A multi-physics coupling simulation optimization method and system

Abstract

This invention discloses a multiphysics coupled simulation optimization method and system, belonging to the field of multiphysics coupled simulation technology. In the iterative optimization process, the method synchronously outputs local unreliability by a surrogate model. This local unreliability is calculated by fusing the prediction variance and the local curvature modulus. Each iteration constructs an adversarial optimization problem by weighting the difference between the prediction performance term and the local unreliability term, and generates a search path along the negative gradient direction, selecting the point with the highest local unreliability on the path as the simulation sampling point. The adversarial coefficient is adaptively adjusted according to the change in the local unreliability field. This invention achieves an active correlation between simulation sampling and the optimization search direction, thereby obtaining reliable optimization results within a finite number of simulations.

Description

Technical Field

[0001] This invention relates to the field of multiphysics coupling simulation technology, and more specifically, to a multiphysics coupling simulation optimization method and system. Background Technology

[0002] In structural optimization design under multiphysics coupling, a single high-fidelity coupled simulation is time-consuming and difficult to directly apply to optimization processes requiring numerous iterations. To reduce computational costs, surrogate models are typically used to replace high-fidelity simulators for optimization. This involves establishing an approximate mapping relationship between design variables and physical quantity responses using a small number of simulation samples, and then the optimizer performs optimization on the surrogate model.

[0003] In surrogate model-driven optimization, due to the limited number of initial samples, the surrogate model exhibits prediction bias in most areas of the design space, requiring continuous addition of new simulation samples to improve model accuracy. Existing methods often employ statistical criteria such as expected improvement and maximum variance to guide sampling, which select sampling locations based on the posterior statistics of the surrogate model. However, the sampling decisions of these criteria are independent of the optimizer's current search direction, and statistical evaluation is performed uniformly across the entire design space, failing to target the optimization path that the optimizer is actually focused on for targeted model improvement. This results in simulation computation being consumed in areas outside the optimization search direction, while insufficient samples on the optimization path lead to inaccurate predictions by the surrogate model, affecting the reliability of the optimization results. Therefore, a multiphysics coupling simulation optimization method and system are proposed to address the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a multiphysics coupled simulation optimization method and system, which aims to solve the problem that sampling decision and optimization search direction are disconnected in existing surrogate model optimization methods, resulting in simulation resources not being used in a targeted manner to improve the reliability of surrogate model predictions on the optimization path.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multiphysics coupled simulation optimization method and system transforms simulation sampling decision-making from passive statistical exploration to active adversarial pathfinding, ensuring that each simulation sample is precisely placed in the cognitive blind spot where the surrogate model is most vulnerable in the optimization direction.

[0006] This invention provides a multiphysics coupling simulation optimization method, comprising: S1: Obtain the design variable space, optimization objective and constraints, and generate the initial simulation sample set using experimental design methods.

[0007] These initial samples establish a data foundation between design variables and physical quantity responses, providing support for the construction of subsequent proxy models.

[0008] S2: Construct a surrogate model based on the initial simulation sample set. The surrogate model outputs the predicted values ​​of the structural physical quantity response and simultaneously outputs the local unbelievability.

[0009] The local unreliability is calculated by fusing the prediction variance and the local curvature modulus. The prediction variance reflects the uncertainty caused by the sparse sample size near the point, while the local curvature modulus reflects the degree of curvature of the predicted response surface at that point. Relying solely on the prediction variance is insufficient to comprehensively assess the model's reliability, because even in areas with sample coverage, if the response surface morphology is complex and has high curvature, the model's grasp of the trend at that location may still be inaccurate.

[0010] By combining the two, even if the samples are not sparse in a certain place, as long as the response surface changes drastically, the model will still show a high degree of unreliability, thus more accurately identifying the weak points in the surrogate model that really need to supplement information.

[0011] Furthermore, the surrogate model is a Gaussian process model. The local unreliability is calculated by fusing the posterior standard deviation of the Gaussian process model at a given point with the local curvature modulus of the model's predicted response surface at that point.

[0012] Furthermore, the local unreliability is calculated by multiplying the posterior standard deviation by a correction factor, where the correction factor is a monotonically increasing function with the local curvature modulus as input. The more significant the curvature of the response surface, the larger the correction factor, and the amplified the local unreliability, thereby prompting subsequent sampling to focus on such nonlinearly significant regions.

