An ai-based microphone design optimization method and system

Abstract

This invention relates to the field of microphone design technology, specifically to an AI-based microphone design optimization method and system. The method includes: establishing a design space based on the microphone's structural parameters, material parameters, operating parameters, and performance requirements; determining design variables, optimization objectives, and constraints; generating an initial sample set and performing high-fidelity physical field simulation to obtain target values; performing various preprocessing steps on the target values, constructing an AI surrogate model, and selecting the current surrogate model; embedding the current surrogate model into a multi-objective evolutionary optimization process to obtain a candidate non-dominated solution set; further filtering candidate designs based on prediction uncertainty and performing high-fidelity physical field simulation to update the surrogate model online; and after meeting the stopping condition, verifying the non-dominated solutions and outputting a Pareto solution set that satisfies the constraints. This invention can improve the efficiency and reliability of multi-objective microphone optimization.

Description

Technical Field

[0001] This invention relates to the field of microphone design technology, and in particular to an AI-based microphone design optimization method and system. Background Technology

[0002] As a sound-to-electric conversion device, the structural design of a microphone typically needs to simultaneously consider multiple requirements, including sensitivity, frequency response characteristics, mechanical reliability, driving or detection stability, and manufacturing feasibility. In current design processes, researchers often rely on parameter trial and error, local empirical corrections, and repeated verification through high-fidelity simulations to select solutions. This approach is computationally expensive, especially when there are many design variables, coupling constraints between objectives, and the need to meet specific performance boundaries, easily leading to prolonged design cycles. Furthermore, traditional optimization methods often emphasize a single performance indicator, or struggle to simultaneously consider the quality, quantity, and engineering feasibility of solutions in multi-objective scenarios, thus failing to provide microphone designers with multiple candidate solutions of practical reference value. While some surrogate modeling optimization methods can reduce some computational burden, they still have shortcomings in prediction accuracy, constraint satisfaction capability, and dynamic correction capabilities in key areas of the design space during the optimization process. Summary of the Invention

[0003] In view of the above technical problems, the present invention provides an AI-based microphone design optimization method and system, which aims to solve the problems of high cost of high-fidelity simulation, difficulty in multi-objective collaborative optimization, difficulty in taking into account constraints, and insufficient prediction accuracy of existing surrogate models in key design areas during the microphone design optimization process.

[0004] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0005] According to one aspect of the present invention, an AI-based microphone design optimization method is proposed, the method comprising the steps of: A design space is established based on the structural parameters, material parameters, operating parameters, and performance requirements of the microphone to be optimized, and the design variables, multiple mutually constraining optimization objectives, and the constraints corresponding to the optimization objectives are determined. An initial sample set is generated within the design space, and a high-fidelity physical field simulation is performed on each of the initial sample sets to obtain the target value corresponding to each initial sample. The target value is preprocessed in various ways to form multiple training datasets. Based on each training dataset, an artificial intelligence agent model for the optimization target is constructed. The current agent model is selected according to a preset accuracy evaluation rule. The current agent model is used to output the target prediction value and prediction uncertainty of the candidate design. The current proxy model is embedded into a multi-objective evolutionary optimization process. Based on the target prediction value, the candidate designs are non-dominated, constrained, and diversity is preserved to obtain a set of candidate non-dominated solutions. Based on the predicted uncertainty, candidate designs with predicted uncertainty higher than a preset threshold are selected from the candidate non-dominated solution set. The selected candidate designs are subjected to high-fidelity physical field simulation, and the newly added simulation data is incorporated into the training data corresponding to the current proxy model so as to update the current proxy model online during the optimization process. After the stopping condition is met, the high-fidelity physical field simulation verification is performed on the non-dominated solution obtained by optimization. The solution that satisfies the constraint condition is retained as the microphone design optimization result, and the corresponding Pareto solution set is output.

[0006] Furthermore, the design variables include at least one of the following: microphone geometric parameters, material parameters, operating condition parameters, and boundary condition parameters. The design variables can be in continuous, discrete, or mixed form. The initial sample set is generated by combining Latin hypercube sampling with the maximum-minimum distance criterion, so that the initial sample set has uniform coverage and spatial filling within the design space.

[0007] Furthermore, the various preprocessing steps include: performing normalization, logarithmic transformation, and logarithmic transformation followed by normalization on the target value; for the target value after normalization, a positive shift is used to keep the preprocessed target value positive, so as to reduce the impact of dimensional differences and numerical ranges between different optimization targets on the accuracy of proxy modeling.

[0008] Furthermore, the artificial intelligence agent model is a Gaussian process regression model established for each optimization objective; the Gaussian process regression model uses the squared exponential covariance kernel to characterize the correlation between design samples, and determines the model hyperparameters by maximizing the marginal likelihood, so as to output the target prediction mean and standard deviation corresponding to each candidate design.

[0009] Furthermore, the preset accuracy evaluation rules include: performing cross-validation on the AI ​​agent models corresponding to each training dataset, calculating the average absolute percentage error of each optimization objective, and combining the agent models with the lowest prediction error as the current agent model.

[0010] Furthermore, the multi-objective evolutionary optimization process employs a non-dominated sorting genetic algorithm, which generates new candidate designs through parent selection, crossover, and mutation operations, and retains individuals by combining non-dominated sorting with crowding distance. During the individual retention process, boundary constraints, performance threshold constraints, inequality constraints, and equality constraints of design variables are simultaneously determined to output a non-dominated solution set.

[0011] Furthermore, the online update includes: in each generation of the multi-objective evolutionary optimization process, firstly, using the current surrogate model to predict the target prediction value and prediction uncertainty of individuals in the population, then selecting candidate designs with prediction uncertainties higher than the preset threshold from those that satisfy the constraints and belong to non-dominated solutions, in descending order of prediction uncertainty, and performing the high-fidelity physical field simulation; each time new simulation data is obtained, the new simulation data is immediately incorporated into the training data and the current surrogate model is retrained, and then the candidate design screening for the current generation continues; when there are no candidate designs that satisfy the preset threshold in the current generation, the unused high-fidelity physical field simulation times of the current generation are transferred to subsequent generations.

[0012] Furthermore, the review process includes: The high-fidelity physical field simulation is performed on each of the optimized non-dominated solutions to obtain the true target value; The non-dominated solutions that still satisfy the constraints after review are determined to be valid non-dominated solutions. The reliability of the optimization is evaluated based on the average absolute percentage error between the true target value and the predicted target value. Then, the quality of the solution set is evaluated using the hypervolume index, wherein the reference point of the hypervolume index is determined based on the worst target point in the joint solution set composed of multiple solution sets to be compared.

