Simulation drive intelligent design method and system based on unified data flow

Abstract

The invention provides a simulation drive intelligent design method and system based on a unified data stream, and relates to the technical field of computer simulation design. The method comprises the steps that the number of design variables is obtained, and a physical value interval is set; determining the number of total samples; distributing the number of the training samples and the number of the test samples according to a preset proportion; an integrated simulation environment is imported, batch simulation is started, and various kinds of high-dimensional physical field data corresponding to each sample are obtained; determining scalar output response; determining a standardized scalar output response; automatically associating the design variable with the standardized scalar output response, and dividing the design variable and the standardized scalar output response into a training data set and a test data set; obtaining a plurality of trained agent models; obtaining a predicted value of the trained agent model; determining the confidence of the proxy model; and performing global optimization. According to the method and the device, simulation design can be optimized according to the number of samples, data processing and the confidence coefficient of the proxy model, and the efficiency and the accuracy of the simulation design are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer simulation design, and in particular to a simulation-driven intelligent design method and system based on unified data flow. BACKGROUND

[0002] In the related art, although the simulation speed is improved by joint simulation and reconfigurable distributed data flow, the influence of sample quantity, data processing and proxy model confidence on simulation design is not considered, that is, the simulation design cannot be optimized according to sample quantity, data processing and proxy model confidence.

[0003] The information disclosed in the background section of this application is only intended to deepen the understanding of the general background of the application and should not be regarded as acknowledging or implying in any form that this information constitutes prior art known to those skilled in the art. SUMMARY

[0004] The present application provides a simulation-driven intelligent design method and system based on unified data flow, which can solve the technical problem that the related art cannot optimize the simulation design according to sample quantity, data processing and proxy model confidence.

[0005] According to a first aspect of the present application, a simulation-driven intelligent design method based on unified data flow is provided, comprising: obtaining the number of design variables and setting the physical value interval for the design variables; determining the total number of samples according to the number of design variables and a preset sampling density factor, wherein the total samples include training samples and test samples; allocating the number of training samples and the number of test samples according to a preset ratio according to the total number of samples; importing the total samples into an integrated simulation environment, starting and performing batch simulation, and obtaining a plurality of high-dimensional physical field data corresponding to each sample; determining a scalar output response according to the high-dimensional physical field data; determining a standardized scalar output response according to the scalar output response; automatically associating the design variables with the standardized scalar output response and dividing them into a training data set and a test data set; training a plurality of proxy models through the training data set to obtain a plurality of trained proxy models; inputting the test samples into the trained proxy models to obtain the predicted values of the trained proxy models; determining the proxy model confidence according to the test data set and the predicted values; and performing global optimization according to the trained proxy model with the maximum proxy model confidence and the physical value interval.

[0006] Further, the total number of samples is determined based on the number of design variables and the preset sampling density factor, including: taking the number of design variables as the basic scale parameter, combining squaring with logarithmic scaling, multiplying with the preset sampling density factor, and then processing with an up-rounding function to determine the total number of samples.

[0007] Further, determining the scalar output response based on the high-dimensional physical field data includes: determining the values ​​and gradients of multiple samples and multiple physical fields at multiple spatial locations based on the high-dimensional physical field data; obtaining the unit control volume of multiple samples and multiple physical fields at multiple spatial locations; and determining the scalar output response based on the values, the gradients, and the unit control volume.

[0008] Further, determining the scalar output response based on the numerical value, the gradient, and the unit control volume includes: using the ratio of the gradient magnitude of each unit's physical field to the maximum gradient magnitude of the entire field as a weighting coefficient for the numerical value; and performing a weighted average of the unit control volume based on the weighting coefficient of the numerical value and the numerical value to determine the scalar output response.

[0009] Further, determining the standardized scalar output response based on the scalar output response includes: obtaining the mean, standard deviation, maximum value, and minimum value of each scalar output response across all samples; setting preset upper and lower bounds of the target normalization interval; and determining the standardized scalar output response based on the mean, the standard deviation, the maximum value, the minimum value, the upper and lower bounds of the target normalization interval, and the scalar output response.

[0010] Further, determining the standardized scalar output response based on the mean, the standard deviation, the maximum value, the minimum value, the upper and lower bounds of the target normalization interval, and the scalar output response includes: standardizing the scalar output response and then mapping it to the target normalization interval.

