A Multidisciplinary Optimization Method and System for Vehicle Collision and Lightweighting Based on Deep Learning Agent Model
By employing a deep learning proxy model for multidisciplinary optimization of vehicle collision and lightweighting, this approach addresses the challenges of high computational costs, long iteration cycles, and difficulties in multidisciplinary collaboration in vehicle body optimization. It achieves an efficient and precise balance between vehicle body lightweighting and safety performance, providing engineering decision support.
Patent Information
- Application Number
- CN202611602815.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing vehicle body optimization methods are computationally expensive, have lengthy iteration cycles, and are difficult to coordinate across multiple disciplines. Traditional surrogate models have poor fitting accuracy, making it difficult to achieve an effective balance between lightweighting and safety performance.
We employ a lightweight multidisciplinary optimization method for vehicle collision based on a deep learning agent model, including data preparation, agent model training, construction of a multidisciplinary optimization framework, collaborative evaluation, and multi-objective optimization solution. We utilize parallel training of deep learning networks and the OpenMDAO framework to achieve efficient collaborative optimization.
It significantly shortens the R&D cycle, improves prediction accuracy and model robustness, achieves a systematic balance between lightweight and safety performance, and provides high-quality solutions to assist engineering decision-making.
Smart Images

Figure CN123116458A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of automotive body structure design and performance optimization technology, and in particular relates to a multidisciplinary optimization method and system for vehicle body collision and lightweighting based on a deep learning agent model. Background Technology
[0002] With the rapid development of the automotive industry, the safety and lightweight design of vehicle body structures have become core concerns in the automotive R&D process. In the development of vehicle body safety performance, the Moving Progressive Deformable Barrier (MPDB) crash test and the Side Pole Impact (SP) are two important scenarios for evaluating vehicle crash safety. Meanwhile, achieving vehicle body lightweighting while meeting safety performance requirements is of great significance for reducing energy consumption and emissions.
[0003] Existing MPDB and SP vehicle body safety optimization typically employ a single-condition cyclical trial calculation method. Specifically, engineers first set up design of experiment (DOE) samples for the geometry, plate thickness, and material combination of the front-end structure and the side structure of the vehicle body. Then, they run full-vehicle finite element (CAE) simulations for MPDB and SP conditions respectively. Finally, they manually judge the merits of the proposed solutions based on indicators such as collision intrusion, acceleration, speed, and mass.
[0004] However, current mainstream vehicle body optimization methods have the following problems in practical engineering applications: First, the computational cost is high. A single MPDB and SP full-vehicle finite element simulation typically requires several hours or even tens of hours of computation time, and each design change requires a new simulation verification, resulting in a lengthy iterative optimization cycle that is difficult to meet the needs of rapid R&D in modern automobiles.
[0005] Secondly, multidisciplinary collaboration is difficult. When considering multiple disciplines' objectives such as MPDB, SP, and lightweighting, traditional methods require iterative iterations and trade-offs across multiple operating conditions, resulting in extremely low optimization efficiency and making it difficult to achieve an effective balance between lightweighting and safety performance.
[0006] Finally, traditional surrogate physical domains suffer from poor accuracy. To reduce computational costs, existing technologies often use traditional response surface methodology (RSM) and standard Kriging models to replace CAE simulations. However, these surrogate models have inherent limitations in fitting accuracy and are poorly suited for multi-condition optimization of vehicle collisions. Vehicle MPDB and SP collisions possess complex mechanical characteristics such as high-dimensional variables, strong nonlinearity, large instantaneous deformation, and abrupt contact changes. The surrogate models exhibit significant fitting deviations for core indicators such as intrusion amount, collision acceleration, and weight reduction, making it difficult to meet engineering accuracy requirements. Summary of the Invention
[0007] In view of this, the present invention aims to propose a vehicle body MPDB, SP and lightweight multidisciplinary optimization platform based on a deep learning proxy model, in order to solve the technical problems in the prior art, such as high computational cost leading to long iteration cycle, difficulty in multidisciplinary collaboration making it difficult to achieve an effective balance between lightweight and safety performance, and poor fitting accuracy of traditional proxy models when dealing with high-dimensional strong nonlinear collision conditions.
[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows: This invention provides a multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model, comprising the following steps: Data preparation steps: Read the multidimensional design input file and parse out multiple geometric variable columns, multiple thickness variable columns, and multiple material variable columns; merge the thickness variable columns and material variable columns into multiple integer combination variables according to the preset legal combination table, and together with the multiple geometric variable columns, they constitute multiple actual optimization variables; read the response labels of the moving progressive deformable barrier (MPDB) and the side column impact (SP), align them according to the sample identifier, perform normalization processing, and save the center value and half-range parameter; Proxy model training steps: Train multiple candidate deep learning networks in parallel, automatically select the best one based on the validation set index, obtain the MPDB proxy model and the SP proxy model, verify the input and output dimensions, and set the model to a gradient-free inference state. Steps for constructing a multidisciplinary optimization framework: Construct subject components based on the multidisciplinary optimization framework, analyze the boundaries and mapping relationships of multiple actual optimization variables, and reconstruct the candidate designs into shared input vectors according to the original header order; Collaborative evaluation steps: The shared input vector is fed into the MPDB and SP surrogate models in parallel. The physical quantities are recovered using the central value and half-range parameter. The first and second directional coefficients with opposite signs are used to optimize the direction. The non-mass response of SP uses the first directional coefficient, while the MPDB response and mass response use the second directional coefficient. The maximum value of the two disciplines' mass prediction values is taken as the global mass. Multi-objective optimization solution steps: Construct objective function and constraints, automatically select multi-objective optimization algorithm according to the number of objectives, use mixed variable operator for iterative solution, and calculate constraint violation degree; Sensitivity analysis and output steps: Perform sensitivity analysis on the recommended scheme based on the global sensitivity method, calculate the absolute mean and standard deviation of the basic effect, and select the Pareto optimal solution according to the three-level priority rules of feasibility, constraint violation degree and objective function sum.
[0009] Furthermore, the multidimensional design input file contains 69 columns of design input, including 25 columns of multiple geometric variables, 22 columns of multiple thickness variables, 22 columns of multiple material variables, and 47 columns of multiple actual optimization variables; the output dimension of the MPDB proxy model is 12-dimensional, and the output dimension of the SP proxy model is 13-dimensional; the first directional coefficient is -1, and the second directional coefficient is +1.
[0010] Furthermore, in the data preparation step, the thickness and material variable columns are not treated as independent optimization variables, but rather share a combined index. In the multidisciplinary optimization framework construction step, boundary pruning is performed on continuous variables, and valid integer index verification is performed on combined variables. The integer indexes are then decoded into thickness values and material grades and written into the corresponding columns to maintain the consistency of the model column order with that during training. The reconstruction into a shared input vector specifically includes: traversing the original header, reading continuous values for the geometry column and linearly mapping them to the normalized interval; for the thickness and material columns, retrieving the corresponding thickness values and material grades from the combined dictionary based on the combined index and normalizing them; and filling in default values for non-design variable columns.