[0013] Furthermore, the high-fidelity physical field simulation is a finite element multiphysics simulation, which is established based on the working mechanism of the microphone and is used to output at least two of the stress, displacement, natural frequency, temperature rise and electrical response as the true evaluation results of the optimization objective.

[0014] According to another aspect of the present invention, an AI-based microphone design optimization system is provided, comprising: The design space construction module is used to establish a design space based on the structural parameters, material parameters, operating parameters and performance requirements of the microphone to be optimized, and to determine the design variables, multiple mutually restrictive optimization objectives and the constraints corresponding to the optimization objectives. The initial sample simulation module is used to generate an initial sample set within the design space and perform high-fidelity physical field simulation on each initial sample set to obtain the target value corresponding to each initial sample. The proxy model construction module is used to perform various preprocessing on the target value to form multiple training datasets, construct artificial intelligence proxy models for the optimization target based on each training dataset, and select the current proxy model according to the preset accuracy evaluation rules. The current proxy model is used to output the target prediction value and prediction uncertainty of the candidate design. The multi-objective optimization solution module is used to embed the current surrogate model into the multi-objective evolutionary optimization process, and perform non-dominated ranking, constraint determination and diversity preservation on the candidate designs based on the target prediction values ​​to obtain a set of candidate non-dominated solutions. An adaptive update module is used to select candidate designs with prediction uncertainties higher than a preset threshold from the candidate non-dominated solution set based on the prediction uncertainty, perform the high-fidelity physical field simulation on the selected candidate designs, and incorporate the newly added simulation data into the training data corresponding to the current proxy model, so as to update the current proxy model online during the optimization process. The result verification output module is used to perform the high-fidelity physical field simulation verification on the optimized non-dominated solution after the stopping condition is met, retain the solution that satisfies the constraint condition as the microphone design optimization result, and output the corresponding Pareto solution set.

[0015] The technical solution of the present invention has the following beneficial effects: Compared with existing technologies, this invention combines artificial intelligence agent modeling with a multi-objective optimization process and introduces a dynamic update mechanism for candidate solutions with high uncertainty during the optimization process. This reduces the number of calls to high-fidelity physical field simulations while improving the prediction reliability of key areas, making the optimization process more suitable for handling complex scenarios in microphone design where multiple performance indicators are mutually constrained and must meet constraints. At the same time, this invention can output a verified non-dominated solution set, providing designers with multiple sets of candidate designs with engineering significance, rather than just a single result. Therefore, it is beneficial to improve design efficiency, enhance the reliability of optimization results, and increase the flexibility of selecting the final solution under practical application requirements. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an AI-based microphone design optimization method as described in the embodiments of this specification. Figure 2 This is a structural block diagram of an AI-based microphone design optimization system as described in the embodiments of this specification. Detailed Implementation

[0017] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make the invention more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention may be practiced with one or more of these specific details omitted, or other methods, components, systems, steps, etc., may be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of the invention.

[0018] Furthermore, the accompanying drawings are merely illustrative of the invention. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0019] This invention provides an AI-based microphone design optimization method. (Refer to...) Figure 1 The diagram shown is a flowchart of an AI-based microphone design optimization method according to an embodiment of the present invention. The method may specifically include the following steps S101-S106: In step S101, a design space is established based on the structural parameters, material parameters, operating parameters, and performance requirements of the microphone to be optimized, and the design variables, multiple mutually restrictive optimization objectives, and constraints corresponding to the optimization objectives are determined.

[0020] The design variables include at least one of the following: microphone geometric parameters, material parameters, operating condition parameters, and boundary condition parameters. The design variables can be in continuous, discrete, or mixed form. The initial sample set is generated by combining Latin hypercube sampling with the maximum-minimum distance criterion, so that the initial sample set has uniform coverage and spatial filling within the design space.

[0021] In step S101, a design space is first established around the microphone to be optimized. The design space characterizes the possible range of design parameters and their combinations. Essentially, it maps parameters affecting microphone performance into a unified parameter domain for subsequent sample generation, performance evaluation, and optimization searches. Parameters in the design space can include geometric parameters, material parameters, operating condition parameters, and boundary condition parameters. Geometric parameters describe the dimensional characteristics of the microphone's sensitive structure, backplate structure, support structure, cavity structure, or acoustic channel structure; material parameters describe the material properties of the various functional components of the microphone; operating condition parameters describe the electrical, thermal, mechanical, or acoustic excitation conditions applied to the microphone during operation; and boundary condition parameters describe the fixing method, connection method, constraint method, and the effects of the external environment. Design parameter optimization is typically performed before device manufacturing. The optimization objects can include geometric structure, material properties, and parameters related to actuation or working mechanisms. Therefore, incorporating these parameters into the design space provides clear variable boundaries and an evaluation basis for the subsequent optimization process.

[0022] After establishing the design space, the correspondence between design variables, optimization objectives, and constraints is further determined. Design variables can take continuous, discrete, or hybrid forms. Continuous forms are suitable for continuously adjustable quantities such as length, width, thickness, gap, scaling factor, elastic parameters, or voltage; discrete forms are suitable for quantities that can only take discrete values, such as the number of turns, layers, number of units, and arrangement levels; hybrid forms are suitable for design problems that simultaneously contain continuous and discrete variables. This type of optimization method is suitable for handling continuous, combined, and hybrid variables, without requiring additional constraints on the convexity or differentiability of the objective function, thus adapting to scenarios with multiple types of parameters in microphone design. For some parameters, ratio variables can also be used to describe relative structural relationships, thereby expressing structural coupling relationships without directly limiting absolute dimensions, improving design expressiveness.

[0023] When determining the optimization objectives, multiple performance indicators that are mutually constrained in the microphone design are considered as joint evaluation objects. These objectives often exhibit an inverse relationship; improvement in one objective may lead to degradation in another. Therefore, multi-objective characterization is necessary within a unified design space. The multi-objective optimization problem can be written as: ; Its constraints can be written as: ; ; ; Where M represents the number of optimization objectives, J represents the number of inequality constraints, K represents the number of equality constraints, n represents the number of design variables, and x represents the design variable vector. Let m be the optimization objective. Denotes the j-th inequality constraint. This represents the k-th equality constraint. and Let represent the lower and upper bounds of the i-th design variable, respectively. A design solution satisfying inequality constraints, equality constraints, and variable boundary conditions can be considered a feasible solution. In this way, multiple performance requirements of the microphone can be transformed into computable objective and constraint terms, allowing subsequent optimization searches to focus not only on performance improvement but also on engineering feasibility.