[0011] Further, the surrogate model confidence is determined based on the test standardized scalar output response, the predicted maximum value, the test minimum value, and the predicted value, including: determining the surrogate model confidence by taking the root mean square of the difference between the predicted value and the test standardized scalar output response and mapping the range to between 0 and 1.

[0012] According to a second aspect of the present invention, a simulation-driven intelligent design system based on a unified data flow is provided, comprising: a design variable module for acquiring the number of design variables and setting physical value ranges for the design variables; a total sample quantity module for determining the total sample quantity based on the number of design variables and a preset sampling density factor, wherein the total sample quantity includes training samples and test samples; an allocation module for allocating the number of training samples and the number of test samples according to a preset ratio based on the total sample quantity; a simulation module for importing the total sample quantity into an integrated simulation environment, starting and performing batch simulation to obtain multiple high-dimensional physical field data corresponding to each sample; and a scalar output response module for determining a scalar based on the high-dimensional physical field data. The system comprises the following modules: an output response module; a standardized scalar output response module for determining the standardized scalar output response based on the scalar output response; an association module for automatically associating the design variables with the standardized scalar output response and dividing them into training and test datasets; a trained surrogate model module for training multiple surrogate models using the training dataset to obtain multiple trained surrogate models; a prediction value module for inputting the test samples into the trained surrogate models to obtain the predicted values ​​of the trained surrogate models; a surrogate model confidence module for determining the confidence of the surrogate models based on the test dataset and the predicted values; and a global optimization module for performing global optimization based on the trained surrogate model with the highest confidence and the physical value range.

[0013] Technical Effects: According to this invention, by setting design variables and their physical value ranges, intelligent design can be performed without being limited to a specific design object. The number of total samples reduces waste or inadequacy caused by manual experience-based settings. Converting high-dimensional physical field data into scalar output responses and standardizing these responses better adapts to the requirements of various surrogate model training algorithms, improving the training efficiency of surrogate models. Automated process design and surrogate model training and evaluation processes reduce manual intervention. Global optimization using the trained surrogate model with the highest confidence level quickly finds the optimal design scheme, shortening simulation time and enhancing the efficiency and accuracy of simulation design. When determining the total number of samples, the number of design variables and a preset sampling density factor can be used to control computational resource input, intelligently determine the required number of simulations, and reduce waste or inadequacy caused by manual experience-based settings. When determining the scalar output response, numerical values, gradients, and cell control volumes can be used to determine the scalar output response. Based on adaptive spatial weights of the physical field gradient, automated and lossless compressed feature extraction of high-dimensional physical field data is achieved. Key regions with drastic changes in physical quantities are automatically identified and assigned higher weights, making the extracted scalar output response more reflective of the physical essence related to the design objectives. When determining the standardized scalar output response, the mean, standard deviation, maximum, minimum, upper and lower bounds of the target normalization interval, and the scalar output response can be used to determine the standardized scalar output response. Standardization and target normalization interval mapping of the scalar output response eliminate differences between different samples caused by factors such as units and numerical ranges. This better adapts to the requirements of various surrogate model training algorithms, improves the convergence speed and stability of surrogate models, and reduces surrogate model bias or training difficulties caused by inconsistent data scales. When determining the confidence level of a surrogate model, the trained surrogate model can be evaluated using test samples to determine the confidence level. This takes into account the characteristics and advantages of different surrogate models, reduces the potential limitations of a single surrogate model, and allows us to find the most suitable surrogate model for the current design problem, thereby improving the accuracy and reliability of the surrogate model.

[0014] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Other features and aspects of the invention will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0016] Figure 1 An exemplary flowchart of a simulation-driven intelligent design method based on unified data flow according to an embodiment of the present invention is shown.

[0017] Figure 2 A flowchart illustrating the calculation of a scalar output response according to an embodiment of the present invention is shown by way of example;

[0018] Figure 3 An exemplary flowchart illustrating the calculation of a normalized scalar output response according to an embodiment of the present invention is shown;

[0019] Figure 4 A flowchart illustrating the calculation of agent model confidence is shown as an example according to an embodiment of the present invention.