[0011] Furthermore, in the surrogate model training step, candidate deep learning networks include multilayer perceptrons, one-dimensional convolutional neural networks, bidirectional long short-term memory networks, or self-attention networks; the selection process compares the validation loss, test set R2 score, and physical domain prediction accuracy in turn to select the optimal model; the physical domain prediction accuracy is obtained by calculating the relative error between the predicted value and the true value, and the result is cropped to a preset interval to suppress interference from abnormal samples.
[0012] Furthermore, in the multi-objective optimization solution step, the constraints include objective value tolerance constraints and quality consistency constraints; the quality consistency constraint is set to the absolute value of the difference between the two subject quality prediction values being less than or equal to the quality difference tolerance; the objective function includes MPDB subject performance indicators, SP subject performance indicators and global quality objectives, and the global quality supports minimization mode or target tracking mode.
[0013] Furthermore, in the multi-objective optimization solution step, the NSGA-II algorithm is used when the number of objectives is less than or equal to a preset threshold, and the NSGA-III algorithm is used when the number of objectives is greater than the preset threshold. For mixed variables containing continuous real numbers and discrete integers, mixed variable sampling is used to process the initial population. For continuous variables, simulated binary crossover and polynomial mutation strategies are used, and for integer variables, two-point crossover and polynomial mutation strategies are used.
[0014] Furthermore, in the sensitivity analysis and output steps, the three-level priority rules are as follows: the first priority is to prioritize feasible solutions, and individuals with constraint violation degrees less than or equal to the preset feasibility threshold are classified as feasible solutions; the second priority is to prioritize solutions with smaller constraint violation degrees; the third priority is to prioritize solutions with smaller sums of objective function values; after screening, the recommended solutions are re-run for collaborative evaluation to verify consistency, and the actual thickness, material grade, and physical response are output.
[0015] Based on the same concept, this invention also provides a multidisciplinary optimization system for vehicle collision and lightweighting based on a deep learning agent model, including: The data preparation module is used to perform data preparation steps, read the multidimensional design input shared by MPDB and SP, parse out multiple continuous structural variables and multiple thickness-material combination variables to form multiple actual optimization variables, and read the response labels to form a training dataset. The deep learning proxy model module is used to execute proxy model training steps, train or load MPDB proxy models and SP proxy models, verify the weight dimensions, and set the model to a gradient-free inference state. The shared design variable module is used to execute the multidisciplinary optimization framework construction steps, parse variables from the input header, merge the thickness and material columns into integer combination variables, copy the baseline input row according to the candidate design, and generate shared input by overlaying the header mapping. The MPDB-SP collaborative evaluation module is used to perform collaborative evaluation steps, reconstruct the same candidate design into a shared input, predict the response separately, recover the physical quantities using the center value and half-range parameters and unify the optimization direction, and calculate the global mass. The multi-objective constraint optimization module is used to execute multi-objective optimization solution steps, configure target tracking mode or minimum quality mode, construct response tolerance and quality consistency constraints, and automatically select a mixed variable multi-objective optimization algorithm according to the number of targets; The scheme output and sensitivity analysis module is used to perform sensitivity analysis and output steps. It filters the last recommended schemes based on feasibility, constraint violation degree and objective function, outputs the actual thickness, material grade and physical response, and performs global sensitivity analysis.
[0016] Based on the same concept, the present invention also provides an electronic device, which includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement a multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model.
[0017] Based on the same concept, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model.
[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. Significantly improves optimization efficiency and shortens the R&D cycle: This invention constructs and applies a deep learning proxy model, utilizing GPUs for millisecond-level inference, significantly replacing traditional CAE simulations that rely on CPUs for computation. This drastically reduces computational costs, enabling large-scale, multi-objective vehicle safety and lightweighting optimizations to be completed within an acceptable timeframe, effectively solving the problem of lengthy iteration cycles in traditional methods.
[0019] 2. Improved Prediction Accuracy and Model Robustness: This invention employs four heterogeneous architectures for parallel training: multilayer perceptron, one-dimensional convolutional neural network, bidirectional long short-term memory network, and self-attention model, and automatically selects the best model based on validation set performance. This strategy avoids the bias and limitations of a single model and effectively overcomes the problem of significant fitting deviations in traditional response surface methodology or basic surrogate models such as Kriging when dealing with complex mechanical features such as high-dimensional variables, strong nonlinearity, and instantaneous large deformations in whole-vehicle collisions. This significantly improves the prediction accuracy and robustness across different response indices.
[0020] 3. Achieving efficient multidisciplinary collaborative optimization: This invention constructs multidisciplinary components based on the OpenMDAO unified framework. Through a variable merging mechanism, it ensures that MPDB and SP disciplines share the same set of design variables. Furthermore, through an input vector reconstruction method, it automatically maps the design variables to the standardized inputs of each discipline's model, eliminating data alignment issues between disciplines and achieving true multidisciplinary collaborative optimization.
[0021] 4. Achieving a systematic balance between lightweight design and safety performance: This invention sets a specific global quality objective and introduces quality consistency constraints, supporting both quality minimization and fixed-value optimization modes. This mechanism ensures that quality prediction deviations across different disciplines are within tolerable limits, avoiding inconsistent solutions where some conditions are lightweight while others are heavyweight, thus achieving an effective balance between lightweight design and safety performance in engineering.
[0022] 5. Assisting engineering decision-making and improving solution usability: This invention provides flexible fixed-value target constraint functions to ensure that the optimized solution meets the mandatory engineering indicators; at the same time, it combines Morris global sensitivity analysis to accurately identify key parameters that are sensitive to multiple working conditions and interdisciplinary conflict parameters, and combines three-level priority rules to screen Pareto solution sets, providing engineers with quantitative basis and high-quality candidate solutions for final engineering selection and trade-offs. Attached Figure Description
[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the data preparation and preprocessing process provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the construction process of the OpenMDAO multidisciplinary optimization framework provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the objective and constraint architecture of a multi-objective optimization problem provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the overall architecture process provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the deep learning agent model training and verification process provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the unified evaluation and selection process for the four proxy models provided in this embodiment of the invention; Figure 7 This is a schematic diagram of the Morris global sensitivity analysis process provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the overall process of the MPDB-SP deep learning multidisciplinary optimization platform provided in this embodiment of the invention; Figure 9 This is a schematic diagram of the MPDB-SP data preparation module process provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the MPDB-SP deep learning agent model module provided in an embodiment of the present invention; Figure 11 This is a flowchart illustrating the MPDB-SP shared design variable module provided in an embodiment of the present invention; Figure 12 This is a schematic diagram of the MPDB-SP collaborative evaluation module provided in an embodiment of the present invention; Figure 13 This is a schematic diagram of the MPDB-SP multi-objective constraint optimization module provided in an embodiment of the present invention; Figure 14 This is a flowchart illustrating the MPDB-SP recommendation scheme output module provided in an embodiment of the present invention. Figure 15 This is a schematic diagram of the MPDB-SP Morris sensitivity module provided in an embodiment of the present invention. Detailed Implementation
[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0025] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.
