Multi-material structure reverse design method driven by machine learning method

Abstract

The invention discloses a reverse design method of a multi-material structure driven by a machine learning method. The reverse design method comprises the following steps of S1, determining input and output of a model; s2, constructing a data set according to input and output of the model; s3, constructing a forward model by using the data set; s4, on the basis of the forward model, constructing and training a reverse model; s5, after reverse model training is completed, a multi-material entity is manufactured and subjected to performance verification; the method solves the problems that an existing method is extremely high in calculation cost and difficult to cover all potential optimization schemes, and the optimization mode of manually adjusting parameters is extremely low in efficiency and low in cost when facing the characteristics that multiple materials are multiple in variety, the structural performance relation is complex and the interface effect cannot be neglected. The problems of long research and development period and high cost due to the fact that the existing method is difficult to consider multi-target performance requirements and needs to reconstruct a theoretical model or adjust core parameters of numerical simulation for different target performances or application scenes are solved.

Description

Technical Field

[0001] This invention belongs to the field of reverse design methods for multi-material structures, specifically involving a reverse design method for multi-material structures driven by machine learning. Background Technology

[0002] Multimaterial structures refer to novel functional structures formed by artificially combining two or more substrates with different physical and mechanical properties according to a specific spatial distribution pattern. The core characteristic of these structures is that their macroscopic mechanical properties not only depend on the intrinsic properties of a single material, but also on the selection of different materials, optimization of spatial layout, and proportions to achieve functional customization that is difficult to achieve with a single material (such as anisotropic response, gradient hardness, and localized strengthening). Furthermore, compared to traditional single-material structures, multimaterial structures have the advantage that their performance depends not only on the intrinsic properties of a single material, but also on the selection of material types, spatial distribution patterns, interfacial bonding states, and the proportions of each component. Therefore, the core of multimaterial structure design is to establish a precise mapping relationship between the influence of material combinations and spatial distribution on macroscopic target performance, and current research primarily focuses on this direction.

[0003] Currently, the design methods for multi-material structures still follow traditional technical paths. These traditional paths mainly include theoretical analysis and numerical simulation. However, theoretical analysis methods cannot accurately describe the nonlinear relationships of mechanical responses using simplified models, and are only applicable to basic scenarios with simple structures and few material components. Numerical simulation methods require extensive iterative simulations to verify performance, resulting in extremely high computational costs and difficulty in covering all potential optimization schemes. Furthermore, existing design methods are extremely inefficient when dealing with the characteristics of multi-material structures, the variety of materials, the complex relationships between structural performance, and the significant interface effects. Manual parameter adjustments are insufficient and fail to meet multiple performance requirements. Additionally, different performance targets or application scenarios necessitate the reconstruction of theoretical models or adjustments to core numerical simulation parameters, leading to poor versatility, long development cycles, and high costs.

[0004] To address the limitations of existing methods, this invention proposes a multi-material structure reverse design method driven by machine learning. By training a large amount of material structure performance data, a high-performance prediction and design model is established through machine learning, providing a new approach for efficient multi-material design and enabling precise customization of material selection and spatial distribution structure based on target performance. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a multi-material structure reverse design method driven by machine learning. This method solves the problems of existing methods having extremely high computational costs, difficulty in covering all potential optimization schemes, extremely low efficiency of manual parameter adjustment when facing a variety of materials, complex structural performance relationships, and significant interface effects, and difficulty in meeting multiple performance requirements. Furthermore, it requires rebuilding theoretical models or adjusting core parameters of numerical simulations for different target performance or application scenarios, resulting in poor versatility, long R&D cycles, and high costs.

[0006] The technical solution of the present invention is as follows:

[0007] A machine learning-driven reverse design method for multi-material structures includes the following steps:

[0008] Step S1: Define the input and output of the model;

[0009] Step S2: Construct the dataset based on the model's input and output;

[0010] Step S3: Build a forward model using the dataset;

[0011] Step S4: Based on the forward model, construct and train the inverse model;

[0012] Step S5: After the reverse model training is completed, create and verify the performance of the multi-material entity.

[0013] Preferably, step S1 specifically includes:

[0014] During the training of the forward prediction model, the input data uses two multi-material structural design parameters, and the output uses the stress and strain parameters of the material in three directions. When performing the reverse prediction, the mechanical parameters are used as input and the structural design parameters of the material are used as output.