[0020] Figure 5 A block diagram of a simulation-driven intelligent design system based on a unified data flow according to an embodiment of the present invention is shown as an example. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0022] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0023] Figure 1An exemplary flowchart of a simulation-driven intelligent design method based on a unified data flow according to an embodiment of the present invention is shown. The method includes: Step S1, obtaining the number of design variables and setting physical value ranges for the design variables; Step S2, determining the total number of samples based on the number of design variables and a preset sampling density factor, wherein the total number of samples includes training samples and test samples; Step S3, allocating the number of training samples and the number of test samples according to a preset ratio based on the total number of samples; Step S4, importing the total number of samples into an integrated simulation environment, starting and performing batch simulation to obtain multiple high-dimensional physical field data corresponding to each sample; Step S5, based on the high-dimensional... Step S6: Determine the scalar output response based on the scalar output response; Step S7: Automatically associate the design variables with the standardized scalar output response and divide them into training datasets and test datasets; Step S8: Train multiple surrogate models using the training dataset to obtain multiple trained surrogate models; Step S9: Input the test samples into the trained surrogate models to obtain the predicted values ​​of the trained surrogate models; Step S10: Determine the confidence level of the surrogate models based on the test dataset and the predicted values; Step S11: Perform global optimization based on the trained surrogate model with the highest confidence level and the physical value range.

[0024] The simulation-driven intelligent design method based on unified data flow, according to embodiments of the present invention, allows for intelligent design not limited to specific design objects by setting design variables and their physical value ranges. By increasing the total number of samples, it reduces waste or inadequacy caused by manual experience-based settings. Converting high-dimensional physical field data into scalar output responses and standardizing these responses better adapts to the requirements of various surrogate model training algorithms, improving the training efficiency of surrogate models. Automated process design and surrogate model training and evaluation processes reduce manual intervention. Global optimization using the trained surrogate model with the highest confidence level quickly finds the optimal design scheme, shortening simulation time and enhancing the efficiency and accuracy of simulation design.

[0025] According to one embodiment of the present invention, in step S1, the desired goals of the entire design project are defined. For example, when designing a new type of automobile engine, the design goals may be to improve fuel efficiency, enhance power output, or reduce exhaust emissions. Experts in relevant fields are consulted to set relevant design variables. For example, in automobile engine design, variables that significantly affect engine performance, such as cylinder diameter, piston stroke, and fuel injector orifice diameter, can be set as design variables. All determined design variables are counted to obtain the number of design variables. For example, in the above automobile engine design example, if three design variables—cylinder diameter, piston stroke, and fuel injector orifice diameter—are determined, the number of design variables is 3. Various constraints faced during the design process are considered, such as physical limitations, safety requirements, and manufacturing process limitations, to set reasonable value ranges for the design variables. For example, in the optimized design of an aircraft wing, for the wing length L, considering the overall size of the aircraft and flight performance requirements, its value range can be 5m ≤ L ≤ 15m; for the airfoil thickness t, according to aerodynamic design requirements, its value range can be 0.1m ≤ t ≤ 0.5m. By designing variables and their physical value ranges, intelligent design can be performed without being limited to specific design objects.

[0026] According to an embodiment of the present invention, in step S2, the total number of samples is determined based on the number of design variables and a preset sampling density factor. The total number of samples includes training samples and test samples. The training samples are generated by the Latin hypercube sampling method, and the test samples are generated independently by the random sampling method. The preset sampling density factor reflects the density of sampling in the design space, for example, 1.8.

[0027] According to an embodiment of the present invention, step S2 includes: taking the number of design variables as the basic scale parameter, combining squaring with logarithmic scaling, multiplying with a preset sampling density factor, and then processing with an up-rounding function to determine the total number of samples.

[0028] According to one embodiment of the present invention, determining the total number of samples based on the number of design variables and a preset sampling density factor includes: determining the total number of samples N according to formula (1). (1), in, Here, k is the preset sampling density factor, and k is the number of design variables. To round up, both k and N are positive integers.

[0029] According to an embodiment of the present invention, in formula (1), The square of the number of design variables represents a scaling factor applied to the number of design variables. The more design variables there are, the more complex their interactions become, requiring more samples to capture the impact of these interactions on the design objectives. For example, when the number of design variables increases from two to three, the dimension of the design space increases from two to three dimensions, significantly increasing the possible combinations of variables. Therefore, the square operation is needed to appropriately increase the sample size. To Performing natural logarithmic calculations helps to adjust the values ​​after square amplification, reflecting the non-linear impact of the increase in the number of design variables on the sample size. Presetting the sampling density factor controls the allocation of computational resources. , and The total number of samples can be obtained by multiplying the results together and rounding them up.