[0013] S3: Perform iterative sampling optimization. A single iteration includes: Constraint optimization is performed on the current agent model to obtain the constraint-optimal design point. This point represents the best choice under the current agent model's perception and will serve as the starting position for adversarial pathfinding.

[0014] Determine the adversarial coefficient. The adversarial coefficient is determined by the ratio of the current maximum value of the local unbelievability field to the initial maximum value of the local unbelievability field. This ratio signifies a comparison between the current cognitive state of the surrogate model and its initial state. The lower the current maximum unreliability relative to the initial value, the more complete the model's overall understanding, and the weaker the adversarial strength. As iterations progress, the agent model's understanding of key areas gradually becomes clearer, the maximum value of the local unreliability field naturally decreases, and the adversarial coefficient decreases accordingly, making the initial exploration in pathfinding more proactive, and then stabilizing in the later stages.

[0015] Furthermore, the resistance coefficient is determined by multiplying the aforementioned ratio by a preset coefficient.

[0016] An adversarial optimization problem is constructed. The objective function of the adversarial optimization problem is the weighted difference between the prediction performance term and the local unreliability term, where the weight of the local unreliability term is the adversarial coefficient. Conventional optimization only seeks the optimal prediction solution on the surrogate model, but this solution may be located in a region where the model's cognition is weak, and its prediction value is unreliable. By subtracting the local unreliability term from the prediction value, the optimizer, while pursuing lower prediction values, is forced to move in the direction that the surrogate model considers most unreliable. The optimization process is directly related to the conflict of model cognition.

[0017] Starting from the constrained optimal design point, solve the adversarial optimization problem, continuously move along the negative gradient direction to generate a search path, and select the point with the highest local unreliability on the search path as the sampling point for this iteration.

[0018] Furthermore, during the movement, points that exceed the constraint boundaries are projected to ensure that the entire path remains within the feasible region; the sequence of recorded trajectory points forms the search path, and the point with the maximum local unbelievability on this search path is selected as the sampling point.

[0019] The reason for not directly taking the optimal solution of the adversarial optimization problem, but instead finding the peak point of local unreliability on the path, is that this point is where seemingly superior performance and extreme model distrust converge. It is the weak link where the surrogate model is most likely to give misleading predictions. Adding simulation here can maximize the model's credibility in key directions with minimal computational cost.

[0020] Multiphysics high-fidelity simulation is performed on the sampling points to obtain the true response values. The sampling point and its true response pair are then added to the sample set to update the surrogate model. Through this precise sampling, the cognitive conflict of the surrogate model in this key direction is eliminated, the model forms a more accurate understanding of the input-output relationship in this region, and the prediction reliability is enhanced in a targeted manner.

[0021] S4: Stop the iteration when the maximum local unreliability on the search path generated in the current iteration is lower than the preset threshold.

[0022] Furthermore, the preset threshold is a pre-defined tolerance value for local unreliability. Iteration stops when the local unreliability of all points along the entire search path is lower than this threshold. When the unreliability along the entire path is lower than the threshold, it means that the model no longer has significant cognitive blind spots along the entire optimization direction from the current optimal design point to the adversarial optimization solution. At this point, the model's prediction reliability meets engineering requirements, and there is no need to add more simulation samples.

[0023] Perform non-confrontational constraint optimization on the final proxy model to obtain the optimal design variables, and output the optimal design variables and their simulation verification values.

[0024] Furthermore, the non-adversarial constraint optimization involves minimizing the predicted value output by the surrogate model while satisfying the constraints. Since the model is already sufficiently reliable in the optimization direction, it no longer requires the guidance of adversarial terms, and a high-quality optimal solution can be obtained through conventional optimization.

[0025] The present invention also provides a multiphysics coupled simulation optimization system, comprising: The initial modeling module is used to obtain the design variable space, optimization objectives and constraints, and generate an initial simulation sample set based on the experimental design, providing a data foundation for the construction of subsequent surrogate models.