[0026] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0027] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0028] like Figures 1 to 15 As shown, a multidisciplinary optimization method and system for vehicle collision and lightweighting based on a deep learning agent model is presented. The multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model includes the following steps: Data preparation steps: Read the multidimensional design input file and parse out multiple geometric variable columns, multiple thickness variable columns, and multiple material variable columns; merge the thickness variable columns and material variable columns into multiple integer combination variables according to the preset legal combination table, and together with the multiple geometric variable columns, they constitute multiple actual optimization variables; read the response labels of the moving progressive deformable barrier (MPDB) and the side column impact (SP), align them according to the sample identifier, perform normalization processing, and save the center value and half-range parameter; Proxy model training steps: Train multiple candidate deep learning networks in parallel, automatically select the best one based on the validation set index, obtain the MPDB proxy model and the SP proxy model, verify the input and output dimensions, and set the model to a gradient-free inference state. Steps for constructing a multidisciplinary optimization framework: Construct subject components based on the multidisciplinary optimization framework, analyze the boundaries and mapping relationships of multiple actual optimization variables, and reconstruct the candidate designs into shared input vectors according to the original header order; Collaborative evaluation steps: The shared input vector is fed into the MPDB and SP surrogate models in parallel. The physical quantities are recovered using the central value and half-range parameter. The first and second directional coefficients with opposite signs are used to optimize the direction. The non-mass response of SP uses the first directional coefficient, while the MPDB response and mass response use the second directional coefficient. The maximum value of the two disciplines' mass prediction values is taken as the global mass. Multi-objective optimization solution steps: Construct objective function and constraints, automatically select multi-objective optimization algorithm according to the number of objectives, use mixed variable operator for iterative solution, and calculate constraint violation degree; Sensitivity analysis and output steps: Perform sensitivity analysis on the recommended scheme based on the global sensitivity method, calculate the absolute mean and standard deviation of the basic effect, and select the Pareto optimal solution according to the three-level priority rules of feasibility, constraint violation degree and objective function sum.
[0029] The multidisciplinary optimization system for vehicle collision and lightweighting based on a deep learning agent model consists of six core modules: data preparation module, deep learning agent model module, shared design variable module, MPDB-SP collaborative evaluation module, multi-objective constraint optimization module, and scheme output and sensitivity analysis module. Each module is orchestrated by a unified controller.
[0030] The data preparation module is used to perform data preparation steps, read the multidimensional design input shared by MPDB and SP, parse out multiple continuous structural variables and multiple thickness-material combination variables to form multiple actual optimization variables, and read the response labels to form a training dataset. The deep learning proxy model module is used to execute proxy model training steps, train or load MPDB proxy models and SP proxy models, verify the weight dimensions, and set the model to a gradient-free inference state. The shared design variable module is used to execute the multidisciplinary optimization framework construction steps, parse variables from the input header, merge the thickness and material columns into integer combination variables, copy the baseline input row according to the candidate design, and generate shared input by overlaying the header mapping. The MPDB-SP collaborative evaluation module is used to perform collaborative evaluation steps, reconstruct the same candidate design into a shared input, predict the response separately, recover the physical quantities using the center value and half-range parameters and unify the optimization direction, and calculate the global mass. The multi-objective constraint optimization module is used to execute multi-objective optimization solution steps, configure target tracking mode or minimum quality mode, construct response tolerance and quality consistency constraints, and automatically select a mixed variable multi-objective optimization algorithm according to the number of targets; The scheme output and sensitivity analysis module is used to perform sensitivity analysis and output steps. It filters the last recommended schemes based on feasibility, constraint violation degree and objective function, outputs the actual thickness, material grade and physical response, and performs global sensitivity analysis.
[0031] The programmatic implementation of this embodiment corresponds to the optimization main program and the public platform program; the MPDB model weight file, the SP model weight file and the label normalization files of the two disciplines are loaded at an independent entry point, and the original multidisciplinary optimization engineering basic class is reused without rewriting the original engineering code.
[0032] The specific implementation method is as follows: Step 1: Data Preparation and Preprocessing like Figure 1 and Figure 9 As shown, this step mainly involves reading, matching, filtering, and cleaning the data to provide standardized data input for subsequent model training. The process begins by reading the CSV input file and the Excel label file, then matching samples according to the vehicle type (Sedan or SUV) and numbering rules, followed by filtering the valid label columns (Elongation or Disp), and finally removing NaN samples and handling outliers.
[0033] In one specific embodiment of the invention, the input data structure includes a CSV physical header containing 69 columns of data, used to retain complete design information and maintain alignment with the column order of the original CAE sample. For example... Figure 9 As shown, the system first checks if the number of header columns is 69. If not, it outputs the actual number of columns and stops; if so, it saves the original header and fixed column order. The optimized code, during the variable construction phase, parses the geometric variable G / P column and the thickness variable T column from these 69 columns as optimizable design variables. The material variable M column is not treated as an independent variable but shares a material thickness combination index with the corresponding T column.
[0034] It should be noted that the 69-column CSV header, data loader output, surrogate model input, and optimization component reconstruction vector all use the same dimension and column order. Although the thickness T column and the material M column share a combined index at the optimization variable layer, they still occupy their respective column positions at the model input layer. Therefore, there is no situation where the 69 columns coexist with other input dimensions.
[0035] The design variable input data is stored in a CSV file, containing 69 columns of features and three types of variables: The first category consists of 25 columns of geometric variables (G / P), which are part node geometric position parameters identified by / G / or / P / .
[0036] The second category is the thickness variable column T, which consists of 22 columns, with part thickness parameters identified by / T / .
[0037] The third category is the material variable column M, which consists of 22 columns and is identified by / M / ; each M column shares a composite index with the corresponding T column.
[0038] Both MPDB and SP optimization entry points use an input CSV file as a 69-dimensional input file, with each row corresponding to an experimental design sample; the original header order of this file is maintained during training, optimization, and sensitivity analysis.
[0039] Tag data is stored in Excel files and loaded separately according to dataset classes. For example... Figure 9 As shown, the system reads the MPDB label table and SP label table respectively, aligns them with the 69-dimensional input table according to the sample identifiers, and deletes missing, infinite, and labeled outlier samples to form MPDB 69-12 and SP 69-13 supervised samples. Then, with a fixed random seed of 2025, the dataset is divided into 70%, 20%, and 10% portions, and stored by subject and response center. and half-range The final output includes shared input, dual-discipline labels, and independent scale parameters.
[0040] For MPDB tag processing logic, the data preparation module executes the following sub-operations in sequence: Sub-operation 1: Read the Excel label file. The first column is the sample identifier column, and the rest are performance response columns.
[0041] Sub-operation 2: Call the filtering function to filter valid label columns. For columns that start with elongation or displacement, only keep columns that end with the final value and filter those that end with the maximum value; retain velocity, acceleration, and mass columns to ensure that training, optimization, and sensitivity analysis use consistent output variables.
[0042] Sub-operation 3: Parse the vehicle number and serial number from the sample identifier column, calculate the CSV row index by vehicle model data block, and match each label row with the corresponding design input row; record the reason for samples that cannot be parsed or are out of bounds and skip them.
[0043] Sub-operation 4: Skip all label samples that are empty or contain infinite values, and combine the remaining feature sequences with the label values to form valid sample pairs.
[0044] The MPDB proxy model ultimately outputs 12 dimensions, including 2 peak velocities, 8 final spring elongation values, 1 crew cabin acceleration evaluation value, and 1 mass; during denormalization, the MPDB normalization table is matched only by these 12 final label names.