[0015] Preferably, step S2 includes the following steps:

[0016] Step S21: For a designed multi-material structure, compressive simulation is performed on the structure in the X, Y, and Z directions to obtain the stress-strain curves of the corresponding multi-material in the three directions. Then, the values ​​corresponding to t points are uniformly taken on the curve as the characteristic values ​​of stress and strain. These characteristic values ​​are the mechanical characteristics of the corresponding multi-material structure.

[0017] Step S22: After obtaining the mechanical characteristics of a multi-material structure, other multi-material structures and their corresponding mechanical characteristics can be obtained in the same way. By corresponding these mechanical data with the spatial vectors represented by their structures, the dataset can be obtained.

[0018] Step S23: After obtaining the corresponding stress and strain parameters through simulation, the corresponding data needs to be output as CSV format samples. During data preprocessing, the CSV format samples are converted into 32-bit floating-point arrays to eliminate data format differences. The training set and validation set are randomly divided in a 7:3 ratio to ensure the model's generalization ability.

[0019] Preferably, step S3 includes the following steps:

[0020] Step S31: The architecture of the forward model adopts a fully connected neural network, with the material type and material distribution vector of the cubic cell as input and the stress-strain curve as output. After processing by the hidden layer, the output layer outputs the stress-strain state of the material under pressure in the X, Y, and Z directions.

[0021] Step S32: The hidden layer has two fully connected layers, namely a ReLU activation layer and a normalization layer. This makes full use of the dataset obtained based on finite element simulation to discover the relationship between material type and distribution and the stress-strain curve of the material, so as to better train the positive prediction model and ensure the accuracy of mechanical response prediction.

[0022] Step S33: After constructing the forward model, perform multiple rounds of training;

[0023] Step S34: In each round of training, the dataset must be randomly divided into a 70% training set and a 30% validation set. Then, a callback function is configured to adaptively adjust the learning rate, and the model is saved for each round.

[0024] Step S35: In this training process, mean squared error is used as the loss function to evaluate the error between the predicted curve coordinates and the true curve coordinates, and stochastic gradient descent is used as the optimizer to improve convergence efficiency.

[0025] Step S36: During the training process, it is necessary to detect the training status and save the model with the best convergence after training is completed for auxiliary training of the inverse model.

[0026] Preferably, step S4 includes directly training the inverse model and a multi-training model assisted by the fusion of the forward model.

[0027] Preferably, step S5 includes the following steps:

[0028] Step S51: After the inverse model training is completed, input the 21 feature values ​​corresponding to the stress-strain curve of the target into the model. The inverse model outputs a 1×27 binary matrix to represent the design parameters of the multi-material structure.

[0029] Step S52: After obtaining the predicted distribution, substitute it back into the forward model, and the corresponding mechanical characteristic values ​​will be output again. Restore it to the stress-strain curve to obtain the predicted performance of the multi-material structure.

[0030] Step S53: Using the 1×27 binary matrix output by the reverse model as the optimal design parameters, use FDM printing to create a multi-material solid, remove the support structure and trim the edges and corners, and then conduct compression experiments on it in the X, Y and Z directions to obtain the actual mechanical property curves of the material.

[0031] Step S54: Compare the mechanical performance curve with the target curve to verify the degree of matching between the design parameters and the target mechanical requirements.

[0032] Compared with existing technologies, the multi-material structure reverse design method based on machine learning methods of the present invention has the following advantages:

[0033] 1. Compared with existing methods such as topology optimization and geometric parameter optimization, this method can significantly reduce the difficulty of constructing material constitutive relations in multi-material design, and does not require repeated manual adjustments, thus shortening the material development time.

[0034] 2. This method uses simulation data as the dataset, which ensures the authenticity of the dataset while avoiding the cost of conducting experiments that require a lot of manpower and resources.

[0035] 3. This method does not require readjusting the model architecture when facing different mechanical conditions in different scenarios. It can obtain a multi-material model that meets the requirements simply by changing the input parameters. The method is highly versatile and adaptable to various application scenarios. Attached Figure Description

[0036] To more clearly illustrate the purpose, design concept, and innovation of the multi-material structure reverse design method based on machine learning proposed in this invention, the invention will be described in detail below with reference to the accompanying drawings and tables.

[0037] Figure 1 This is a schematic diagram of the multi-material model structure of the present invention.

[0038] Figure 2 This is a diagram of the joint loss function structure of the present invention.

[0039] Figure 3 This is a diagram of the reverse design and performance prediction model architecture of the present invention.