[0024] Constraints are set to ensure that the optimization results meet specific application requirements and design limitations. Constraints can originate from performance thresholds, structural safety requirements, natural frequency requirements, temperature rise requirements, response requirements, electrical load requirements, and manufacturing feasibility requirements, among others. The optimization framework emphasizes the handling of objective constraints, ensuring that the resulting design is not merely a numerically optimal point, but rather an effective set of solutions formed under given application constraints. This avoids searching the design space for parameter combinations that, while locally optimal, are impractical.

[0025] After defining the design space, an initial sample set is generated. The purpose of this initial sample set is to provide representative training data for subsequent high-fidelity simulations and surrogate model training. To ensure that the training samples accurately represent the actual model, the samples need to be distributed as uniformly as possible throughout the design space. While random sampling has lower computational cost, it often cannot guarantee sufficient coverage of the design space. Therefore, Latin hypercube sampling is used to generate sample points. The Latin hypercube sampling process is as follows: if n sample points need to be extracted from the design space, each dimension of the design space is divided into n segments, and stratified sampling is performed along each dimension. This ensures that when traversing the hypercube in any dimension, a sample point at the same level will not be encountered repeatedly. This method improves the uniformity of distribution across each dimension.

[0026] Considering that Latin hypercube sampling alone cannot fully guarantee spatial filling, the maximum-minimum distance criterion is further incorporated into the sample generation process. The core of the maximum-minimum distance criterion is to maximize the minimum distance between sample points, thereby promoting a more uniform distribution of samples within the design space. By combining Latin hypercube sampling with the maximum-minimum distance criterion, both hierarchical coverage and spatial filling can be considered, making the initial sample set more representative. The resulting samples not only reduce local clustering but also mitigate the risk of undersampling in certain areas, providing a more stable data foundation for subsequent high-fidelity simulations and surrogate model fitting.

[0027] The size of the initial sample set is not necessarily better the larger it is. Increasing the number of samples usually improves model accuracy, but it also increases the computational cost of high-fidelity simulation. Therefore, the sample size needs to be balanced between model accuracy and computational efficiency. In this step, by generating an initial sample set with uniform coverage and space-filling properties within the design space, the mapping relationship between microphone design parameters and performance can be reflected as completely as possible with limited simulation overhead, laying the foundation for subsequent target value acquisition, surrogate model construction, and multi-objective optimization solutions.

[0028] In step S102, an initial sample set is generated within the design space, and a high-fidelity physical field simulation is performed on each of the initial sample sets to obtain the target value corresponding to each initial sample.

[0029] The high-fidelity physical field simulation is a finite element multiphysics simulation. The finite element multiphysics simulation is established based on the working mechanism of the microphone and is used to output at least two of the following: stress, displacement, natural frequency, temperature rise, and electrical response, as the true evaluation result of the optimization objective.

[0030] In step S102, after generating an initial sample set within the design space, high-fidelity physical field simulation is performed on each sample design in the initial sample set to obtain the true target value corresponding to that sample design. The purpose of this step is to transform the design points in the parameter space into evaluation results that characterize the actual performance of the device, serving as the data foundation for subsequent data preprocessing, surrogate model training, and optimization solutions. Since design parameter optimization is usually performed before device manufacturing, and evolutionary optimization algorithms require a large number of function evaluations during the search process, high-fidelity physical field simulation is used for offline evaluation of the samples. This allows for the establishment of a mapping relationship between design variables and performance indicators without actual fabrication. For multi-physics coupled devices, a single simulation can consume a significant amount of computation time. When the sample size is large, the overall computational burden increases significantly. Therefore, this step typically uses a relatively limited number of initial samples to obtain the most effective training data possible.

[0031] High-fidelity physics simulations are established using the finite element multiphysics (FEM) approach. Based on the microphone's working mechanism, the FEM simulation calculates the response of the sample design under operating conditions by jointly modeling the device structure, material properties, external excitations, and boundary constraints. This type of simulation is not limited to a single physical domain but selects the appropriate physics analysis type based on the target objective. For example, for targets related to structural safety or mechanical load-bearing capacity, solid mechanics analysis can be used to obtain stress distribution and displacement response; for targets related to dynamic characteristics, characteristic frequency analysis can be used to obtain natural frequencies; for targets related to thermal stability, thermal analysis can be used to obtain temperature rise distribution; and for targets related to driving, detection, or impedance characteristics, electrical analysis can be combined to obtain current, voltage, capacitance changes, impedance changes, or other electrical responses. At least two of the stress, displacement, natural frequencies, temperature rise, and electrical responses can be output as the true evaluation results of the optimization objective.

[0032] The modeling content of finite element multiphysics simulation is matched with the specific mechanism of the microphone. When the target value involves structural response, the simulation model includes at least the device geometry model, material elastic parameters, mass parameters, and boundary fixation relationships. When the target value involves dynamic performance, characteristic frequencies are further solved based on the aforementioned model. When the target value involves temperature rise or thermal effects, heat conduction conditions and thermal boundaries are further introduced into the simulation model. When the target value involves electrical response, electrical excitation, current paths, potential distribution, or equivalent electrical boundaries are further introduced into the simulation model. For DC-driven devices with relatively small temperature differences, thermal analysis can focus on Joule heating and heat conduction. Under microscale conditions, thermal radiation has a smaller impact, and thermal conduction can dominate. For devices employing thermal deformation mechanisms, differences in cross-sectional dimensions lead to different temperature rises in different regions, resulting in differential expansion and structural deformation. These mechanisms can also be incorporated into the unified solution of finite element multiphysics simulation.

[0033] To ensure the reliability of the target values ​​in the simulation, appropriate mesh generation is necessary. An overly coarse mesh may lead to large errors or even solution failure; while an overly fine mesh can improve local accuracy, it significantly increases computation time. Therefore, mesh settings need to strike a balance between convergence, consistency, and computational efficiency, ensuring they adequately reflect the true response of the design sample while avoiding excessive time consumption in the sample evaluation process. The computation time for a single simulation may vary significantly for design objects with different degrees of physical coupling and structural complexity; more complex designs with more coupled physics typically require more solution time.