[0045] The SP label processing logic is similar to that of MPDB, with the difference lying in the label filtering rules. A filtering function is called to retain columns that begin with the elongation or displacement and end with the final value. SP labels have 13 output dimensions, including 12 spring elongations and 1 mass. The SP denormalized table contains 25 variables.
[0046] The data preparation module's cleaning sub-process is as follows: any sample containing null values in the input and labels is skipped; features and labels are converted into floating-point tensors; outliers are identified and processed by a dedicated preprocessing program before training.
[0047] Both continuous inputs and outputs from each subject use the center-half-range parameter for scaling and physical quantity recovery; during the optimization phase, only the scaling parameters saved during the training phase are loaded without refitting, as shown in formula (1): (1); in, Indicates the first The central value of each variable, and These represent the upper and lower limits, respectively. Indicates half-range. This represents the normalized value. Represents the original physical value. Indicates the first The first discipline The physical recovery value of each output, and These represent the center value and half-range of the output, respectively. This represents the normalized value of the model output.
[0048] It should be noted that this invention uses the center-half-range parameter for scaling transformation. Compared with the traditional Min-Max normalization, this method uses the center of the data distribution as the benchmark, which can effectively avoid the distortion of the normalization interval caused by individual outliers, thereby improving the training stability of deep learning models on highly nonlinear data such as collision scenarios.
[0049] Step 2: Proxy Model Training like Figure 5 , Figure 6 and Figure 10 As shown, this step involves building and training a deep learning model and performing automatic optimization.
[0050] like Figure 10 As shown, the process begins by configuring MLP, CNN1D, LSTM, and SelfAttention candidate models for MPDB and SP. Subsequently, the training set undergoes batch forward computation, backpropagation, and parameter updates. The optimal epoch is selected based on the validation loss, and the weights are saved. The error is calculated on the test set, and the final model for each subject is determined. The entry point is optimized by instantiating the model, loading the optimal weights, and cleaning up key name prefixes. Next, dimensionality verification is performed: it checks if MPDB satisfies 69 inputs and 12 outputs; if not, it blocks and reports the weight dimension. If yes, it checks if SP satisfies 69 inputs and 13 outputs; if not, it blocks and reports the weight dimension. After successful verification, both models switch their eval state and disable gradients, registering output names, direction coefficients, center and half-range, ultimately outputting a parallel-callable MPDB / SP deep learning proxy model.
[0051] In one specific embodiment of the present invention, the platform trains four candidate networks in parallel: a multilayer perceptron, a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a self-attention model, and automatically selects the best one. The MPDB model has an input dimension of 69 and an output dimension of 12, including 11 performance responses and 1 mass; the SP model has an input dimension of 69 and an output dimension of 13, including 12 spring elongations and 1 mass.
[0052] Those skilled in the art should understand that candidate deep learning networks are not limited to the four types mentioned above, but also include other variants of recurrent neural networks (RNN), gated recurrent units (GRU), or Transformers. Any deep learning model capable of handling high-dimensional nonlinear mappings falls within the protection scope of this invention.
[0053] After training four models in parallel, the system does not manually specify a particular network, but automatically selects the best one based on its performance on the validation set. For example... Figure 5 As shown, the training script divides the data into 70% training set, 20% validation set, and 10% test set, using a fixed random seed of 2025 to ensure reproducibility. After each training round, the validation loss, R² for each output dimension, and the inversely normalized physical domain accuracy are calculated.
[0054] like Figure 6 The optimization process shown first involves a unified evaluation and selection, comparing training and validation losses to choose the checkpoint with the lowest validation loss; then, comparing the R² on the test set; and finally, comparing the model's prediction accuracy (Accc). Ultimately, the surrogate model with the lowest validation loss and the highest R² and Acc on the test set is selected. If validation losses are similar but stability varies significantly across different output dimensions, the model with a more balanced R² across multiple output metrics and higher physical domain accuracy is prioritized. This process demonstrates the complementary advantages of the four models.
[0055] The four model architectures are as follows: Model 1: Multilayer Perceptron. The first hidden layer maps the 69-dimensional input to 512 dimensions, followed by a ReLU activation function and a 20% probability Dropout; the second hidden layer compresses the 512-dimensional input to 256 dimensions and uses the same regularization, and the output layers are mapped to either the 12-dimensional MPDB or the 13-dimensional SP.
[0056] Model 2: One-dimensional Convolutional Neural Network. First, a linear expansion layer extends the 69-dimensional input to 256 dimensions. Then, two one-dimensional convolutional layers are used: the first layer has a kernel size of 3, expanding from 1 channel to 64 channels, followed by a max-pooling layer with a stride of 2; the second layer expands the number of channels from 64 to 128, also followed by max-pooling. After flattening, a 7936-dimensional feature vector is obtained. Finally, a fully connected layer reduces the dimension from 7936 to 128, and ReLU with Dropout is applied to output the feature vector.
[0057] Model 3: Bidirectional Long Short-Term Memory Network. The model reshapes the 69-dimensional input into a batch multiplied by 1 and then by 69 sequences, employing a two-layer bidirectional long short-term memory network and 256-dimensional hidden states; the final bidirectional 512-dimensional representation is mapped to the corresponding output dimension through a 64-dimensional fully connected layer.
[0058] Model 4: Self-Attention Model. Based on the Transformer encoder architecture, the 69-dimensional input is first mapped to a 128-dimensional embedding space through an embedding layer, and then passed through 4 Transformer encoder layers. Each encoder layer includes a multi-head attention mechanism with 4 attention heads, a 128-dimensional embedding dimension, layer normalization and residual connections, and a feedforward network that reduces the dimensions from 128 to 512 and then back to 128 using ReLU activation.
[0059] The training hyperparameters are configured as follows: 100 training epochs; batch size MPDB = 4, SP = 8; AdamW optimizer; learning rate... Weight decay coefficient The learning rate scheduling strategy uses cosine annealing with a period of 100; gradient clipping is enabled with a maximum gradient norm of 1.0; and the loss function is mean squared error.
[0060] like Figure 5 The training loop shown is as follows: In each training round, the system performs the following operations sequentially: First, it calculates the model's predicted output through batch forward propagation on the training set, calculates the loss using mean squared error, calculates the gradient through backpropagation, performs gradient pruning, and then updates the parameters using the AdamW optimizer. Then, it calculates the validation loss on the validation set in a gradient-free mode. If the current validation loss is lower than the historical best value, the model parameters are saved as the optimal checkpoint. After each round, the cosine annealing scheduler automatically adjusts the learning rate. After 100 rounds of training, the model with the lowest validation loss is selected as the final model. Figure 5 The judgment logic in the test set is used to determine whether the physical domain accuracy is greater than or equal to 90%. If it is, the model is output; otherwise, it is returned for retraining.
[0061] The evaluation metrics employ a dual evaluation approach. The first metric is the R-squared value in the normalized domain, calculated using a correlation function over the normalized range [-1, 1]. The second metric is the physical domain accuracy, which is calculated by first inversely normalizing the predicted and actual values column-by-column to restore them to the physical domain, and then calculating the relative accuracy. The calculation formula is as follows: (2); in, Indicates the precision of the physical domain. Represents the true value. This represents the predicted value. The result is cropped to the [0,1] interval to suppress interference from outliers.