[0026] The surrogate model module is used to train a surrogate model using a sample set. The surrogate model is configured to output predicted values ​​and a local unreliability calculated by fusing the prediction variance and the local curvature modulus. This module integrates information from two dimensions: sample sparsity and response surface curvature. This allows the local unreliability to more comprehensively identify areas of cognitive weakness in the surrogate model, providing accurate cognitive conflict signals for subsequent adversarial pathfinding.

[0027] Furthermore, the proxy model module adopts a Gaussian process model, and the local unreliability is generated by fusing the model's posterior standard deviation with the local curvature modulus of the predicted response surface.

[0028] The adversarial pathfinding module is used to perform constraint optimization on the current proxy model to obtain the constraint optimal design point, determine the adversarial coefficient, construct an adversarial optimization problem with the objective function being the weighted difference between the prediction performance term and the local unreliability term, solve the adversarial optimization problem with the constraint optimal design point as the starting point, generate a search path along the negative gradient direction, and select the peak point of local unreliability on the search path as the sampling point.

[0029] The adversarial coefficient is determined by the ratio of the current maximum value of the local unbelievability field to the initial maximum value of the local unbelievability field, and is adaptively adjusted according to the cognitive level of the model. The selected peak point is the most vulnerable cognitive blind spot of the surrogate model in the current optimization direction. Sampling at this point can most effectively improve the credibility of the model on the optimization path.

[0030] Furthermore, the adversarial coefficient is determined by multiplying the above ratio by a preset coefficient, and the search path is generated by continuously moving along the negative gradient direction of the objective function and projecting points that exceed the constraint boundary, ensuring that the path is always within the feasible region.

[0031] The simulation update module is used to call the multiphysics solver to perform high-fidelity simulation on the sampling point to obtain the real response value, and feed the simulation results back to the surrogate model module to update the model, thereby eliminating the cognitive blind spot of the surrogate model near the sampling point.

[0032] The convergence determination module is used to monitor the maximum local unreliability on the search path generated in each iteration. When the value is lower than the set threshold, a convergence signal is issued, indicating that the surrogate model has sufficient credibility in the optimization focus direction and can enter the final optimization stage.

[0033] The output module is used to receive the convergence signal, perform non-confrontational constraint optimization on the latest surrogate model, and output the final design variables and their simulation verification results.

[0034] The technical effects and advantages of this invention are as follows: In this invention, the surrogate model outputs not only the predicted value but also a local unreliability score, obtained by fusing the prediction variance and the local curvature modulus. This index integrates the influence of sample sparsity and response surface curvature on the model's prediction reliability, enabling the model to identify weak prediction regions that cannot be detected by variance alone.

[0035] In each iteration, an adversarial optimization objective is constructed using the weighted difference between the predicted performance term and the local unbelievability term. A search path is generated along the negative gradient direction, starting from the constrained optimal design point, and local unbelievability peak points are selected as sampling locations along the path. This approach actively links sampling decisions with the optimization search direction, concentrating simulation calculations on the locations along the optimization path where the model has the least confidence. The adversarial coefficient is determined by the ratio of the current maximum value of the local unbelievability field to the initial maximum value of the local unbelievability field, and it adaptively adjusts as the overall confidence of the model improves.

[0036] Through the above process, the cognitive blind spots of the surrogate model on the optimization path are gradually eliminated, and the predictive reliability of the model in the optimization direction is improved in a targeted manner, so as to obtain more reliable optimization results with fewer simulations. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the system module composition of the present invention; Figure 2 This is a schematic diagram of the system workflow of the present invention; Figure 3 This is a flowchart illustrating the proxy model construction and local unbelievability generation process of the present invention. Figure 4 This is a diagram showing the internal unit division of the anti-pathfinding module of the present invention. Detailed Implementation

[0038] 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.

[0039] Example 1 As attached Figure 2 and Figure 4 The multiphysics coupling simulation optimization method shown below has the following specific implementation details: Step 101: Obtain the design variable space, optimization objective, and constraints.

[0040] Design variables are denoted as vectors. , dimension Each dimension corresponds to an adjustable geometric parameter in the structure. The optimization objective is to minimize the physical quantity response at a critical location of the structure under multiphysics coupling, and this response value is denoted as a scalar. The constraint condition is that the structural mass or volume does not exceed a preset upper limit.