[0034] After the simulation of the sample design is completed, the obtained target values ​​are the true evaluation results, not approximate prediction results. The true evaluation results can be directly mapped to the optimization objectives; for example, the displacement objective corresponds to the structural displacement output, the frequency objective corresponds to the characteristic frequency output, the mechanical reliability objective corresponds to the stress output, the thermal stability objective corresponds to the temperature rise output, and the electrical performance objective corresponds to the electrical response output. If a sample design corresponds to multiple optimization objectives, multiple target values ​​are recorded simultaneously for the same sample, forming multi-output training sample data. If the optimization problem includes performance constraints, the evaluation quantities related to the constraints can also be read synchronously in this step to determine whether the sample falls within the feasible region. In this way, each sample design in the initial sample set can form a one-to-one correspondence between the design variable inputs and the target value outputs, providing basic data for subsequently building surrogate models for each optimization objective.

[0035] This step emphasizes performing high-fidelity evaluations on each initial sample, rather than directly replacing real simulations with approximate models. This is because subsequent surrogate models essentially fit real performance based on a limited number of high-fidelity samples, and the authenticity and representativeness of the initial samples directly determine the accuracy of subsequent predictions. While the number of high-fidelity simulation samples is limited, their target values ​​have high reliability and can provide a realistic calibration during subsequent training. This approach preserves the ability of finite element multiphysics simulation to characterize complex coupled responses while creating conditions for reducing large-scale repetitive simulations later.

[0036] In step S103, the target value is subjected to various preprocessing steps to form multiple training datasets. Based on each training dataset, an artificial intelligence agent model for the optimization target is constructed. The current agent model is selected according to a preset accuracy evaluation rule. The current agent model is used to output the target prediction value and prediction uncertainty of the candidate design.

[0037] The various preprocessing methods include: performing normalization, logarithmic transformation, and logarithmic transformation followed by normalization on the target value; for the target value after normalization, a positive shift is used to keep the preprocessed target value positive, so as to reduce the impact of dimensional differences and numerical ranges between different optimization targets on the accuracy of proxy modeling.

[0038] The artificial intelligence agent model is a Gaussian process regression model established for each optimization objective. The Gaussian process regression model uses the squared exponential covariance kernel to characterize the correlation between design samples and determines the model hyperparameters by maximizing the marginal likelihood, so as to output the target prediction mean and standard deviation corresponding to each candidate design.

[0039] The preset accuracy evaluation rules include: performing cross-validation on the AI ​​agent models corresponding to each training dataset, calculating the average absolute percentage error of each optimization objective, and combining the agent models with the lowest prediction error as the current agent model.

[0040] Specifically, in step S103, the target value obtained in step S102 undergoes various preprocessing steps to form multiple corresponding training datasets with different scales. The purpose of this processing is to reduce the impact of dimensional differences and numerical ranges between different optimization objectives on the accuracy of surrogate modeling, and to compare the effects of different preprocessing methods on prediction performance. Three preprocessing methods are used: normalization, logarithmic transformation, and logarithmic transformation followed by normalization. Normalization can be expressed as: ; in, Represents the original target value. and These represent the minimum and maximum values ​​of the target in the initial sample, respectively. This processing scales the target values ​​to a uniform range, which helps to reduce the order-of-magnitude differences between different targets. Considering that results exceeding the initial sample range may occur during subsequent optimization, some target values ​​after normalization may become negative, thus affecting the stability of the surrogate model evaluation and optimization process. Therefore, a positive shift can be performed on the normalized results to keep all preprocessed results positive, thereby ensuring that the value range falls within the positive range. Logarithmic transformation is used to compress the distribution of target values ​​with large spans, making the modeling of targets that vary across orders of magnitude smoother. Normalization after logarithmic transformation balances order-of-magnitude compression and range uniformity, and is suitable for situations where the numerical ranges of multiple targets differ significantly. For design objects with different complexities and different combinations of targets, the preprocessing method has a significant impact on prediction accuracy, especially when the targets span multiple orders of magnitude. The combination of normalization and logarithmic transformation is often more beneficial to improving prediction performance.

[0041] After multiple training datasets are generated, AI surrogate models for each optimization objective are constructed based on each training dataset. The AI ​​surrogate models employ Gaussian process regression. Gaussian process regression is suitable for handling complex, nonlinear mapping relationships between design variables and target values. Its basic idea is to treat function values ​​as a set of random variables following a joint Gaussian distribution. For noisy observation data, the objective function can be expressed as: ; in, Indicates design variable input, Indicates the observed target value. Represents the noise term. Function With noise They all follow a Gaussian distribution. Gaussian process regression can not only output the predicted values ​​corresponding to the candidate designs, but also provide the prediction uncertainty, which is usually expressed as variance or standard deviation. This makes it suitable for uncertainty determination in the subsequent optimization stage.

[0042] Gaussian process regression models use the covariance function to characterize the correlation between different sample points. The covariance function uses a squared exponential kernel, and its expression is: ; in, Indicates process variance. This represents the length scale hyperparameter. This represents the squared Euclidean distance between two input samples. The squared exponential kernel reflects that the closer the sample points are, the stronger the correlation of their function values; the farther apart they are, the weaker the correlation. Using this kernel function, the covariance matrix between training samples and between training samples and the sample to be predicted can be constructed. For new input... The predicted mean and predicted variance of the Gaussian process regression are as follows: ; ; in, This represents the covariance matrix between the new input and the training input. This represents the covariance matrix between the training inputs. This represents the covariance matrix between the training input and the new input. Let I represent the covariance matrix among the new inputs, and let I represent the identity matrix. This represents the training target value. The predicted mean can be used as the target predicted value for the candidate design, and the predicted variance or the standard deviation derived from it can be used as the predicted uncertainty of the candidate design. In this way, the surrogate model can both simulate the performance response of the real model and quantify the reliability of the prediction result.

[0043] Model hyperparameters are determined by maximizing the marginal likelihood. The marginal likelihood can be written as: ; in, ; This represents the set of hyperparameters, including length scale, process variance, and noise variance. By maximizing the marginal likelihoods mentioned above, the kernel function parameters can be better matched to the distribution of the training data, thereby improving prediction accuracy and the reasonableness of uncertainty estimation. Since there is usually more than one optimization objective, a Gaussian process regression model is trained for each optimization objective, forming a combination of surrogate models split by objective. This avoids interference between the statistical characteristics of different objectives and preserves the data distribution characteristics of each objective.