[0030] In this way, the total number of samples can be determined by designing the number of variables and setting the sampling density factor. This allows for control over the input of computing resources, intelligent determination of the required number of simulations, and reduction of waste or inadequacy caused by manual experience settings.

[0031] According to one embodiment of the present invention, in step S3, the number of training samples and the number of test samples are allocated according to a preset ratio (e.g., 4:1), that is, 80% of the total number of samples (training samples) are used to train the agent model, and 20% of the total number of samples (test samples) are used to test the performance of the agent model.

[0032] According to one embodiment of the present invention, in step S4, a parallel batch simulation task is initiated in the simulation environment. The high-dimensional physical field data may include various physical field information such as displacement field, stress field, strain field, temperature field, and velocity field.

[0033] According to one embodiment of the present invention, in step S5, a scalar output response is determined based on the high-dimensional physical field data.

[0034] Figure 2 A flowchart illustrating the calculation of a scalar output response according to an embodiment of the present invention is shown.

[0035] According to an embodiment of the present invention, step S5 includes: step S51, determining the values ​​and gradients of multiple physical fields of multiple samples at multiple spatial locations based on the high-dimensional physical field data; step S52, obtaining the unit control volume of multiple physical fields of multiple samples at multiple spatial locations; and step S53, determining the scalar output response based on the values, the gradients, and the unit control volume.

[0036] According to one embodiment of the present invention, a spatial grid is set in the space of the physical field, which can be divided into multiple unit spaces. The coordinates of each unit space in the physical field are its spatial position. The spatial grid can be regular (e.g., uniform grid) or irregular (e.g., non-uniform grid) for discretizing the spatial region. For example, in finite element analysis, triangular or quadrilateral grids are usually used to discretize the structure, while in computational fluid dynamics, hexahedral or tetrahedral grids may be used to discretize the fluid region. The physical field values ​​at each sample's spatial position are extracted from the high-dimensional physical field data. For example, for stress field data, the average value of the stress components (e.g., normal stress, shear stress) at each sample's spatial position is extracted. The physical field gradient at each sample's spatial position is calculated using numerical differentiation methods (e.g., forward difference, backward difference, or central difference). The gradient represents the rate of change of the physical field in space and is crucial for understanding the distribution and variation of the physical field. For example, for temperature field data, calculating the temperature gradient can reveal the direction and intensity of heat transfer. The unit control volume is the volume of the unit space. The scalar output response is scalar data transformed from high-dimensional physical field data.

[0037] According to an embodiment of the present invention, step S53 includes: using the ratio of the physical field gradient magnitude of each unit to the maximum gradient magnitude of the entire field as a weighting coefficient of the numerical value, performing a weighted average of the unit control volume based on the weighting coefficient of the numerical value and the numerical value, and determining the scalar output response.

[0038] According to an embodiment of the present invention, determining the scalar output response based on the numerical value, the gradient, and the unit control volume includes: determining the scalar output response of the t-th sample and the s-th physical field according to formula (2). , (2), in, The unit control volume for the t-th sample, the s-th physical field, at the e-th spatial location. Let be the gradient of the s-th physical field of the t-th sample at the e-th spatial location. Let be the value of the s-th physical field of the t-th sample at the e-th spatial location. Let e ​​be the total number of discretized units of the t-th sample and the s-th physical field, ≤ And t, s, e and All are positive integers, and max is the function to find the maximum value.

[0039] According to an embodiment of the present invention, in formula (2), Let be the ratio between the magnitude of the gradient of the s-th physical field at the e-th spatial location of the t-th sample and the maximum magnitude of the gradient of the s-th physical field at multiple spatial locations of the t-th sample, representing the adaptive spatial weight of the physical field gradient. , and The product represents the output response of the t-th sample and the s-th physical field at the e-th spatial location. The larger the gradient change, the larger the physical field value, and the larger the unit control volume, the larger the output response. By averaging the output responses of the t-th sample and the s-th physical field at multiple spatial locations, a scalar output response can be obtained, thus converting high-dimensional physical field data into scalar data.