[0040] Figure 4 This is a schematic diagram of the multi-material structure completed for the design of this invention.

[0041] Figure 5 This is a compression stress-strain diagram of Embodiment 1 of the present invention. Detailed Implementation

[0042] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0043] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0044] A machine learning-driven reverse design method for multi-material structures includes the following steps:

[0045] Step S1: Define the model's inputs and outputs;

[0046] Step S2: Construct the dataset based on the model's input and output;

[0047] Step S3: Build a forward model using the dataset;

[0048] Step S4: Based on the forward model, construct and train the inverse model;

[0049] Step S5: After the reverse model training is completed, create and verify the performance of the multi-material entity.

[0050] Step S1 of this implementation plan is as follows:

[0051] During the training of the forward prediction model, the input data uses two multi-material structural design parameters, and the output uses the stress and strain parameters of the material in three directions. When performing the reverse prediction, the mechanical parameters are used as input and the structural design parameters of the material are used as output.

[0052] Step S2 of this implementation plan includes the following steps:

[0053] Step S21: For a designed multi-material structure, compressive simulation is performed on the structure in the X, Y, and Z directions to obtain the stress-strain curves of the corresponding multi-material in the three directions. Then, the values ​​corresponding to t points are uniformly taken on the curve as the characteristic values ​​of stress and strain. These characteristic values ​​are the mechanical characteristics of the corresponding multi-material structure.

[0054] Step S22: After obtaining the mechanical characteristics of a multi-material structure, other multi-material structures and their corresponding mechanical characteristics can be obtained in the same way. By corresponding these mechanical data with the spatial vectors represented by their structures, the dataset can be obtained.

[0055] Step S23: After obtaining the corresponding stress and strain parameters through simulation, the corresponding data needs to be output as CSV format samples. During data preprocessing, the CSV format samples are converted into 32-bit floating-point arrays to eliminate data format differences. The training set and validation set are randomly divided in a 7:3 ratio to ensure the model's generalization ability.

[0056] Step S3 of this implementation plan includes the following steps:

[0057] Step S31: The architecture of the forward model adopts a fully connected neural network, with the material type and material distribution vector of the cubic cell as input and the stress-strain curve as output. After processing by the hidden layer, the output layer outputs the stress-strain state of the material under pressure in the X, Y, and Z directions.

[0058] Step S32: The hidden layer has two fully connected layers, namely a ReLU activation layer and a normalization layer. This makes full use of the dataset obtained based on finite element simulation to discover the relationship between material type and distribution and the stress-strain curve of the material, so as to better train the positive prediction model and ensure the accuracy of mechanical response prediction.

[0059] Step S33: After constructing the forward model, perform multiple rounds of training;

[0060] Step S34: In each round of training, the dataset must be randomly divided into a 70% training set and a 30% validation set. Then, a callback function is configured to adaptively adjust the learning rate, and the model is saved for each round.

[0061] Step S35: In this training process, mean squared error is used as the loss function to evaluate the error between the predicted curve coordinates and the true curve coordinates, and stochastic gradient descent is used as the optimizer to improve convergence efficiency.

[0062] Step S36: During the training process, it is necessary to detect the training status and save the model with the best convergence after training is completed for auxiliary training of the inverse model.

[0063] Step S4 of this implementation plan includes directly training the inverse model and using a multi-training model assisted by the fusion of the forward model.

[0064] Step S5 of this implementation plan includes the following steps:

[0065] Step S51: After the inverse model training is completed, input the 21 feature values ​​corresponding to the stress-strain curve of the target into the model. The inverse model outputs a 1×27 binary matrix to represent the design parameters of the multi-material structure.

[0066] Step S52: After obtaining the predicted distribution, substitute it back into the forward model, and the corresponding mechanical characteristic values ​​will be output again. Restore it to the stress-strain curve to obtain the predicted performance of the multi-material structure.

[0067] Step S53: Using the 1×27 binary matrix output by the reverse model as the optimal design parameters, use FDM printing to create a multi-material solid, remove the support structure and trim the edges and corners, and then conduct compression experiments on it in the X, Y and Z directions to obtain the actual mechanical property curves of the material.

[0068] Step S54: Compare the mechanical performance curve with the target curve to verify the degree of matching between the design parameters and the target mechanical requirements.