[0041] Step 102: Generate an initial simulation sample set using experimental design methods.

[0042] The Latin hypercube experimental design method is used to generate [variables] within the design variable space. One initial sample point, The value ranges from 10 to 20. For each sample point The multiphysics solver is invoked to perform a complete coupled simulation calculation to obtain the corresponding true response values. .

[0043] The multiphysics solver receives the design variable vector. As input, the physical quantity response value is calculated and output. If a simulation fails or returns an invalid value, a replacement sample point is selected for the simulation. All valid sample points and their response values ​​constitute the initial simulation sample set. , .

[0044] Step 103: Construct a proxy model based on the initial simulation sample set.

[0045] A surrogate model is used to approximate the input-output relationship in multiphysics simulations. As one implementation, the surrogate model employs a Gaussian process model, derived from the mean function. Sum of covariance functions Definition. Mean function The covariance function is set to zero. The covariance function uses a squared exponential kernel function, expressed as follows: ,in This is the signal variance hyperparameter. This is a length scale hyperparameter. The sample set... Input the logarithmic marginal likelihood function, and solve for the hyperparameters by maximizing this function. and If the optimization fails to converge, adjust the initial values ​​of the hyperparameters and re-optimize.

[0046] As another implementation method, the covariance function can be the Matérn3 / 2 kernel function, expressed as follows: The mean function can be taken as a linear combination of the design variables. Similarly, all hyperparameters are solved by maximizing the logarithmic marginal likelihood function.

[0047] Step 104: Construct the output and local untrustworthiness of the proxy model.

[0048] After training, for any input point The proxy model outputs the predicted mean. As the predicted value, the posterior variance is also output. Posterior standard deviation Take as The square root of.

[0049] In addition to the predicted values, the surrogate model also outputs the local unreliability, denoted as . . The construction is divided into two steps: first, calculate the local curvature modulus. Then and To integrate.

[0050] Local curvature modulus Reflecting the predicted response surface at point The degree of curvature at that point. (Regarding...) Seeking information about The second-order partial derivatives are used to construct the Hessian matrix. , its first Line 1 Column elements are .

[0051] As one implementation method, Take as The absolute value of Gaussian curvature is calculated as follows: ,in Represents a determinant. It is the gradient magnitude.

[0052] As another implementation method Take as The spectral norm, i.e., for After performing singular value decomposition, take the largest singular value.

[0053] Converged computing will and Combination generation As one implementation method, a product form is used. ,in As a correction factor, it is based on The input is a monotonically increasing function. Specifically, , A positive coefficient is preset, determined based on the magnitude of the physical quantity response and the design spatial scale, with a typical value range of 0.1 to 10. .or, , and To preset non-negative coefficients, at this time As another implementation method, fusion computing employs a linear weighted form. , and Preset positive weighting coefficients and .

[0054] Step 105: Calculate and store the maximum value of the initial local unbelievability field.

[0055] Based on the initial surrogate model constructed in steps 103 and 104, within the feasible region that satisfies the design constraints, random selection is performed according to a uniform distribution. An initial point, The value range is from 50 to 200.

[0056] Perform gradient ascent search independently for each initial point. Starting from that point, proceed along the gradient ascent path within the feasible region. Move along the gradient direction to find The local maximum. Gradient ascent step size. The method for determining it is as follows: ,in For the first The length of the value range of the design variable. This is the proportionality coefficient. Take a value between 0.01 and 0.1.

[0057] During the search process, when a certain point Less than the preset small amount (Pick When the gradient is calculated using the finite difference method, the analytical gradient calculation is skipped, and the gradient direction is estimated using the finite difference method. If a point exceeds the feasible region boundary, it is projected onto the boundary. At the feasible region boundary, Calculations are performed directly using boundary point coordinates, without additional extension. If the gradient magnitude falls below a preset value during a search starting from a certain point... (Pick to When the condition is met, stop the search at that starting point.

[0058] Record the local maximum points converged to from each starting point and their corresponding values. Value. If multiple starting points converge to the same point, only one is retained. The largest of all local maxima is taken as the initial local unbelievability field maximum value. When the calculation yields Less than the preset lower limit (Pick When ), Set as This is to prevent errors in subsequent calculations. Stored for use in subsequent iterations.