[0040] In this way, the scalar output response can be determined by numerical values, gradients, and unit control volumes. Based on the adaptive spatial weights of the physical field gradient, the automatic and lossless compression feature extraction of high-dimensional physical field data is realized. Key regions with drastic changes in physical quantities are automatically identified and given higher weights, so that the extracted scalar output response can better reflect the physical essence related to the design objectives.

[0041] According to an embodiment of the present invention, in step S6, a normalized scalar output response is determined based on the scalar output response.

[0042] Figure 3 A flowchart illustrating the calculation of a normalized scalar output response according to an embodiment of the present invention is shown.

[0043] According to an embodiment of the present invention, step S6 includes: step S61, obtaining the mean, standard deviation, maximum value and minimum value of each scalar output response in all samples; step S62, setting a preset upper and lower bound of the target normalization interval; step S63, determining the standardized scalar output response based on the mean, the standard deviation, the maximum value, the minimum value, the upper and lower bounds of the target normalization interval and the scalar output response.

[0044] According to one embodiment of the present invention, the mean reflects the average level of the scalar output response in the overall sample, and the standard deviation measures the dispersion of the scalar output response across all samples. The larger the standard deviation, the more dispersed the scalar output response. The scalar output response values ​​of all samples are arranged in ascending order to obtain the maximum and minimum values. A preset upper and lower bound for the target normalization interval is set, for example, [−1, 1]. Standardizing the scalar output response maps all physical responses to a unified dimensionless interval, eliminating the interference of dimensional differences on the training of the surrogate model.

[0045] According to an embodiment of the present invention, step S63 includes: using the ratio of the physical field gradient magnitude of each unit to the maximum gradient magnitude of the entire field as a weighting coefficient of the numerical value, performing a weighted average of the unit control volume based on the weighting coefficient of the numerical value and the numerical value, and determining the scalar output response.

[0046] According to an embodiment of the present invention, determining the standardized scalar output response based on the mean, the standard deviation, the maximum value, the minimum value, the upper and lower bounds of the target normalization interval, and the scalar output response includes: determining the standardized scalar output response of the scalar output response of the t-th sample according to formula (3). , (3), in, Let be the scalar output response of the t-th sample and the s-th physical field. Let be the mean of the scalar output response of the s-th physical field across all samples. Let be the standard deviation of the scalar output response of the s-th physical field across all samples. The upper bound of the preset target normalization interval, The lower bound of the pre-defined target normalization interval, Let be the maximum value of the scalar output response of the s-th physical field across all samples. Let be the minimum scalar output response of the s-th physical field across all samples. The t and s are both positive integers.

[0047] According to one embodiment of the present invention, in formula (3), Let be the ratio of the scalar output response of the t-th sample to the mean of the scalar output response of the s-th physics across all samples, and the standard deviation of the scalar output response of the s-th physics across all samples. This represents the standardization of the scalar output response. It is a positive number, for example, 0.0001. Scaling factor Let be the range of values ​​for the scalar output response of the s-th physics field across all samples, calculated numerically only. The length of the target normalization interval is determined by the scaling factor and... Multiplying, can By mapping to the target normalization interval, a standardized scalar output response is obtained, which can uniformly map all scalar output responses from different sources and with different physical meanings to the same numerical scale.

[0048] In this way, the standardized scalar output response can be determined by the mean, standard deviation, maximum value, minimum value, upper and lower bounds of the target normalization interval, and scalar output response. The scalar output response is then standardized and mapped to the target normalization interval, eliminating the differences between different samples caused by factors such as units and numerical ranges. This approach can better adapt to the requirements of various surrogate model training algorithms, improve the convergence speed and stability of the surrogate model, and reduce surrogate model bias or training difficulties caused by inconsistent data scales.

[0049] According to one embodiment of the present invention, in step S7, a matrix form (e.g., an array created using the NumPy library) is used to store the design variables and standardized scalar output responses. The correspondence between the design variables and the standardized scalar output responses is established through simulation results. In Python, a loop structure can be used to traverse all samples, matching and storing the design variable value of each sample with its corresponding standardized scalar output response value. Different identity labels can be assigned to training samples and test samples. These identity labels can identify whether a sample is a training sample. Through these identity labels, the design variables and standardized scalar output responses are automatically divided into training datasets and test datasets.