[0069] In implementing this scheme, taking an anisotropic, mechanically controllable multi-material structure as an example, we selected a cubic structure as the target material for design. First, the cube is divided into an array of 27 units, each unit being made of either a hard or soft material, as illustrated in the diagram below. Figure 1 As shown. Figure 1 This is a schematic diagram of a multi-material model structure (0 and 1 represent random distribution). When the materials of each unit of the entire cube are distributed according to different patterns, the whole will exhibit different mechanical properties in the X, Y, and Z directions. Based on the above design concept, if a multi-material structure is designed directly, the possible structures could reach 2... 27 There are many types of materials, and each design layout has different material properties, making differentiation and selection extremely labor-intensive. Using machine learning methods for design would significantly reduce this workload. Based on the above analysis, the design requires two parameters: the type of material and the location of its distribution. These two parameters correspond to the mechanical properties of the designed multi-material mixture. Through forward reasoning and reverse design processes using a machine learning model, materials that meet the required mechanical performance specifications can be obtained.

[0070] The specific implementation is as follows:

[0071] (1) Define the model input and output

[0072] During forward prediction model training, the input data uses two multi-material structural design parameters, and the output uses the stress and strain parameters of the material in three directions. During inverse prediction, the mechanical parameters are used as input, and the structural design parameters of the material are used as output. In actual simulation and model training, the two structural design parameters can be transformed into a 3×3×3 spatial vector, where the elements are only 0 and 1, with 0 corresponding to hard materials and 1 corresponding to soft materials. The position of the element in the matrix corresponds to the distribution location of the material, such as... Figure 1As shown. This spatial vector can be expanded into a 1×27 binary matrix as the input and output of the model.

[0073] (2) Constructing the dataset

[0074] The dataset is obtained by simulating models based on some multi-material design results. The specific steps for obtaining the dataset are as follows:

[0075] First, for a designed material structure, compression simulations are performed in the X, Y, and Z directions to obtain the stress-strain curves of the corresponding material in these three directions. Then, the values ​​corresponding to t points uniformly selected on the curves are taken as characteristic values ​​of stress and strain. These characteristic values ​​represent the mechanical characteristics of this multi-material structure. After obtaining the mechanical characteristics of one multi-material structure, the same approach can be used to obtain the mechanical characteristics of other multi-material structures. By mapping these mechanical data to the spatial vectors represented by their structures, a dataset can be obtained.

[0076] In our design case, using this method, we obtained a dataset of 10,000 samples using a script. Each dataset consists of two parts: a 3×3×3 spatial vector corresponding to two multi-material structural design parameters and 63 eigenvalues ​​corresponding to the stress-strain curves of the designed structure in three directions.

[0077] In addition, after obtaining the corresponding stress and strain parameters through simulation, the corresponding data needs to be output as CSV format samples. During data preprocessing, the samples are converted into 32-bit floating-point arrays to eliminate data format differences. The training set and validation set are randomly divided in a 7:3 ratio to ensure the model's generalization ability.

[0078] (3) Construction of the forward model

[0079] The forward model architecture employs a fully connected neural network, taking the material type and distribution vector of a cubic unit cell as input and the stress-strain curve as output. After processing through hidden layers, the output layer outputs the stress-strain state of the material under pressure in the X, Y, and Z directions. The hidden layers consist of two fully connected layers: a ReLU activation layer and a normalization layer. This allows full utilization of the dataset obtained from finite element simulations to discover the relationship between material type and distribution and the stress-strain curve, thus better training the forward prediction model and ensuring the accuracy of mechanical response prediction.

[0080] After building the forward model, multiple rounds of training are performed. In each round, the dataset is randomly divided into a 70% training set and a 30% validation set. A callback function is then configured to adaptively adjust the learning rate. Finally, the model from each round is saved. During this training process, mean squared error (MSE) is used as the loss function to evaluate the error between the predicted and true curve coordinates, and stochastic gradient descent (SGD) is used as the optimizer to improve convergence efficiency. Additionally, the training status needs to be monitored during the training process, and the model with the best convergence is saved after training is complete for auxiliary training of the inverse model.

[0081] (4) Construction and training of the inverse model

[0082] The core objective of the inverse model is to deduce the material distribution with corresponding mechanical characteristics from the stress-strain curve of the material. In order to ensure training accuracy, the present invention adopts the following two training schemes for the inverse training process.

[0083] ① Directly train the inverse model

[0084] Similar to the forward model, the inverse model is also designed as a neural network. The input layer receives data from the material's mechanical characteristic curves. To extract these curve features, the inverse model increases the number of hidden layers to four, and the output layer has 27 output heads, corresponding to two types of design parameters (type and location) for various materials. For training, 200 training iterations are performed, and the model with the optimal validation loss is saved.