[0059] Step 106: Enter the iterative sampling optimization stage and set the number of iterations. .

[0060] Step 106.1: Obtain the constrained optimal design point.

[0061] In the current proxy model, the sequential quadratic programming method is used to solve the constrained optimization problem, that is, to minimize the constraint under the condition of satisfying the constraint. When solving, input the current prediction function. Given the constraints, the output is the constrained optimal solution, denoted as... .when The absolute value of each component exceeding the boundary or violating implicit constraints is less than (Pick to When a problem occurs, it is considered a slight out-of-bounds error, and the point is first projected to satisfy the constraints. If the solution fails, the solution is rolled back to the optimal design point from the previous iteration, or a point is selected from the current sample set. The smallest feasible sample point is used as .

[0062] Step 106.2: Determine the adversarial coefficient.

[0063] Based on the current agent model, calculate the maximum value of the local unbelievability field within the feasible region. The calculation method is the same as in step 105, including the calculation of... Too small and Too small a processing size.

[0064] Read the data stored in step 105 Resistance coefficient , A positive coefficient is preset, determined based on the magnitude of the optimization objective, with a typical value range of 0.1 to 10. When the calculated... Exceeding the preset limit (Pick to )season .

[0065] As one implementation method, the following can also be adopted: , The smoothing exponent takes values ​​greater than 0 and not greater than 1. Exceeding Cut off at that time.

[0066] Step 106.3: Construct an adversarial optimization problem and generate a search path.

[0067] Construct the objective function ,in is the weighting coefficient, set to 1. In the objective function, the prediction performance term is... The partial unreliability item is The weights are the adversarial coefficients obtained in step 106.2. .

[0068] The result obtained in step 106.1 Starting from the point, along The search path is generated by continuously moving along the negative gradient direction. The gradient is calculated. ,in This is provided by the proxy model parsing.

[0069] for The calculation, in exponential form For example, the calculation formula is: .

[0070] in, ,when The value of this item is zero. With the current point Sample set The hyperparameters, already trained, are used as input, and the output is... The gradient vector is a dimensional vector formed by the covariance kernel function. right The variance is obtained by combining the partial derivatives and the formula for predicting variance using the Gaussian process. Through the The expression is obtained by differentiating it. When When taking the absolute value form of Gaussian curvature, Through molecular and denominator The derivatives are obtained by taking the derivatives separately and combining them according to the quotient rule, which involves... The third partial derivative. When The absolute value is less than When this is the case, the derivative of this part is calculated using the numerical difference method. When the calculated value exceeds the preset upper limit, it is truncated to that upper limit.

[0071] If a polynomial or linear weighted form is used to construct... Its gradient Then, by differentiating the corresponding expression, we can obtain the result.

[0072] The path generation uses the gradient descent method. Let the current point be... The test point is obtained along the negative gradient direction. Step length The method for determining it is as follows: ,in This is the proportionality coefficient, ranging from 0.001 to 0.01.

[0073] examine Does the design constraint condition meet? If it does, then let... If not satisfied, then... Projection processing is performed iteratively, alternating between component boundary projection and implicit constraint projection: for each component, if it exceeds the upper bound of the design variable, the upper bound value is used; if it falls below the lower bound, the lower bound value is used. For violations of implicit constraints such as mass or volume, the component is adjusted to the constraint boundary using a binary search along the gradient direction of the constraint function. The binary search terminates when the absolute value of the constraint function at the current point is less than a preset tolerance. (Pick to The projection process continues until a point simultaneously satisfies all constraints, or reaches a preset maximum number of iterations. The resulting point is obtained after projection. .

[0074] when and The distance between them is less than the preset value (Pick If the path is deemed to be stalled, path generation is terminated prematurely.

[0075] Repeat the above update process until any of the following conditions are met: Condition 1: Reach the preset maximum number of iterations. , The value range is from 30 to 100; Condition two, Lower than the preset value (Pick to ); Condition 3: The path becomes stagnant.

[0076] Recording point sequence ,in This sequence is the search path generated in this iteration.