[0062] The inverse normalization mechanism is the key transformation that restores the model output in the normalization domain [-1,1] to the physical domain. It must use mapping parameters that are completely consistent with those in the data preparation module.
[0063] First, the system call function reads the denormalized interval configuration table. This file contains three columns: variable name, lower limit, and upper limit. The function reads each row sequentially, building a mapping dictionary using the variable name as the key.
[0064] Then, the system calls a function to construct a column-by-column denormalization function. This function iterates through all output variable names, and for each output column, it looks up the corresponding upper and lower limits in the mapping dictionary according to the substring matching rules.
[0065] The mathematical formula for inverse normalization is: (3); This formula linearly maps the interval [-1, 1] to the lower and upper limits.
[0066] Step 3: Construction of a Multidisciplinary Optimization Framework like Figure 2 and Figure 11 As shown, this step is based on OpenMDAO to build a multidisciplinary optimization framework, including design variable analysis, subject component construction, target direction transformation and input vector reconstruction.
[0067] like Figure 11As shown, the process begins by scanning the 69 column headers in a fixed order, identifying 25 continuous G / P columns, 22 T columns, and 22 M columns. Each T column is paired with its corresponding M column using a mapping table, and each valid thickness and material combination is encoded as an integer index, forming 25 + 22 = 47 actual optimization variables. Candidate x is then received and the baseline 69-dimensional input is copied. It is determined whether the current variable is continuous; if so, the G / P column is overwritten; otherwise, the combined index is decoded and the T / M paired columns are overwritten. The 69-dimensional tensor is output in the saved original column header order. Next, it is determined whether there are unmapped or out-of-bounds variables; if so, the variable name is reported and the candidate is rejected; otherwise, all valid 69-dimensional candidates are simultaneously fed into the MPDB and SP models.
[0068] In one specific embodiment of the invention, a system call function automatically generates an optimized design variable space from the CSV header and configuration file. The parsing process is divided into three categories based on column type: The first type involves parsing the G / P column of geometric variables: Iterating through each column of the CSV header, converting the column names to uppercase, and then searching in the mapping table. If a match is found, the original key name is retrieved, the upper and lower bounds corresponding to that key are read from the configuration, and it is registered as a continuous real-number variable in the variable dictionary.
[0069] The second type involves parsing the thickness variable column T: After converting the column name to uppercase, it is searched in the mapping table. Upon successful matching, the original thickness key name is retrieved, and the corresponding material-thickness combination list is obtained from the combination dictionary configuration. This is then registered as a discrete integer variable in the variable dictionary, with a value range from 0 to the number of combinations minus 1.
[0070] The third category involves parsing the M column of material variables: For material columns identified by / M / , the system does not register design variables separately. Each M column shares the same combined index with its corresponding T column. For example, when the combined index of the thickness variable is 2, the system simultaneously reads the material grade of the second combination in the combined dictionary, enabling linked selection of material and thickness.
[0071] Variable merging: When MPDB and SP disciplines share the same set of input CSVs, the system call function merges the variable dictionaries of the two disciplines. This function takes the variable set of the first discipline as a base, appends the newly added variables from the second discipline, and retains only one copy of the existing variables, ensuring the uniqueness of the global design variable space.
[0072] Furthermore, 25 continuous variables and 22 thickness-material combination indices form a 47-dimensional actual design vector, and a 69-dimensional shared input is generated by the reconstruction operator with a fixed header, as shown in formula (4): (4); in, This represents a 47-dimensional actual design vector. Representing geometric variables, Indicates the index of the combined variable. Indicates the decoding function. and These represent thickness and material, respectively. This represents the reconstructed 69-dimensional input vector. This represents the reconstruction operator.
[0073] The design variable configuration parameters are derived from the configuration file: there are 47 pairs of upper and lower limit mappings for geometric parameters, including 25 geometric variables and 22 material-thickness combination variables.
[0074] Subject Component Construction: Each subject is constructed as a component, inheriting from OpenMDAO's explicit component base class. Taking the MPDB subject as an example, the following key parameters are passed during component initialization: subject name, complete CSV header list (69 columns), geometry column index list, thickness column index list, material column index list, design configuration dictionary, pre-trained surrogate model, computing device, output label list, and input normalization flag set to true. Each component defines two output variables: Y represents the surrogate model's original prediction value in the normalized domain; F represents the target value after direction sign transformation.
[0075] like Figure 12 As shown, the same 69-dimensional candidate tensor is fed into two deep learning proxy models in parallel. MPDB channel output. The 12-term physical response is recovered using the MPDB center-half-range parameters, with the response direction coefficients kept at +1 and the MPDB_ prefix added. SP channel output. The 13-term physical response was recovered using the SP center-half-range parameter, with the response direction coefficients for non-mass responses remaining at -1 and mass responses remaining at +1, and the SP_ prefix was added. Subsequently, the following was obtained: and Calculate global quality The final output includes 25 subject responses, optimization direction values, and overall quality.
[0076] The same reconstructed input is fed into the MPDB network and the SP network, which have independent parameters, respectively, to obtain standardized responses of different dimensions, as shown in formula (5): (5); in, and These represent the normalized outputs of the MPDB and SP networks, respectively. Representing the network model, Indicates reconstruction input, Indicates the model weights.
[0077] Target direction conversion mechanism: Since different disciplines have different optimization direction requirements for performance indicators, the system assigns a sign to each output dimension through a function, with a value of +1 or -1, unifying all targets into a minimization form.
[0078] For the MPDB security discipline, all performance metrics follow the principle that the smaller the value, the safer the result, and all outputs are assigned a value of +1.
[0079] For the SP subject, a differentiated processing strategy is adopted: the quality index is assigned a value of +1; the non-quality index, i.e., the spring elongation, is assigned a value of -1, and after being converted to maximize, they are unified into the minimize target.
[0080] When processing the response direction, the MPDB response is kept in the positive direction, the SP non-mass response is taken in the negative direction, and the SP mass response is kept in the positive direction, as shown in formula (6): (6); in, This represents the target value after transformation. Indicates the direction coefficient. This represents the normalized output of the model.
[0081] Input vector reconstruction method: Input vector reconstruction is the key bridge connecting the design variable space and the surrogate model input. The system reconstructs the 69-dimensional input vector column by column according to the column order of the original CSV header. The specific processing logic is divided into the following four cases: Case 1: Processing the G / P column (Geometry Column). Search for the uppercase form of this column in the mapping table. Upon successful matching, construct the design variable key name and read the corresponding continuous value from the design variable dictionary. If input normalization is enabled, use the normalization formula to linearly map the values from physical space to the [-1, 1] interval. The normalization formula is as follows: (7); in, This represents the normalized value. and These represent the upper and lower bounds, respectively.
[0082] Case 2: Processing the thickness column T. Search for the column name in the mapping table. If a match is found, construct a composite key name and retrieve the composite index value from the design variable dictionary. Extract the corresponding thickness value based on the composite index. If input normalization is enabled, normalize the thickness value from 0.5 to 5.5 to the interval [-1, 1].
[0083] Case 3: Material column M processing. The system replaces / M / in the column name with / T / to construct a paired candidate key name for column T. If it exists, the same composite key name is constructed, and the same composite index is obtained, realizing the combined sharing of columns T and M. The material grade is retrieved through this index, and if normalization is enabled, the range from 5000 to 65000 is mapped to [-1,1].