[0044] After all proxy models have been trained, the current proxy model is selected according to the preset accuracy evaluation rules. Accuracy evaluation uses cross-validation, and five-fold cross-validation can be used to assess the predictive ability of the proxy models corresponding to each training dataset. The evaluation metric is the mean absolute percentage error, expressed as: ; in, Represents the actual value. Let represent the predicted value, and n represent the number of samples. The mean absolute percentage error (MASE) is used to represent the prediction error as a percentage, facilitating comparison between targets at different scales. The MASE is calculated for each surrogate model trained using each preprocessing method. The surrogate models with the lowest prediction errors are then combined and used as the current surrogate model in subsequent optimization stages. This approach avoids pre-defining a specific preprocessing method and instead uses cross-validation results to filter different scaling strategies at the data level, making the current surrogate model more closely match the current target distribution.

[0045] Once the surrogate model is determined, candidate designs in the design space can be quickly evaluated. Compared with high-fidelity physics simulation, the surrogate model predicts new design points significantly faster and can output the target prediction values ​​and prediction uncertainties for each candidate design in a shorter time. Therefore, it is suitable for embedding into subsequent multi-objective optimization processes as an evaluation function.

[0046] In step S104, the current surrogate model is embedded into a multi-objective evolutionary optimization process, and the candidate designs are non-dominated, constrained, and diversity-preserving based on the target prediction values ​​to obtain a candidate non-dominated solution set.

[0047] The multi-objective evolutionary optimization process employs a non-dominated sorting genetic algorithm, which generates new candidate designs through parent selection, crossover, and mutation operations, and retains individuals by combining non-dominated sorting and crowding distance. During the individual retention process, design variable boundary constraints, performance threshold constraints, inequality constraints, and equality constraints are simultaneously evaluated to output a non-dominated solution set.

[0048] In step S104, the current surrogate model is embedded into the multi-objective evolutionary optimization process. This allows the fitness evaluation of candidate designs to no longer directly rely on high-fidelity physics simulation, but instead, the current surrogate model rapidly predicts the objective values ​​for each candidate design. This approach replaces the computationally expensive objective evaluation process with a more computationally efficient approximate evaluation process, enabling the optimizer to perform a large number of function evaluations within a limited time, making it suitable for iterative searching the microphone design space. Since the current surrogate model can output corresponding objective prediction values ​​for each candidate design, the multi-objective evolutionary optimization process can directly compare, filter, and retain candidate designs based on the prediction results.

[0049] The multi-objective evolutionary optimization process employs a non-dominated sorting genetic algorithm. This algorithm is suitable for handling multi-objective optimization problems and design scenarios involving continuous, discrete, and mixed variables. It does not impose strict restrictions on the convexity, differentiability, or constraint form of the objective function, thus adapting to the search requirements of various parameter combinations in microphone design. During the optimization process, the best-performing individuals in the current population participate in parent selection, followed by crossover and mutation operations to generate new candidate designs. Crossover is used to recombine design variables among different parent individuals to form new parameter combinations; mutation is used to perturb some design variables to expand the search range and prevent the search process from prematurely concentrating on local regions. Through repeated iterations, the design individuals in the population gradually achieve a balance between convergence and dispersion, thereby forming a candidate solution set oriented towards multiple objectives.

[0050] The merits of candidate designs are compared using a non-dominated ranking mechanism. When multiple optimization objectives exist simultaneously, if a candidate design is superior to another candidate design in at least one objective and no worse than the other in the remaining objectives, then the candidate design can be considered to dominate the other candidate design. Based on this comparison, individuals within the population can be divided into multiple non-dominated levels, with individuals at higher priority levels exhibiting better overall objective performance. Through non-dominated ranking, the trade-offs between multiple objectives can be preserved without simply merging them into a single index, ensuring that the optimization process output is not limited to a single design point but forms a non-dominated solution set with engineering selection significance.

[0051] In the individual retention process, crowding distance is introduced to maintain the diversity of the solution set. Crowding distance characterizes the density of an individual among its neighbors in the target space. For multiple individuals in the same non-dominated layer, individuals with larger crowding distances are usually located in sparser regions, which is more conducive to maintaining a uniform distribution of the solution set in the target space; individuals with smaller crowding distances are more likely to cluster with surrounding individuals. By combining non-dominated ordering and crowding distance for individual retention, we can ensure that the solution set converges to the Pareto front while avoiding excessive concentration of retained candidate designs in local regions, thereby improving the coverage and diversity of the candidate non-dominated solution set.

[0052] Throughout the optimization iteration process, constraint determination is performed simultaneously. Constraint determination includes at least design variable boundary constraints, performance threshold constraints, inequality constraints, and equality constraints. Design variable boundary constraints limit the values ​​of each design variable to within preset upper and lower bounds, ensuring that candidate designs always remain within the design space. Performance threshold constraints ensure that the predicted performance indicators meet application requirements, such as lower frequency limits, upper response limits, upper temperature rise limits, upper stress limits, or other performance boundaries. Inequality and equality constraints express general limiting relationships in the design problem. Candidate designs that satisfy the above constraints are considered feasible, while those that do not satisfy the constraints have their priority reduced or are eliminated during the ranking and retention process. This ensures that the output results are competitive not only at the target value level but also meet actual design requirements.

[0053] After embedding the optimizer into the current surrogate model, the non-dominated sorting genetic algorithm evaluates individuals in the population based on the target prediction value in each generation. It then retains individuals according to the non-dominated level and crowding distance, generating the next generation's population. As generational iterations progress, the population gradually clusters towards desirable regions that satisfy the constraints, while maintaining the dispersion of the solution set in the target space. After a predetermined number of generations or when the stopping condition is met, a candidate non-dominated solution set is obtained, consisting of multiple non-dominated candidate designs. This candidate non-dominated solution set reflects the different trade-offs between multiple mutually constraining objectives and can serve as the basis for further screening, supplementary evaluation, and result verification.

[0054] In step S105, based on the prediction uncertainty, candidate designs with prediction uncertainties higher than a preset threshold are selected from the candidate non-dominated solution set. The high-fidelity physical field simulation is performed on the selected candidate designs, and the newly added simulation data is incorporated into the training data corresponding to the current surrogate model so as to update the current surrogate model online during the optimization process.