[0085] ②Multiple training assisted by incorporating positive models

[0086] This step builds a comprehensive model combining a reverse design model and a forward prediction model. During the forward training process, each input and its corresponding weight are retained. The mechanical characteristics of the material are input into the input layer of the comprehensive model, the reverse model outputs the material design parameters (a 1×27 binary matrix), and then the forward model predicts the material's mechanical characteristics. Regarding the training strategy, 100 training epochs are performed, and the model with the smallest loss function value in each epoch is saved.

[0087] When predicting materials using the model, it can be observed that a high proportion of soft materials leads to a decrease in structural stiffness. However, some multi-material structures can still maintain extremely high mechanical properties because the soft materials form local support regions in space. Conversely, a large proportion of hard materials can also cause multi-material structures to exhibit flexible characteristics, because their deformation regions absorb most of the strain during compression. Therefore, when using stress-strain curves for inverse design tasks, the design results may contain structures with a high proportion of soft materials or a high proportion of hard materials. To explicitly control the material proportions in the design results, we propose a joint loss function (…). Figure 2The specific format is as follows:

[0088]

[0089] This loss function L incorporates a material proportioning constraint term into the original objective and uses a weighting coefficient n to adjust the priority of the proportioning constraint. It also considers curve fitting accuracy and the volume fractions of both hard and soft materials. contraint It is a loss function used to constrain the material ratio, while MSE is the mean squared error loss function used in the inverse model training process.

[0090] After introducing the new loss function, increasing the weighting coefficient n leads to a decrease in the matching degree between the performance curve and the target curve of the designed multi-material structure. Therefore, we introduce a new strategy: for multi-material structures that excessively pursue matching material properties with target properties, the m voxels with the lowest probability are replaced with the opposite material until the material ratio strictly meets the constraint conditions. This mechanism ensures that the designed multi-material structure meets the material ratio requirements while avoiding the degradation of mechanical properties caused by an excessively high n value.

[0091] The above describes the model building and training process for a single direction. To achieve anisotropy in multi-material structures while maintaining customizable mechanical properties in each direction, we introduce a weighted joint loss function:

[0092]

[0093] In this model, weights α, β, and γ prioritize stress features in specific directions. By adjusting the weights of the loss function in these three directions, the model can be controlled to prioritize the mechanical features in a particular direction. x loss Y loss Z It is the mean squared error loss function (MSE) used in the X, Y, and Z directions during training.

[0094] (5) Multi-material physical fabrication and performance verification

[0095] After the inverse model training is complete, the 21 eigenvalues ​​corresponding to the target stress-strain curve are input into the model. The model can output a 1×27 binary matrix representing the design parameters of the multi-material structure. After obtaining the predicted distribution, it is substituted back into the forward model, which will re-output the corresponding mechanical eigenvalues. By restoring these to the stress-strain curve, the predicted performance of the multi-material structure can be obtained. The specific structure is as follows: Figure 3 As shown.

[0096] Using the 1×27 binary matrix output by the above inverse model as the optimal design parameters, multi-material solids can be fabricated using FDM printing. The supporting structure can be removed and the edges and corners can be trimmed. Then, compression tests can be performed on the solids in the X, Y, and Z directions to obtain the actual mechanical property curves (stress-strain curves) of the materials. By comparing these curves with the target curves, the degree of matching between the design parameters and the target mechanical requirements can be verified.

[0097] Example 1:

[0098] Step 1: Model training;

[0099] Step 2: Input the mechanical curves into the model as input parameters, run the model, and obtain the multi-material structure design parameters predicted by the model;

[0100] Step 3: Design a multi-material structure according to the design parameters;

[0101] Step 4: Import the multi-material model into the 3D printing slicing software for slicing;

[0102] Step 5: Print multi-material structural entities using FDM printing technology;

[0103] Step 6: Set up a test platform, fix the material entity, select a direction as the X direction, perform a compression test on the material in this direction, and record its stress-strain curve data;

[0104] Step 7: Perform compression tests on the other two directions of the multi-material according to the right-hand coordinate system and record the mechanical data;

[0105] Step 8: Compare the actual stress-strain curve of the designed material with the target curve to prove that the stress-strain curve of the designed multi-material structure is in high agreement with the target curve.