[0084] Case 4: Handling other columns. For columns that do not belong to any of the above categories, simply enter 0.0.
[0085] After reconstruction, output a floating-point vector of length 69; each position is filled with normalized values strictly according to the shared CSV header.
[0086] Step 4: Multi-objective optimization solution like Figure 3 and Figure 13 As shown, this step defines a multi-objective problem, configures the optimization algorithm, and performs individual evaluations.
[0087] like Figure 13 As shown, the process begins by reading the subject, response name, target value, and tolerance from the target JSON. It checks if the configured response name exists; if not, it lists all available ones. If it does, it checks if it's in target tracking mode; if not, it constructs a global quality target and collision response constraints; if so, it constructs a deviation target and two-sided tolerance. Then, it adds... The equivalent double inequality is used to assemble a continuous boundary and integer set of 47 mixed variables. Next, it checks if the target number is greater than 5. According to the logic in the text, if so, NSGA-III is used; otherwise, NSGA-II is used. After initiating the iteration, sampling, dual-model evaluation, and feasibility selection crossover mutation are performed. It checks if the termination generation or termination criterion is met. If not, the next generation is generated and the evaluation is returned; if so, the Pareto last generation population and the integrity constraint violation are output.
[0088] In one specific embodiment of the present invention, the system defines a problem class based on the pymoo library and bridges the OpenMDAO model to the multi-objective optimization framework.
[0089] The objective function is composed of: Let the target vector of MPDB subject after symbolic transformation be... The target vector of the SP subject after symbolic transformation is All components have been unified to the direction of minimization. Let... and If the predicted quality values are for two subjects, then the global quality is taken as the maximum value: (8); In minimize mode, the quality objective is written as: (9); In target tracking mode, the quality target is written as: (10); in, and Indicates the upper and lower limits of quality. Indicates target quality. This represents the quality tolerance. The final optimization objective is F, which includes... , and .
[0090] Constraints consist of: The target value constraint consists of the subject name, label name, target value, and tolerance. MPDB and SP both use a symmetrical tolerance band and convert it into a pair of inequalities of less than or equal to 0: (11); (12); In target tracking mode, the platform constructs a dimensionless deviation target and upper and lower tolerance constraints, as shown in formula (13): (13); in, Indicates the target value. Indicates tolerance. Indicates the predicted value. This represents the smallest positive number that prevents the denominator from being zero (e.g., ).
[0091] The second type is quality consistency constraints, which ensure that the differences between quality predictions across disciplines are within tolerance limits. Constraints are imposed on each pair of disciplines participating in the quality comparison: (14); Right now: (15); (16); In the quality objectives and consistency constraints, the maximum value of the two disciplines' quality is taken as the global quality, and the quality difference tolerance is expanded into a double inequality, as shown in formula (17): (17); in, This indicates the tolerance for poor quality.
[0092] Optimize algorithm configuration: The system selects the algorithm based on the number of optimization objectives: NSGA-II is used when the number of objectives is less than or equal to 5; NSGA-III is used when the number of objectives is greater than 5. Those skilled in the art should understand that multi-objective optimization algorithms are not limited to the NSGA series, but also include multi-objective evolutionary algorithms based on Pareto dominance or decomposition mechanisms, such as SPEA2 and MOEA / D.
[0093] To address the common issue in vehicle body design where continuous geometric variables coexist with discrete material-thickness combinations, the system employs mixed variable sampling for the initial population and crossover operations for mixed variable mating. Continuous variables utilize simulated binary crossover and polynomial mutation strategies, while integer variables use two-point crossover and polynomial mutation. Duplication elimination of mixed variables ensures the uniqueness of each generation's population. This mechanism avoids invalid solutions (such as non-existent material-thickness combinations) resulting from forcibly making discrete variables continuous, ensuring the feasibility of the optimization results in engineering manufacturing.
[0094] Individual assessment process: For each individual in the population, the system evaluates them in the following order: The first step is to set the design variables. Write the value of each design variable for each individual into the OpenMDAO model.
[0095] The second step is to run the OpenMDAO model. Calling the run function automatically triggers the computation methods of the MPDB and SP components. Within each discipline's computation method, the following steps are executed sequentially: calling the reconstruction function to reconstruct the 69-dimensional input vector based on the current design variables; performing forward inference prediction using a pre-trained surrogate model; and multiplying the network output by the sign vector to convert it into a target value with a unified minimization direction.
[0096] The third step is to extract the target values for each subject. The output vectors for each subject are read from the OpenMDAO model, and the non-mass components are extracted by index and concatenated into the parts of the multi-objective function.
[0097] The fourth step is to calculate the global quality and quality target. Extract quality values from the outputs of MPDB and SP respectively, and take the maximum value as the global quality; calculate the quality target value based on the quality model.
[0098] The fifth step is to calculate the constraint violation rate. For each fixed-value objective constraint, the corresponding output is first denormalized from the normalization domain to the physical quantity domain, and then compared with the objective value and tolerance to convert it into an inequality constraint value in the form of less than or equal to 0. For each pair of quality disciplines, the quality consistency constraint value is calculated by comparing the quality difference with the tolerance threshold.
[0099] After evaluation, the pymoo framework performs selection, crossover, and mutation operations based on the individual's target value and constraint violation degree, according to the non-dominated sorting and crowding / reference direction mechanism of NSGA-II or NSGA-III, to generate the next generation population until the termination condition is met.
[0100] Step 5: Sensitivity Analysis and Post-processing like Figure 7 and Figure 15 As shown, this step uses the Morris global sensitivity method to identify key design variables and interdisciplinary conflict parameters.
[0101] like Figure 15 As shown, the process begins by reading the recommended design and the definitions of 47 variables. A Morris random trajectory and variable grading points are constructed. Only one variable is changed at a time, while the other variables retain their recommended values. It is determined whether the current variable is continuous; if not, it jumps between valid combination indices; if so, it undergoes grading perturbations within upper and lower bounds. The MPDB-SP co-evaluation is rerun after each perturbation. Subsequently, the basic effect EE and absolute mean are calculated. and standard deviation Rank the variables by sensitivity for each response. Identify key variables, main effects, and potential nonlinearities / interactions. The final output includes a sensitivity CSV, a ranking plot, and a list of engineering conflict parameters.
[0102] In one specific embodiment of the present invention, sensitivity analysis is implemented based on a running function, using the Morris global sensitivity method library to identify key design variables and interdisciplinary conflict parameters. The process is as follows: The first step is to construct the problem definition. The design variable dictionary of the OpenMDAO model is converted into a standard format: extract the name list and the lower and upper bound range lists for each variable to form the problem object.
[0103] The second step is Morris sampling. The sampling method is called, setting the number of trajectories to 60 and the number of discrete levels to 4, to generate the Morris sampling matrix. The total number of sampling points is the number of variables plus 1 multiplied by the number of trajectories.
[0104] The third step is batch evaluation for each sample point. For each sample point, type correction is first performed, continuous variables are pruned to the legal range, and integer variables are rounded and then pruned. Then, the OpenMDAO model's running function is called to run a complete evaluation, collecting the target output values for each subject. The outputs of all sample points are stacked into a matrix.