[0055] The online update includes: in each generation of the multi-objective evolutionary optimization process, firstly, using the current surrogate model to predict the target prediction value and prediction uncertainty of individuals in the population, then selecting candidate designs with prediction uncertainties higher than the preset threshold from those that satisfy the constraints and belong to non-dominated solutions, in descending order of prediction uncertainty, and performing the high-fidelity physics simulation; each time new simulation data is obtained, the new simulation data is immediately incorporated into the training data and the current surrogate model is retrained, and then the candidate design screening for the current generation continues; when there are no candidate designs that satisfy the preset threshold in the current generation, the unused high-fidelity physics simulation times of the current generation are transferred to subsequent generations.

[0056] Specifically, in step S105, the prediction uncertainty output by the current surrogate model is used to further screen the candidate non-dominated solution set. Candidate designs with prediction uncertainties exceeding a preset threshold are identified as candidates for supplementary evaluation, and high-fidelity physical field simulations are performed on these candidate designs. This process does not recalculate all candidate designs, but only supplements the evaluation of design points that meet the constraints, have entered the range of non-dominated solutions, and have high uncertainties. Its purpose is to concentrate high-fidelity simulation resources on more promising areas where the model has insufficient control, thereby reducing invalid simulations and improving the modeling accuracy of key areas. Static surrogate models are established with fixed initial samples, which can easily lead to insufficient sampling in small areas where the optimal solution may be concentrated; online updates, on the other hand, continuously supplement data in these areas, improving the prediction reliability of those areas.

[0057] The prediction uncertainty is given by the Gaussian process regression model and is expressed as variance or standard deviation. Within the intergenerational cycle of multi-objective evolutionary optimization, the surrogate model simultaneously outputs the target prediction value and prediction uncertainty to each individual in the population. It then selects the candidate design with the highest uncertainty from the non-dominated solutions that satisfy the constraints for high-fidelity physics simulation, but this selection is predicated on the uncertainty of the candidate design exceeding a preset threshold. Using non-dominance, constraint satisfaction, and a large standard deviation as screening conditions ensures that new samples come from more promising search regions, rather than from ordinary regions that contribute little to the optimization. This dynamic sampling strategy essentially guides the direction of sample replenishment during the optimization process using uncertainty. The preset threshold serves two purposes. First, it ensures that high-fidelity physics simulations are triggered only when the surrogate model's uncertainty is high, thus avoiding repeated evaluations of samples that only provide a small amount of new information. Second, since the kernel function configuration of the Gaussian process regression model typically results in low standard deviations near training samples, indicating that the model already has high confidence in that region, the preset threshold prevents these well-learned design points from being repeatedly selected. Thus, online updates do not average out all candidate designs, but rather perform targeted corrections to the most uncertain and effective regions of the surrogate model. Online updates are performed sequentially within each generation. Upon entering a generation, the current surrogate model first predicts the individual population. Then, from candidate designs that meet the constraints and belong to non-dominated solutions, they are checked one by one in descending order of prediction uncertainty. When a candidate design meets the uncertainty threshold, a high-fidelity physical field simulation is performed on that candidate design to obtain a new target value. The new target value, along with the corresponding design variables, is then incorporated into the training data, and the current surrogate model is immediately retrained. The updated surrogate model is then used to continue the screening and judgment of the remaining candidate designs in the current generation. This sequential update method allows each new simulation data to immediately affect the prediction results and uncertainty distribution of subsequent candidate designs. Especially when the model uncertainty is high in the early stages of optimization, each supplementary sample may significantly correct the prediction results. Therefore, point-by-point updates are more conducive to timely correction of the search direction than centralized updates after the end of a generation.

[0058] When multiple high-fidelity physics simulations are planned within a generation, the method of evaluating each sample sequentially is still adopted, rather than completing all new simulations at once and then updating the model uniformly. Although this approach increases the number of training and prediction rounds for the surrogate model, the training and prediction costs of the surrogate model are lower than those of high-fidelity physics simulations, thus reducing the additional computational burden. After completing all new simulations for the current generation, the updated prediction results are then used to guide the non-dominated sorting genetic algorithm in sorting and retaining individuals entering the next generation. Therefore, the online update process is not an additional step independent of the optimization process, but rather embedded within intergenerational evolution, forming a closed loop with candidate design selection, sorting, and retention.

[0059] When no candidate design meets the preset threshold condition in the current generation, it indicates that the prediction uncertainty of feasible non-dominated solutions in the current generation is generally low, and the surrogate model has a high degree of learning in this region. At this time, the unused high-fidelity physics simulations in the current generation are no longer forcibly consumed, but are reserved for subsequent generations. This avoids making evaluations with low information gain in order to exhaust the budget, and allows subsequent generations to still use the remaining simulation resources when new high-uncertainty valid solutions appear. The flowchart reflects this scheduling logic by reserving simulation counts, that is, when no candidate design is found in the current generation that meets the non-dominated, constraint-satisfied, and standard deviation exceeding the threshold, the remaining simulation counts are transferred to subsequent generations for continued use.

[0060] Through the aforementioned online update method, the surrogate model can continuously converge towards more critical design regions during the optimization process. This results in the optimizer being less likely to be misled into suboptimal directions by early, inaccurate surrogate predictions, reducing the deviation between the predicted and true solution sets, and leading to a higher number and quality of effective non-dominated solutions. For scenarios with significant coupling between objectives, strict constraints, or discrete variables in the design, online learning is even more effective in improving optimization accuracy and search efficiency.

[0061] In step S106, after the stopping condition is met, the high-fidelity physical field simulation verification is performed on the optimized non-dominated solution, the solution that satisfies the constraint condition is retained as the microphone design optimization result, and the corresponding Pareto solution set is output.

[0062] The verification process includes: performing high-fidelity physical field simulations on each of the optimized non-dominated solutions to obtain the true target value; determining the non-dominated solutions that still satisfy the constraints after verification as valid non-dominated solutions; evaluating the reliability of the optimization based on the average absolute percentage error between the true target value and the predicted target value; and then using the hypervolume index to evaluate the quality of the solution set, wherein the reference point of the hypervolume index is determined based on the worst target point in the joint solution set composed of multiple solution sets to be compared.