[0077] As an alternative implementation, a linear search gradient descent method with backoff can also be used to generate the search path. Start, calculate the search direction ,along With initial trial step size Try to obtain trial points and perform constraint projection, then check the Armijo conditions (where the constant is taken as...). If the condition is not met, multiply the step size by the shrinkage factor of 0.5 and try again.

[0078] Alternatively, the conjugate gradient method can be used to generate the search path, with the search direction updated according to the Fletcher-Reeves formula. Line search employs strong Wolfe conditions, and the parameters... Take 0.1, Take 0.9.

[0079] The guiding principle for choosing which path generation method to use is: when the design space dimension is low ( When the objective function exhibits significant numerical oscillations on the search path, the fixed step size gradient descent method is preferred; when the objective function exhibits significant numerical oscillations on the search path, the line search gradient descent method is used; and when faster convergence to the adversarial optimization solution is required, the conjugate gradient method is employed.

[0080] Regardless of the path generation method used, after each update step, the same projection processing must be performed on the out-of-bounds points, and the same stall judgment must be executed.

[0081] Step 106.4: Select sampling points.

[0082] The search path generated in step 106.3 Calculate the local unbelievability at each point. Select The point with the largest value is taken as the sampling point for this iteration, denoted as . If multiple points have the same maximum If the value is selected, then choose one of them. The smallest point is used as .

[0083] Step 106.5: Perform high-fidelity simulation and update the agent model.

[0084] Call the multiphysics solver to Perform high-fidelity simulation to obtain the corresponding real response values. When the multiphysics solver returns an invalid value or the simulation fails, the sample point is marked as infeasible, its response value is set to a preset large value, which is 1.5 to 2 times the maximum response value in the current sample set, or 10 to 100 times the typical order of magnitude of the optimization target, and it is still added to the sample set to update the surrogate model.

[0085] New sample Add to sample set The agent model is retrained using the updated sample set, and the hyperparameters are updated. , , , Each function. If the log marginal likelihood decreases or the hyperparameter optimization fails to converge after retraining, retain the model parameters before the update, adjust the initial values ​​of the hyperparameter optimization, and then retrain.

[0086] Step 106.6: Check the convergence conditions.

[0087] Obtain the search path generated in this iteration Calculate the local unreliability of all points on the path. When all points All values ​​are below the preset threshold. When the iteration stops, proceed to step 107.

[0088] when At the same time, the following stall criteria are also checked: if continuous In this iteration, the result calculated in step 106.2 is... relatively Change All are lower than the preset value The iteration stops and proceeds to step 107, where... Take 3 to 5, Pick to .

[0089] If the above convergence condition is not met, then let Increment by 1, return to step 106.1 to proceed to the next iteration. Preset threshold. This is a pre-defined tolerance value for local unreliability.

[0090] Thus, through the loop from steps 106.1 to 106.6, the proxy model is updated in a targeted manner in each iteration, and its cognitive blind spots are gradually eliminated.

[0091] Step 107: Perform non-confrontational constraint optimization and output the results.

[0092] After the iteration convergence in step 106, perform non-adversarial constraint optimization on the final surrogate model, that is, minimize the predicted value while satisfying the constraints. The sequential quadratic programming method is used to solve this problem. The input is... Given the constraints, the output is the optimal solution. Call the multiphysics solver to... Perform a high-fidelity simulation to obtain simulation verification values. If the simulation fails, then select from the simulated sample points already simulated during the iteration process. The smallest feasible point is taken as the alternative optimal solution, and its corresponding simulation response value is used as... Output optimal design variables. and its simulation verification values The optimization process is now complete.

[0093] System Implementation Process As attached Figure 1 and Figure 3 As shown, this invention also provides a multiphysics coupled simulation optimization system, including an initial modeling module, a proxy model module, an adversarial pathfinding module, a simulation update module, a convergence determination module, and an output module. The system operates according to the following steps.

[0094] Step 201: The initial modeling module obtains the design variable space, optimization objective, and constraints. It generates an initial sample point set according to the Latin hypercube experimental design and calls an external multiphysics solver to simulate each sample point. The input to the multiphysics solver is the design variable vector. The output is the physical quantity response value. The initial modeling module will generate the design variable vectors for each sample point. With simulation response value Combined into an initial simulation sample set and will Pass it to the proxy model module.