[0105] The fourth step is to perform Morris analysis on each output dimension. For each column of the matrix, the analysis function is called to calculate two key sensitivity indicators based on the sampling matrix and the target column: the mean of the main effects, which measures the average strength of the variable's influence on the output; and the standard deviation, which measures the degree of interaction effects and nonlinear effects of the variables.
[0106] Step 5: Results Summary. Summarize the sensitivity results for all output dimensions, sort them in descending order by output and main effect mean, and save the results.
[0107] For discrete thickness-material combination variables, the perturbations in the Morris trajectory are first processed within the set of valid combination indices, then decoded into pairs of thickness and material and the co-evaluation is rerun, thereby avoiding the generation of combinations that do not exist in engineering.
[0108] The Morris basic effect for each response and its absolute mean and standard deviation are calculated according to formula (18): (18); in, Indicates the basic effect. Indicates the output function. Indicates the current value of the variable. Indicates the perturbation step size. Represents a unit vector. This represents the absolute mean. It represents the standard deviation.
[0109] The system automatically identifies two types of important parameters: Key parameters: For each output dimension, select the top few variables with the largest mean main effects; these variables constitute the set of key parameters for that output.
[0110] Conflicting parameters: These are variables that appear simultaneously in two or more sets of key output parameters. These parameters have a high impact on multiple performance metrics and may conflict in optimization directions.
[0111] Visualization Output: The system generates a scatter plot for each output dimension. The X-axis represents the main effect, and the Y-axis represents the standard deviation. This visualization helps engineers intuitively determine whether a variable exhibits linear dominance or strong nonlinear / strong interaction effects.
[0112] Step Six: Pareto Solution Set Filtering and Output like Figure 14 As shown, in one specific embodiment of the present invention, the system selects the top N candidate solutions from the final population according to a three-level priority. The process begins by outputting the last generation of the Pareto population and the number of integrity constraint violations. For each candidate, the following calculations are performed: .set up Construct sort key And sorted in ascending order, among which Infeasibility is flagged. The first candidate is selected as the recommended design. The 22 combined indices are decoded into thickness and material. The MPDB-SP co-evaluation is rerun for verification. The 47 variables, 69-dimensional inputs, 12 MPDB responses, and 13 SP responses are summarized. Global quality, constraint margins, algorithm parameters, random seeds, and runtime configuration are saved. The final output includes the recommended solution in JSON / CSV format, a Pareto plot, and a traceability report.
[0113] The sorting rules are as follows: First priority: Feasible solutions take precedence. Individuals with a constraint violation degree less than or equal to the feasibility threshold are classified as feasible solutions and marked with 0; otherwise, they are infeasible solutions and marked with 1. This flag is used as the first sorting key during sorting.
[0114] Second priority: Lower constraint violation rate takes precedence. Within the feasible / infeasible group, individuals with lower violation rates, i.e., closer to 0, are ranked higher.
[0115] Third priority: Prioritize minimizing the sum of the objective function values. All objective values are now aligned in the direction of minimization; a smaller sum indicates better overall optimization.
[0116] The system uses a sorting function to stably sort the feasibility flag, violation degree, objective function, and triplet, and selects the top few individuals in the sorting results as candidate Pareto optimal solutions.
[0117] The recommended scheme constructs the constraint violation degree and lexicographical selection key according to the feasibility priority principle, as shown in formula (19): (19); in, Indicates the degree of constraint violation. This represents the value of the k-th constraint function. Indicates the sort key. Indicates an indicator function, This represents the value of the i-th objective function. This represents the optimal solution.
[0118] Output content: After filtering, four sets of data are output: design variable combination, objective function value, constraint function value, and comprehensive constraint violation degree.
[0119] The platform decodes the combined index in the recommended solution into the actual thickness and material, reruns the collaborative evaluation for consistency verification, and saves the actual variables, model inputs, response values, global quality, constraint margins, algorithm parameters, and runtime configuration.
[0120] Example 1 In a preferred embodiment of the present invention, the E007, E115 and E202 series SUV samples are used as examples for illustration.
[0121] In this embodiment, an input CSV containing the geometry, thickness, and material information of 69 train bodies, as well as label Excel files for the MPDB and SP working conditions, are prepared first. The dataset class selects the corresponding rows from the CSV based on the car number and sample sequence number in the folder, and only retains the valid columns at the end of the final value for the elongation / displacement labels.
[0122] In this embodiment, training scripts are executed to train and select the best of the four types of models in parallel, and a normalization configuration file is used to establish an anti-normalization mapping so that the prediction results can be restored from [-1,1] to the real physical quantity domain.
[0123] In this embodiment, during the optimization phase, the system parses geometrically continuous variables and thickness-material combination discrete variables according to the configuration file, merging the variables shared by MPDB and SP into a unique global design space. For any candidate design, the MPDB and SP components respectively perform input vector reconstruction, model inference, and symbol transformation. The problem class is then combined with the MPDB objective, SP objective, and global quality objective, and the setpoint constraints and quality consistency constraints are calculated.
[0124] In this embodiment, after optimization, the system uses a screening function to select the top N Pareto candidate solutions from the final population according to a three-level rule: feasible solution priority, lower violation degree, and lower objective sum. Morris sensitivity analysis is then run to output the design variables, performance predictions, quality values, key parameters, and conflict parameters of each candidate solution for engineers to make the final engineering selection.
[0125] This invention also provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the above-described multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model.
[0126] This invention also provides a computer-readable storage medium storing computer instructions for causing a computer to execute the above-described multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model.
[0127] Beneficial effects of the present invention In combination with the above specific embodiments, the present invention has the following significant beneficial effects: 1. Significantly improves optimization efficiency and shortens the R&D cycle: This invention constructs and applies deep learning proxy models (such as MLP, CNN1D, LSTM, and SelfAttention) and utilizes GPUs for millisecond-level inference, significantly replacing traditional CAE simulations that rely on CPUs for computation. This greatly reduces computational costs, enabling large-scale, multi-objective vehicle safety and lightweight optimization to be completed within an acceptable timeframe, effectively solving the problem of lengthy iteration cycles in traditional methods.
[0128] 2. Improved prediction accuracy and model robustness: This invention employs four heterogeneous architectures for parallel training and automatically selects the best model based on validation set performance (e.g., ...). Figure 6 (As shown). This strategy avoids the bias and limitations of a single model and effectively overcomes the problem of significant fitting deviations in traditional response surface methodology or basic surrogate models such as Kriging when dealing with complex mechanical characteristics such as high-dimensional variables, strong nonlinearity, and instantaneous large deformation in whole-vehicle collisions. It significantly improves the prediction accuracy and robustness on different response indices.
[0129] 3. Achieving efficient multidisciplinary collaborative optimization: This invention constructs multidisciplinary components based on the OpenMDAO unified framework. Through a variable merging mechanism, it ensures that MPDB and SP disciplines share the same set of design variables, and utilizes input vector reconstruction methods (such as...) Figure 11 As shown, the design variables are automatically mapped to standardized inputs for each discipline's model, eliminating data alignment issues between disciplines and achieving true multidisciplinary collaborative optimization.