[0063] Specifically, in step S106, after the multi-objective evolutionary optimization process meets the stopping condition, high-fidelity physical field simulation is performed on each of the non-dominated solutions obtained through optimization to obtain the true objective value corresponding to each non-dominated solution. The reason for setting up this verification process is that the non-dominated solutions are predicted based on surrogate models during the optimization stage. Although surrogate models can significantly reduce evaluation costs, their output is still an approximation. Only after recalculation through high-fidelity physical field simulation can the true performance of each candidate solution be accurately determined. Therefore, before outputting the final microphone design optimization result, it is necessary to remap the non-dominated solutions selected by the surrogate model onto the true physical evaluation results to eliminate the influence of prediction bias on the final design selection.

[0064] During the review process, high-fidelity physical field simulations are performed on each non-dominated solution, and the actual target value corresponding to the optimization objective is read. If a non-dominated solution still satisfies the constraints after review, it is identified as a valid non-dominated solution; if the actual target value no longer satisfies the constraints after review, it is removed from the final result. The solutions retained in this way not only exhibit non-dominated relationships during the surrogate model evaluation stage but also satisfy performance boundaries and design constraints under the real physical field evaluation, thus serving as practically usable optimization results. The final output is not a single design point but a Pareto solution set composed of multiple valid non-dominated solutions, used to characterize different trade-offs between multiple mutually constraining objectives, thereby providing designers with multiple alternative solutions.

[0065] After verification, the mean absolute percentage error (MASE) between the actual and predicted target values ​​can be used to evaluate the reliability of the optimization. MASE measures the deviation between the surrogate model's predictions and the actual simulation results. A smaller error indicates that the surrogate model's predictions of non-dominated solutions are closer to reality during the optimization phase, and the optimization results are more reliable. A larger error indicates that the surrogate model's estimation of this type of solution has a significant bias during the optimization phase, and the optimization results are more significantly affected by approximation errors. Because this index is expressed as a relative error, it facilitates unified comparisons between targets at different scales, making it suitable for evaluating the reliability of target predictions after verification.

[0066] In addition to evaluating the predictive reliability of individual solutions, the quality of the entire solution set can be evaluated using the hypervolume index. The hypervolume index characterizes the volume covered by the solution set and the reference point in the target space, and its expression is: ; Where A represents the solution set to be evaluated. Indicates the Lebesgue measure. Let represent any solution in the solution set. Indicates a reference point. This indicates that point x is in the solution space. With reference point Between and subject to resolution The dominant region. A larger hypervolume index indicates a wider coverage of the solution set in the target space, generally implying better convergence and distribution quality. For cases where multiple solution sets need to be compared, the reference point can be determined by the worst target point in the joint solution set composed of the compared solutions. This reference point setting allows for comparison of different solution sets under a unified benchmark, avoiding the impact of inconsistent reference point selection on the evaluation results.

[0067] In specific evaluation, the hypervolume index of the solution set predicted by the surrogate model at the end of optimization can be calculated separately, as well as the hypervolume index of the actual solution set verified by high-fidelity physics simulation. If the hypervolume index of the predicted solution set is significantly higher than that of the verified actual solution set, it indicates that the quality of the solution set was overestimated during the optimization stage; if the two are close, it indicates that the surrogate model's judgment of the overall quality of the solution set during the optimization stage is relatively accurate. By combining the mean absolute percentage error with the hypervolume index, we can evaluate both the reliability of individual target predictions and the overall quality of the entire Pareto solution set, thus more comprehensively determining whether the optimization results have practical application value.

[0068] The verification and filtering process after the stopping condition is met ensures that the final Pareto solution set is based on real physical field results, rather than solely on surrogate model predictions. This reduces the likelihood of spurious optimal solutions due to surrogate errors entering the final result and improves the usability and reliability of the output solution set in engineering applications. This verification process is particularly beneficial for microphone optimization tasks with strict objective constraints, a large number of objectives, or complex design spaces, in ensuring the quality of the final result.

[0069] Based on the same line of thought, such as Figure 2 As shown, an AI-based microphone design optimization system is provided, including: The design space construction module 201 is used to establish a design space based on the structural parameters, material parameters, working condition parameters and performance requirements of the microphone to be optimized, and to determine the design variables, multiple mutually constraining optimization objectives and the constraints corresponding to the optimization objectives corresponding to the design space. The initial sample simulation module 202 is used to generate an initial sample set within the design space and perform high-fidelity physical field simulation on each initial sample set to obtain the target value corresponding to each initial sample. The surrogate model construction module 203 is used to perform various preprocessing on the target value to form multiple training datasets, construct artificial intelligence surrogate models for the optimization target based on each training dataset, and select the current surrogate model according to a preset accuracy evaluation rule. The current surrogate model is used to output the target prediction value and prediction uncertainty of the candidate design. The multi-objective optimization solution module 204 is used to embed the current surrogate model into the multi-objective evolutionary optimization process, and perform non-dominated ranking, constraint determination and diversity preservation on the candidate designs based on the target prediction values ​​to obtain a candidate non-dominated solution set. The adaptive update module 205 is used to select candidate designs with prediction uncertainties higher than a preset threshold from the candidate non-dominated solution set according to the prediction uncertainty, perform the high-fidelity physical field simulation on the selected candidate designs, and incorporate the newly added simulation data into the training data corresponding to the current proxy model, so as to update the current proxy model online during the optimization process. The result verification output module 206 is used to perform the high-fidelity physical field simulation verification on the optimized non-dominated solution after the stopping condition is met, retain the solution that satisfies the constraint condition as the microphone design optimization result, and output the corresponding Pareto solution set.

[0070] Compared with existing technologies, this system combines artificial intelligence agent modeling with a multi-objective optimization process and introduces a dynamic update mechanism for candidate solutions with high uncertainty during the optimization process. This reduces the number of calls to high-fidelity physical field simulations while improving the prediction reliability of key areas, making the optimization process more suitable for handling complex scenarios in microphone design where multiple performance indicators are mutually constrained and need to meet constraints. At the same time, this invention can output a verified non-dominated solution set, providing designers with multiple sets of candidate designs with engineering significance, rather than just a single result. Therefore, it helps to improve design efficiency, enhance the reliability of optimization results, and increase the flexibility of selecting the final solution under practical application requirements.