[0106] Figure 4 The schematic diagram of the multi-material structure is shown in blue (the blue part is the soft material TPU, and the green part is the hard material PLA). Figure 5 The figure shows the stress-strain diagram of the compression test in Example 1. The green and purple curves represent the target performance to be achieved, while the gray, orange-red, and blue curves represent the actual performance of the material designed by the model. It can be seen that the performance of the designed material is almost consistent with the target performance in the X and Z directions, and is also very close in the Y direction.

Claims

1. A multi-material structure reverse design method driven by machine learning, characterized in that, Includes the following steps: Step S1: Define the model's inputs and outputs; Step S2: Construct the dataset based on the model's input and output; Step S3: Build a forward model using the dataset; Step S4: Based on the forward model, construct and train the inverse model; Step S5: After the reverse model training is completed, create and verify the performance of the multi-material entity.

2. The multi-material structure reverse design method based on machine learning driven by claim 1, characterized in that, Step S1 specifically involves: During the training of the forward prediction model, the input data uses two multi-material structural design parameters, and the output uses the stress and strain parameters of the material in three directions. When performing the reverse prediction, the mechanical parameters are used as input and the structural design parameters of the material are used as output.

3. The multi-material structure reverse design method based on machine learning driven by claim 1, characterized in that, Step S2 includes the following steps: Step S21: For a designed multi-material structure, compressive simulation is performed on the structure in the X, Y, and Z directions to obtain the stress-strain curves of the corresponding multi-material in the three directions. Then, the values ​​corresponding to t points are uniformly taken on the curve as the characteristic values ​​of stress and strain. These characteristic values ​​are the mechanical characteristics of the corresponding multi-material structure. Step S22: After obtaining the mechanical characteristics of a multi-material structure, other multi-material structures and their corresponding mechanical characteristics can be obtained in the same way. By corresponding these mechanical data with the spatial vectors represented by their structures, the dataset can be obtained. Step S23: After obtaining the corresponding stress and strain parameters through simulation, the corresponding data needs to be output as CSV format samples. During data preprocessing, the CSV format samples are converted into 32-bit floating-point arrays to eliminate data format differences. The training set and validation set are randomly divided in a 7:3 ratio to ensure the model's generalization ability.

4. The multi-material structure reverse design method based on machine learning driven according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: The architecture of the forward model adopts a fully connected neural network, with the material type and material distribution vector of the cubic cell as input and the stress-strain curve as output. After processing by the hidden layer, the output layer outputs the stress-strain state of the material under pressure in the X, Y, and Z directions. Step S32: The hidden layer has two fully connected layers, namely a ReLU activation layer and a normalization layer. This makes full use of the dataset obtained based on finite element simulation to discover the relationship between material type and distribution and the stress-strain curve of the material, so as to better train the positive prediction model and ensure the accuracy of mechanical response prediction. Step S33: After constructing the forward model, perform multiple rounds of training; Step S34: In each round of training, the dataset must be randomly divided into a 70% training set and a 30% validation set. Then, a callback function is configured to adaptively adjust the learning rate, and the model is saved for each round. Step S35: In this training process, mean squared error is used as the loss function to evaluate the error between the predicted curve coordinates and the true curve coordinates, and stochastic gradient descent is used as the optimizer to improve convergence efficiency. Step S36: During the training process, it is necessary to detect the training status and save the model with the best convergence after training is completed for auxiliary training of the inverse model.

5. The multi-material structure reverse design method based on machine learning driven by claim 1, characterized in that, Step S4 includes directly training the inverse model and using a multi-training model assisted by a fusion of the forward model.

6. The multi-material structure reverse design method based on machine learning driven by claim 1, characterized in that, Step S5 includes the following steps: Step S51: After the inverse model training is completed, input the 21 feature values ​​corresponding to the stress-strain curve of the target into the model. The inverse model outputs a 1×27 binary matrix to represent the design parameters of the multi-material structure. Step S52: After obtaining the predicted distribution, substitute it back into the forward model, and the corresponding mechanical characteristic values ​​will be output again. Restore it to the stress-strain curve to obtain the predicted performance of the multi-material structure. Step S53: Using the 1×27 binary matrix output by the reverse model as the optimal design parameters, use FDM printing to create a multi-material solid, remove the support structure and trim the edges and corners, and then conduct compression experiments on it in the X, Y and Z directions to obtain the actual mechanical property curves of the material. Step S54: Compare the mechanical performance curve with the target curve to verify the degree of matching between the design parameters and the target mechanical requirements.

Images