[0095] Step 202: Proxy model module receives A Gaussian process model is then constructed, and after training, the predicted mean function is output. Posterior standard deviation function Local curvature modulus function and local unbelievability function Simultaneously, calculate the maximum value of the initial local unbelievability field as described in step 105. And store. The module provides two interfaces: a prediction interface that receives the design point vector. Return scalar The unreliability query interface receives design point vectors. Return scalar .

[0096] Step 203: The adversarial pathfinding module performs the following operations: Pass the design point vector to the prediction interface and call the sequential quadratic programming method to... Perform constraint optimization to obtain the constrained optimal design point. Call the unbelievability query interface and calculate the current local unbelievability field maximum value as described in step 106.2. Read the data stored in the proxy model module. Then determine the resistance coefficient Or use a form with a smoothing exponent. and in accordance with the method described in step 106.2 Perform upper bound truncation. Construct the objective function. .

[0097] by Starting from the negative gradient direction, a search path is generated. The specific methods of gradient calculation and path generation are the same as in step 106.3, including gradient anomaly handling, constraint projection, step size determination, and stagnation judgment.

[0098] Select path The point with the largest value Output to the simulation update module when multiple identical maximum values ​​exist. The points are selected according to the rules described in step 106.4. Simultaneously, the search path and each point on the path are... The value is provided to the convergence determination module.

[0099] Step 204: Simulation update module receives The true response value is obtained by calling the multiphysics solver. If the simulation fails, proceed as described in step 106.5. Add to sample set This triggers the proxy model module to retrain the model and update all output functions. If the update fails, it is handled as described in step 106.5. After the update is complete, the convergence determination module is notified.

[0100] Step 205: The convergence determination module obtains the search path and the points on the path for this iteration from the adversarial pathfinding module. Value, to determine all points on the path Are all values ​​below the set threshold? .when The system also checks the stagnation judgment condition described in step 106.6. If the convergence condition is met, a convergence signal is sent to the output module; if not, the system returns to step 203 for the next iteration.

[0101] Step 206: After receiving the convergence signal, the output module calls the prediction interface of the surrogate model module to perform prediction under the constraints. To minimize the problem, the optimal design variables are obtained using the sequential quadratic programming method. Call the multiphysics solver once to... Perform simulation to obtain verification values If the simulation fails, select an alternative optimal solution as described in step 107. and As the final output.

[0102] The data flow between modules is as follows: The initial modeling module will Pass it to the proxy model module; The proxy model module provides the adversarial pathfinding module with , Functions and ; Output of the adversarial pathfinding module The simulation update module outputs path data to the convergence determination module; The simulation update module feeds back the new samples to the proxy model module to trigger an update; The convergence determination module sends a convergence signal to the output module; The output module obtains the final model from the proxy model module and completes the output.

[0103] The implementation process of the above system corresponds completely to the implementation process of the aforementioned method. When the algorithm implementation inside the proxy model module or the adversarial pathfinding module adopts different implementation methods, only the corresponding algorithm configuration needs to be replaced, and the calling interface and data flow between modules remain unchanged.

[0104] The above description is merely a specific embodiment of the present invention, and the scope of protection of the present invention is not limited thereto. Any equivalent changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multiphysics coupled simulation optimization method, characterized in that, include: S1: Obtain the design variable space, optimization objective, and constraints, and generate the initial simulation sample set using experimental design methods; S2: Construct a surrogate model based on the initial simulation sample set. The surrogate model outputs predicted values ​​and simultaneously outputs local unreliability. The local unreliability is calculated by fusing the prediction variance and the local curvature modulus. S3: Perform iterative sampling optimization. A single iteration includes: Perform constraint optimization on the current proxy model to obtain the constraint-optimal design point; Determine the adversarial coefficient, which is determined by the ratio of the current maximum value of the local unbelievability field to the initial maximum value of the local unbelievability field; Construct an adversarial optimization problem, wherein the objective function of the adversarial optimization problem is the weighted difference between the prediction performance term and the local unreliability term, and the weight of the local unreliability term is the adversarial coefficient; Starting from the constrained optimal design point, solve the adversarial optimization problem, continuously move along the negative gradient direction to generate a search path, and select the point with the highest local unreliability on the search path as the sampling point for this iteration; Perform multiphysics high-fidelity simulation on the sampling point to obtain the real response value, and add the sampling point and its real response pair to the sample set to update the surrogate model; S4: Stop the iteration when the maximum local unreliability on the search path generated in the current iteration is lower than the preset threshold. Perform non-confrontational constraint optimization on the final proxy model to obtain the optimal design variable and output the optimal design variable and its simulation verification value.