[0050] According to one embodiment of the present invention, in step S8, the surrogate model is a simplified mathematical model used to replace a complex, time-consuming high-fidelity simulation model. Different surrogate models have different data-driven modeling algorithms. For example, Gaussian process regression, neural networks, radial basis functions, etc. Data preprocessing is performed on the training dataset. Each surrogate model is trained using the training dataset. During training, the surrogate model adjusts its internal parameters according to the input design variables and the corresponding standardized scalar output response to minimize prediction error.

[0051] According to one embodiment of the present invention, in step S9, the test sample is input into the trained agent model to obtain the predicted value of the trained agent model.

[0052] According to one embodiment of the present invention, in step S10, the confidence level of the surrogate model is determined based on the test dataset and the predicted value.

[0053] Figure 4 A flowchart illustrating the calculation of agent model confidence is shown as an example according to an embodiment of the present invention.

[0054] According to an embodiment of the present invention, step S10 includes: step S101, obtaining the test normalized scalar output response of the scalar output response of multiple physical fields of multiple test samples based on the test dataset; step S102, obtaining the maximum predicted value and the minimum tested value of the test normalized scalar output response among the predicted values ​​of multiple trained surrogate models; step S103, determining the surrogate model confidence based on the test normalized scalar output response, the maximum predicted value, the minimum tested value, and the predicted value.

[0055] According to one embodiment of the present invention, in the test dataset, the test-normalized scalar output responses of multiple physical fields for multiple test samples are obtained, which are the true simulation values. For example, for a test sample with 3 physical fields (3 types of scalar output responses), the surrogate model will output 3 predicted values, corresponding to the test-normalized scalar output responses of the scalar output responses of these 3 physical fields, respectively. Using a trained surrogate model, the maximum value among the predicted values ​​corresponding to the test-normalized scalar output responses of multiple physical fields in all test samples is the predicted maximum value. Similarly, the predicted minimum value is the minimum value among the predicted values ​​corresponding to the test-normalized scalar output responses of multiple physical fields in all test samples. The accuracy of the surrogate model can be evaluated through the surrogate model confidence score.

[0056] According to an embodiment of the present invention, step S103 includes: determining the surrogate model confidence by performing root mean square analysis on the difference between the predicted value and the test standardized scalar output response and mapping the range to between 0 and 1.

[0057] According to an embodiment of the present invention, determining the surrogate model confidence based on the test standardized scalar output response, the predicted maximum value, the test minimum value, and the predicted value includes: determining the confidence of the i-th surrogate model according to formula (4). , (4), in, For the y-th test sample, input the predicted value corresponding to the scalar output response of the ith trained surrogate model to the s-th physics field. The test normalized scalar output response is the scalar output response of the s-th physics field for the y-th test sample. Let be the maximum predicted value among the predicted values ​​of the i-th trained agent model. To test the minimum value of the standardized scalar output response, n is the number of test samples, S is the number of physical fields, y≤n, s≤S, and i, y, s, n and S are all positive integers, and max is the maximum value function.

[0058] According to one embodiment of the present invention, in formula (4), The root mean square error is the square root of the mean square difference between the predicted value corresponding to the test normalized scalar output response of the scalar output response of the i-th trained surrogate model and the test normalized scalar output response of the s-th physic for the y-th test sample. The larger the root mean square error, the lower the accuracy of the i-th trained surrogate model. The difference between the maximum predicted value among the predictions of the i-th trained surrogate model and the minimum tested value of the normalized scalar output response is used for normalization. and The larger the ratio, the lower the accuracy of the i-th trained agent model. 1 minus The confidence score of the i-th surrogate model can be obtained. The higher the confidence score of the surrogate model, the higher the accuracy of the trained surrogate model and the better the performance of the trained surrogate model.

[0059] In this way, the trained surrogate model can be evaluated using test samples to determine the confidence level of the surrogate model. This approach takes into account the characteristics and advantages of different surrogate models, reduces the potential limitations of a single surrogate model, and finds the most suitable surrogate model for the current design problem, thereby improving the accuracy and reliability of the surrogate model.

[0060] According to one embodiment of the present invention, in step S11, the value ranges of all design variables are organized into a set using the trained surrogate model with the highest confidence level, serving as the search space for global optimization. A suitable global optimization algorithm, such as a genetic algorithm, is selected to quickly find the optimal design scheme by leveraging the predictive power of the surrogate model. Rapid iteration on the surrogate model ultimately yields a theoretically optimal solution. This method does not require extensive practical experiments; instead, it guides design optimization through simulation and surrogate model prediction, significantly saving time and cost. Furthermore, the intelligent optimization process considers the combined influence of multiple design variables, seeking a global optimum rather than a local optimum, thus improving the quality and performance of the design.