[0071] It should be noted that although several modules or units of the device for performing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to exemplary embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0072] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0073] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A microphone design optimization method based on AI, characterized in that, The method includes: A design space is established based on the structural parameters, material parameters, operating parameters, and performance requirements of the microphone to be optimized, and the design variables, multiple mutually constraining optimization objectives, and the constraints corresponding to the optimization objectives are determined. An initial sample set is generated within the design space, and a high-fidelity physical field simulation is performed on each of the initial sample sets to obtain the target value corresponding to each initial sample. The target value is preprocessed in various ways to form multiple training datasets. Based on each training dataset, an artificial intelligence agent model for the optimization target is constructed. The current agent model is selected according to a preset accuracy evaluation rule. The current agent model is used to output the target prediction value and prediction uncertainty of the candidate design. The current proxy model is embedded into a multi-objective evolutionary optimization process. Based on the target prediction value, the candidate designs are non-dominated, constrained, and diversity is preserved to obtain a set of candidate non-dominated solutions. Based on the predicted uncertainty, candidate designs with predicted uncertainty higher than a preset threshold are selected from the candidate non-dominated solution set. The selected candidate designs are subjected to high-fidelity physical field simulation, and the newly added simulation data is incorporated into the training data corresponding to the current proxy model so as to update the current proxy model online during the optimization process. After the stopping condition is met, the high-fidelity physical field simulation verification is performed on the non-dominated solution obtained by optimization. The solution that satisfies the constraint condition is retained as the microphone design optimization result, and the corresponding Pareto solution set is output.

2. The AI-based microphone design optimization method according to claim 1, characterized in that, The design variables include at least one of the following: microphone geometric parameters, material parameters, operating condition parameters, and boundary condition parameters. The design variables may be in continuous, discrete, or hybrid form. The initial sample set is generated by combining Latin hypercube sampling with the maximum-minimum distance criterion, so that the initial sample set has uniform coverage and spatial filling within the design space.

3. The AI-based microphone design optimization method according to claim 1, characterized in that, The various preprocessing methods include: performing normalization, logarithmic transformation, and logarithmic transformation followed by normalization on the target value; for the target value after normalization, a positive shift is used to keep the preprocessed target value positive, so as to reduce the impact of dimensional differences and numerical ranges between different optimization targets on the accuracy of proxy modeling.

4. The AI-based microphone design optimization method according to claim 1, characterized in that, The artificial intelligence agent model is a Gaussian process regression model established for each optimization objective. The Gaussian process regression model uses the squared exponential covariance kernel to characterize the correlation between design samples and determines the model hyperparameters by maximizing the marginal likelihood, so as to output the target prediction mean and standard deviation corresponding to each candidate design.

5. The AI-based microphone design optimization method according to claim 1, characterized in that, The preset accuracy evaluation rules include: performing cross-validation on the AI ​​agent models corresponding to each training dataset, calculating the average absolute percentage error of each optimization objective, and combining the agent models with the lowest prediction error as the current agent model.

6. The AI-based microphone design optimization method according to claim 1, characterized in that, The multi-objective evolutionary optimization process uses a non-dominated sorting genetic algorithm to generate new candidate designs through parent selection, crossover, and mutation operations, and combines non-dominated sorting and crowding distance to preserve individuals. During the individual retention process, the design variable boundary constraints, performance threshold constraints, inequality constraints, and equality constraints are simultaneously determined to output a non-dominated solution set.

7. The AI-based microphone design optimization method according to claim 6, characterized in that, The online update includes: in each generation of the multi-objective evolutionary optimization process, firstly, using the current surrogate model to predict the target prediction value and prediction uncertainty of individuals in the population, then selecting candidate designs with prediction uncertainties higher than the preset threshold from those that satisfy the constraints and belong to non-dominated solutions, in descending order of prediction uncertainty, and performing the high-fidelity physics simulation; each time new simulation data is obtained, the new simulation data is immediately incorporated into the training data and the current surrogate model is retrained, and then the candidate design screening for the current generation continues; when there are no candidate designs that satisfy the preset threshold in the current generation, the unused high-fidelity physics simulation times of the current generation are transferred to subsequent generations.

8. The AI-based microphone design optimization method according to claim 1, characterized in that, During the review process, the following should be included: The high-fidelity physical field simulation is performed on each of the optimized non-dominated solutions to obtain the true target value; The non-dominated solutions that still satisfy the constraints after review are determined to be valid non-dominated solutions. The reliability of the optimization is evaluated based on the average absolute percentage error between the true target value and the predicted target value. Then, the quality of the solution set is evaluated using the hypervolume index, wherein the reference point of the hypervolume index is determined based on the worst target point in the joint solution set composed of multiple solution sets to be compared.

9. The AI-based microphone design optimization method according to claim 1, characterized in that, The high-fidelity physical field simulation is a finite element multiphysics simulation. The finite element multiphysics simulation is established based on the working mechanism of the microphone and is used to output at least two of the following: stress, displacement, natural frequency, temperature rise, and electrical response, as the true evaluation result of the optimization objective.

10. An AI-based microphone design optimization system, characterized in that, include: The design space construction module is used to establish a design space based on the structural parameters, material parameters, operating parameters and performance requirements of the microphone to be optimized, and to determine the design variables, multiple mutually restrictive optimization objectives and the constraints corresponding to the optimization objectives. The initial sample simulation module is used to generate an initial sample set within the design space and perform high-fidelity physical field simulation on each initial sample set to obtain the target value corresponding to each initial sample. The proxy model construction module is used to perform various preprocessing on the target value to form multiple training datasets, construct artificial intelligence proxy models for the optimization target based on each training dataset, and select the current proxy model according to the preset accuracy evaluation rules. The current proxy model is used to output the target prediction value and prediction uncertainty of the candidate design. The multi-objective optimization solution module is used to embed the current surrogate model into the multi-objective evolutionary optimization process, and perform non-dominated ranking, constraint determination and diversity preservation on the candidate designs based on the target prediction values ​​to obtain a set of candidate non-dominated solutions. An adaptive update module is used to select candidate designs with prediction uncertainties higher than a preset threshold from the candidate non-dominated solution set based on the prediction uncertainty, perform the high-fidelity physical field simulation on the selected candidate designs, and incorporate the newly added simulation data into the training data corresponding to the current proxy model, so as to update the current proxy model online during the optimization process. The result verification output module is used to perform the high-fidelity physical field simulation verification on the optimized non-dominated solution after the stopping condition is met, retain the solution that satisfies the constraint condition as the microphone design optimization result, and output the corresponding Pareto solution set.

Images