2. The multiphysics coupling simulation optimization method according to claim 1, characterized in that, The surrogate model described in S2 is a Gaussian process model, and the local unreliability is calculated by fusing the posterior standard deviation of the Gaussian process model at a given point with the local curvature modulus of the model's predicted response surface at that point.

3. The multiphysics coupling simulation optimization method according to claim 2, characterized in that, The local unreliability is calculated by multiplying the posterior standard deviation by a correction factor, where the correction factor is a monotonically increasing function with the local curvature modulus as input.

4. The multiphysics coupling simulation optimization method according to claim 1, characterized in that, The adversarial coefficient described in S3 is determined by multiplying the ratio of the current maximum value of the local unbelievability field to the initial maximum value of the local unbelievability field by a preset coefficient.

5. The multiphysics coupling simulation optimization method according to claim 1 or 4, characterized in that, The process of generating the search path and selecting the point with the highest local unbelievability, as described in S3, is as follows: Starting from the optimal design point of the constraint, the trajectory is formed by continuously moving along the negative gradient direction of the objective function. During the movement, points that exceed the constraint boundary are projected and the sequence of trajectory points is recorded to form a search path. The maximum value of local unbelievability is selected as the sampling point on the search path.

6. The multiphysics coupling simulation optimization method according to claim 1, characterized in that, The preset threshold mentioned in S4 is a pre-set local unreliability tolerance value. The iteration stops when the local unreliability of all points on the entire search path is lower than this threshold.

7. The multiphysics coupling simulation optimization method according to claim 1, characterized in that, The non-adversarial constraint optimization described in S4 is: minimizing the predicted value output by the surrogate model under the constraint conditions.

8. A multiphysics coupled simulation optimization system, characterized in that, include: The initial modeling module is used to obtain the design variable space, optimization objectives and constraints, and generate an initial simulation sample set based on the experimental design; A surrogate model module is used to train a surrogate model using a sample set. The surrogate model is configured to output a predicted value and a local unbelievability calculated by fusing the predicted variance and the local curvature modulus. The adversarial pathfinding module is used to perform constraint optimization on the current proxy model to obtain the constraint optimal design point, determine the adversarial coefficient, construct an adversarial optimization problem with the objective function being the weighted difference between the prediction performance term and the local unbelievability term, solve the adversarial optimization problem with the constraint optimal design point as the starting point, generate a search path along the negative gradient direction, and select the peak point of local unbelievability on the search path as the sampling point, wherein the adversarial coefficient is determined by the ratio of the current maximum value of the local unbelievability field to the initial maximum value of the local unbelievability field; The simulation update module is used to call the multiphysics solver to perform high-fidelity simulation on the sampling points to obtain the real response values, and feed the simulation results back to the proxy model module to update the model. The convergence determination module is used to monitor the maximum local unreliability on the search path generated in each iteration, and issues a convergence signal when the value is lower than a set threshold. The output module is used to receive the convergence signal, perform non-confrontational constraint optimization on the latest surrogate model, and output the final design variables and their simulation verification results.

9. The multiphysics coupled simulation optimization system according to claim 8, characterized in that, The proxy model module adopts a Gaussian process model, and the local unreliability is generated by fusing the model's posterior standard deviation with the local curvature modulus of the predicted response surface.

10. The multiphysics coupled simulation optimization system according to claim 8 or 9, characterized in that, In the adversarial pathfinding module, the adversarial coefficient is determined by multiplying the ratio of the current maximum value of the local unbelievability field to the initial maximum value of the local unbelievability field by a preset coefficient, and the search path is generated by continuously moving along the negative gradient direction of the objective function and projecting points that exceed the constraint boundary.

Images