[0061] The simulation-driven intelligent design method based on unified data flow, according to embodiments of the present invention, allows for intelligent design not limited to specific design objects by setting design variables and their physical value ranges. By controlling the total number of samples, waste or deficiencies caused by manual experience-based settings are reduced. Converting high-dimensional physical field data into scalar output responses and standardizing these responses better adapts to the requirements of various surrogate model training algorithms, improving the training efficiency of surrogate models. Automated process design and surrogate model training and evaluation processes reduce manual intervention. Global optimization using the trained surrogate model with the highest confidence level quickly finds the optimal design scheme, shortening simulation time and enhancing the efficiency and accuracy of simulation design. When determining the total number of samples, the number of design variables and a preset sampling density factor can be used to control computational resource input, intelligently determine the required number of simulations, and reduce waste or deficiencies caused by manual experience-based settings. When determining the scalar output response, numerical values, gradients, and cell control volumes can be used to determine the scalar output response. Based on adaptive spatial weights of the physical field gradient, automated and lossless compressed feature extraction of high-dimensional physical field data is achieved. Key regions with drastic changes in physical quantities are automatically identified and assigned higher weights, making the extracted scalar output response more reflective of the physical essence related to the design objectives. When determining the standardized scalar output response, the mean, standard deviation, maximum, minimum, upper and lower bounds of the target normalization interval, and the scalar output response can be used to determine the standardized scalar output response. Standardization and target normalization interval mapping of the scalar output response eliminate differences between different samples caused by factors such as units and numerical ranges. This better adapts to the requirements of various surrogate model training algorithms, improves the convergence speed and stability of surrogate models, and reduces surrogate model bias or training difficulties caused by inconsistent data scales. When determining the confidence level of a surrogate model, the trained surrogate model can be evaluated using test samples to determine the confidence level. This takes into account the characteristics and advantages of different surrogate models, reduces the potential limitations of a single surrogate model, and allows us to find the most suitable surrogate model for the current design problem, thereby improving the accuracy and reliability of the surrogate model.

[0062] Figure 5An exemplary block diagram of a simulation-driven intelligent design system based on a unified data flow according to an embodiment of the present invention is shown. The system includes: a design variable module, used to acquire the number of design variables and set physical value ranges for the design variables; a total sample quantity module, used to determine the total sample quantity based on the number of design variables and a preset sampling density factor, wherein the total sample quantity includes training samples and test samples; an allocation module, used to allocate the number of training samples and the number of test samples according to a preset ratio based on the total sample quantity; a simulation module, used to import the total sample quantity into an integrated simulation environment, start and perform batch simulation to obtain multiple high-dimensional physical field data corresponding to each sample; and a scalar output response module, used to, based on the high-dimensional physical field data, The system comprises the following modules: a scalar output response determination module, a standardized scalar output response module, an association module, a training dataset and a test dataset, an existing surrogate model module, a training surrogate model module, a prediction value module, a surrogate model confidence module, a global optimization module, and a global optimization module. The first module determines the scalar output response based on the scalar output response. The second module determines the surrogate model confidence based on the test dataset and the predicted values.

[0063] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0064] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments, and any modifications or variations of the embodiments of the present invention may be made without departing from the stated principles.

Claims

1. A simulation-driven intelligent design method based on unified data flow, characterized in that, include: Obtain the number of design variables and set the physical value range for the design variables; The total number of samples is determined based on the number of design variables and a preset sampling density factor, wherein the total number of samples includes training samples and test samples. The number of training samples and test samples are allocated according to a preset ratio based on the total number of samples. The total number of samples is imported into an integrated simulation environment, and batch simulation is started to obtain multiple high-dimensional physical field data corresponding to each sample. A scalar output response is determined based on the high-dimensional physical field data. A standardized scalar output response is determined based on the scalar output response. The design variables and the standardized scalar output response are automatically associated and divided into training datasets and test datasets. Multiple surrogate models are trained using the training dataset to obtain multiple trained surrogate models. The test samples are input into the trained surrogate models to obtain their predicted values. The confidence level of the surrogate models is determined based on the test dataset and the predicted values. Global optimization is performed based on the trained surrogate model with the highest confidence level and the physical value range.

