The AI-coupled agent model enables millisecond-level performance prediction and replaces finite element simulation during the optimization process to reduce computation time.

10. An AI-assisted multiphysics coupled motor optimization design system, characterized in that, include: The system comprises a parameter sampling module, an AI-coupled agent model construction module, an AI-assisted multi-objective optimization module, and a testing and verification module, and is used to execute the method described in any one of claims 1-9.

Images