[0130] 4. Achieving a systematic balance between lightweight design and safety performance: This invention sets a specific global quality objective (Equation 8) and introduces quality consistency constraints (Equations 14-17), supporting both quality minimization and fixed-value optimization modes. This mechanism ensures that the quality prediction deviations of different disciplines are within the tolerable range, avoiding inconsistent solutions where some conditions are lightweight and others are heavy, thus achieving an effective balance between lightweight design and safety performance in engineering.
[0131] 5. Assisting engineering decision-making and improving solution usability: This invention provides flexible fixed-value target constraint functions (Formulas 11-13) to ensure that the optimized solution meets the mandatory engineering indicators; at the same time, it combines Morris global sensitivity analysis (such as... Figure 15 As shown, it accurately identifies key parameters that are sensitive to multiple operating conditions and interdisciplinary conflict parameters, and combines the three-level priority rule (Formula 19) to screen the Pareto solution set, providing quantitative basis and high-quality candidate solutions for engineers' final engineering selection and trade-offs.
[0132] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model, characterized in that, Includes the following steps: Data preparation steps: Read the multidimensional design input file and parse out multiple columns of geometric variables, multiple columns of thickness variables, and multiple columns of material variables; The thickness variable column and the material variable column are merged into multiple integer combination variables according to the preset legal combination table, which together with multiple geometric variable columns constitute multiple actual optimization variables; Read the response labels of the moving progressive deformable barrier (MPDB) and the side column impact (SP), align them according to the sample identifiers, normalize them, and save the center value and half-range parameter. Proxy model training steps: Train multiple candidate deep learning networks in parallel, automatically select the best one based on the validation set index, obtain the MPDB proxy model and the SP proxy model, verify the input and output dimensions, and set the model to a gradient-free inference state. Steps for constructing a multidisciplinary optimization framework: Construct subject components based on the multidisciplinary optimization framework, analyze the boundaries and mapping relationships of multiple actual optimization variables, and reconstruct the candidate designs into shared input vectors according to the original header order; Collaborative evaluation steps: The shared input vector is fed into the MPDB and SP surrogate models in parallel. The physical quantities are recovered using the central value and half-range parameter. The first and second directional coefficients with opposite signs are used to optimize the direction. The non-mass response of SP uses the first directional coefficient, while the MPDB response and mass response use the second directional coefficient. The maximum value of the two disciplines' mass prediction values is taken as the global mass. Multi-objective optimization solution steps: construct objective function and constraints, automatically select multi-objective optimization algorithm according to the number of objectives, use mixed variable operator for iterative solution, and calculate constraint violation degree; Sensitivity analysis and output steps: Perform sensitivity analysis on the recommended scheme based on the global sensitivity method, calculate the absolute mean and standard deviation of the basic effect, and select the Pareto optimal solution according to the three-level priority rules of feasibility, constraint violation degree and objective function sum.
2. The multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model according to claim 1, characterized in that, The multidimensional design input file contains 69 columns of design input, including 25 columns of multiple geometric variables, 22 columns of multiple thickness variables, 22 columns of multiple material variables, and 47 columns of multiple actual optimization variables; the output dimension of the MPDB surrogate model is 12-dimensional, and the output dimension of the SP surrogate model is 13-dimensional; the first directional coefficient is -1, and the second directional coefficient is +1.
3. The multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model according to claim 1, characterized in that, In the data preparation step, the thickness variable column and the material variable column are not treated as independent optimization variables, but share a combined index. In the multidisciplinary optimization framework construction step, boundary pruning is performed on continuous variables, valid integer index verification is performed on combined variables, and the integer index is decoded into thickness value and material grade and written into the corresponding column to keep the model column order consistent with that during training. The reconstruction into a shared input vector specifically includes: traversing the original header, reading continuous values for the geometry column and linearly mapping them to the normalized interval; and for the thickness and material columns, retrieving the corresponding thickness values and material grades from the combination dictionary based on the combination index and normalizing them. Enter default values for columns that are not design variables.
4. The multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model according to claim 1, characterized in that, In the surrogate model training step, candidate deep learning networks include multilayer perceptrons, one-dimensional convolutional neural networks, bidirectional long short-term memory networks, or self-attention networks; the selection process compares the validation loss, test set R2 score, and physical domain prediction accuracy in turn to select the optimal model; the physical domain prediction accuracy is obtained by calculating the relative error between the predicted value and the true value, and the result is cropped to a preset range to suppress the interference of abnormal samples.
5. The multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model according to claim 1, characterized in that, In the multi-objective optimization solution process, the constraints include objective value tolerance constraints and quality consistency constraints. The quality consistency constraint is set to the absolute value of the difference between the two subject quality prediction values being less than or equal to the quality difference tolerance. The objective function includes MPDB subject performance indicators, SP subject performance indicators, and a global quality objective. The global quality objective supports either a minimization mode or an objective tracking mode.
6. The multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model according to claim 1, characterized in that, In the multi-objective optimization solution process, the NSGA-II algorithm is used when the number of objectives is less than or equal to a preset threshold, and the NSGA-III algorithm is used when the number of objectives is greater than the preset threshold. For mixed variables containing continuous real numbers and discrete integers, mixed variable sampling is used to process the initial population. For continuous variables, simulated binary crossover and polynomial mutation strategies are used, and for integer variables, two-point crossover and polynomial mutation strategies are used.
7. The multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model according to claim 1, characterized in that, In the sensitivity analysis and output steps, the three-level priority rules are as follows: the first priority is to prioritize feasible solutions, and individuals with constraint violation degrees less than or equal to the preset feasibility threshold are classified as feasible solutions; the second priority is to prioritize solutions with smaller constraint violation degrees; the third priority is to prioritize solutions with smaller sums of objective function values; after screening, the recommended solutions are re-run for collaborative evaluation to verify consistency, and the actual thickness, material grade and physical response are output.
8. A multidisciplinary optimization system for vehicle collision and lightweighting based on a deep learning agent model, characterized in that, include: The data preparation module is used to perform data preparation steps, read the multidimensional design input shared by MPDB and SP, parse out multiple continuous structural variables and multiple thickness-material combination variables to form multiple actual optimization variables, and read the response labels to form a training dataset. The deep learning proxy model module is used to execute proxy model training steps, train or load MPDB proxy models and SP proxy models, verify the weight dimensions, and set the model to a gradient-free inference state. The shared design variable module is used to execute the multidisciplinary optimization framework construction steps, parse variables from the input header, merge the thickness and material columns into integer combination variables, copy the baseline input row according to the candidate design, and generate shared input by overlaying the header mapping. The MPDB-SP collaborative evaluation module is used to perform collaborative evaluation steps, reconstruct the same candidate design into a shared input, predict the response separately, recover the physical quantities using the center value and half-range parameters and unify the optimization direction, and calculate the global mass. The multi-objective constraint optimization module is used to execute multi-objective optimization solution steps, configure target tracking mode or minimum quality mode, construct response tolerance and quality consistency constraints, and automatically select a mixed variable multi-objective optimization algorithm according to the number of targets; The scheme output and sensitivity analysis module is used to perform sensitivity analysis and output steps. It filters the last recommended schemes based on feasibility, constraint violation degree and objective function, outputs the actual thickness, material grade and physical response, and performs global sensitivity analysis.
9. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multidisciplinary optimization method for vehicle collision and lightweighting based on a deep learning agent model as described in any one of claims 1 to 7.