2. The simulation-driven intelligent design method based on unified data flow according to claim 1, characterized in that, The total number of samples is determined based on the number of design variables and the preset sampling density factor, including: taking the number of design variables as the basic scale parameter, combining squaring with logarithmic scaling, multiplying with the preset sampling density factor, and then processing with an up-rounding function to determine the total number of samples.

3. The simulation-driven intelligent design method based on unified data flow according to claim 1, characterized in that, Determining the scalar output response based on the high-dimensional physical field data includes: determining the values ​​and gradients of multiple physical fields for multiple samples at multiple spatial locations based on the high-dimensional physical field data; obtaining the unit control volume of multiple physical fields for multiple samples at multiple spatial locations; and determining the scalar output response based on the values, the gradients, and the unit control volume.

4. The simulation-driven intelligent design method based on unified data flow according to claim 3, characterized in that, Determining the scalar output response based on the numerical value, the gradient, and the unit control volume includes: using the ratio of the gradient magnitude of each unit's physical field to the maximum gradient magnitude of the entire field as a weighting coefficient for the numerical value; and performing a weighted average of the unit control volume based on the weighting coefficient of the numerical value and the numerical value to determine the scalar output response.

5. The simulation-driven intelligent design method based on unified data flow according to claim 1, characterized in that, Determining the standardized scalar output response based on the scalar output response includes: obtaining the mean, standard deviation, maximum value, and minimum value of each scalar output response across all samples; setting preset upper and lower bounds of the target normalization interval; and determining the standardized scalar output response based on the mean, the standard deviation, the maximum value, the minimum value, the upper and lower bounds of the target normalization interval, and the scalar output response.

6. The simulation-driven intelligent design method based on unified data flow according to claim 5, characterized in that, Determining the standardized scalar output response based on the mean, the standard deviation, the maximum value, the minimum value, the upper and lower bounds of the target normalization interval, and the scalar output response includes: standardizing the scalar output response and then mapping it to the target normalization interval.

7. The simulation-driven intelligent design method based on unified data flow according to claim 1, characterized in that, Determining the surrogate model confidence based on the test dataset and the predicted values ​​includes: obtaining the test normalized scalar output responses of multiple physical fields for multiple test samples based on the test dataset; obtaining the maximum predicted value and the minimum tested value of the test normalized scalar output response among the predicted values ​​of multiple trained surrogate models; and determining the surrogate model confidence based on the test normalized scalar output response, the maximum predicted value, the minimum tested value, and the predicted values.

8. The simulation-driven intelligent design method based on unified data flow according to claim 7, characterized in that, The confidence of the surrogate model is determined based on the test standardized scalar output response, the predicted maximum value, the test minimum value, and the predicted value, including: determining the confidence of the surrogate model by taking the root mean square of the difference between the predicted value and the test standardized scalar output response and mapping the range to between 0 and 1.

9. A simulation-driven intelligent design system based on unified data flow, used to execute the simulation-driven intelligent design method based on unified data flow as described in any one of claims 1-8, characterized in that, include: The design variable module is used to obtain the number of design variables and set the physical value range for the design variables. The system includes a total sample size module, used to determine the total sample size based on the number of design variables and a preset sampling density factor, wherein the total sample size includes training samples and test samples; an allocation module, used to allocate the number of training samples and the number of test samples according to a preset ratio based on the total sample size; a simulation module, used to import the total sample size into an integrated simulation environment, start and perform batch simulation to obtain multiple high-dimensional physical field data corresponding to each sample; a scalar output response module, used to determine the scalar output response based on the high-dimensional physical field data; and a standardized scalar output response module, used to determine the standardized scalar output response based on the scalar output response. The system comprises: a scalar output response module; an association module for automatically associating the design variables with the standardized scalar output response and dividing them into training and test datasets; a trained surrogate model module for training multiple surrogate models using the training dataset to obtain multiple trained surrogate models; a prediction value module for inputting the test samples into the trained surrogate models to obtain the predicted values ​​of the trained surrogate models; a surrogate model confidence module for determining the confidence of the surrogate models based on the test dataset and the predicted values; and a global optimization module for performing global optimization based on the trained surrogate model with the highest confidence and the physical value range.

Images