CN-121997670-A - Multi-material structure reverse design method based on machine learning method driving

CN121997670ACN 121997670 ACN121997670 ACN 121997670ACN-121997670-A

Abstract

The invention discloses a multi-material structure reverse design method driven by a machine learning method, which comprises the following steps of S1, defining input and output of a model; the method comprises the steps of S2, constructing a data set according to input and output of a model, S3, constructing a forward model by utilizing the data set, S4, constructing and training a reverse model on the basis of the forward model, S5, manufacturing and verifying multi-material entities after the reverse model is trained, and the method solves the problems that the existing method is extremely high in calculation cost, difficult to cover all potential optimization schemes, extremely low in efficiency of an optimization mode of manually adjusting parameters when facing the characteristics of multiple multi-material materials, complex in structural performance relationship and non-negligible interface effect, difficult to consider multi-target performance requirements, and needs to reconstruct a theoretical model or adjust core parameters of numerical simulation aiming at different target performances or application scenes, and poor in universality, so that the research and development period is long and the cost is high.

Inventors

WANG MEI
She Yuxiang
HUANG JIGANG
YANG HAOCHENG

Assignees

四川大学

Dates

Publication Date: 20260508
Application Date: 20260206

Claims (6)

1. The multi-material structure reverse design method driven by the machine learning method is characterized by comprising the following steps of: Step S1, defining input and output of a model; s2, constructing a data set according to the input and output of the model; s3, constructing a forward model by utilizing the data set; S4, constructing and training a reverse model on the basis of the forward model; and S5, after the reverse model training is completed, manufacturing and performance verifying the multi-material entity.
2. The machine learning method-driven multi-material structure reverse design method according to claim 1, wherein the step S1 is specifically: when the forward prediction model is trained, two multi-material structural design parameters are used for input data, stress strain parameters of materials in three directions are output, and when the reverse prediction is carried out, the mechanical parameters are used as input, and the structural design parameters of the materials are used as output.
3. The machine learning method driven multi-material structure reverse design method according to claim 1, wherein the step S2 comprises the steps of: step S21, for a designed multi-material structure, compression simulation is carried out on the structure in the X, Y, Z direction to obtain stress-strain curves of the corresponding multi-material in three directions, and then numerical values corresponding to t points are uniformly taken on the curves as characteristic values of stress strain, wherein the characteristic values are mechanical characteristics of the corresponding multi-material structure; Step S22, after the mechanical characteristics of the design machine of a multi-material structure are obtained, other multi-material structures and the mechanical characteristics corresponding to the multi-material structures can be obtained by the same method, and a data set can be obtained by corresponding the mechanical data and the space vectors represented by the structures; Step S23, after the corresponding stress-strain parameters are obtained through simulation, the corresponding data are required to be output as samples in a CSV format, when the data preprocessing is carried out, the samples in the CSV format are used as 32-bit floating point arrays, the data format difference is eliminated, the training set and the verification set are randomly divided according to the ratio of 7:3, and the generalization capability of the model is ensured.
4. The machine learning method driven multi-material structure reverse design method according to claim 1, wherein the step S3 comprises the steps of: s31, adopting a fully-connected neural network as a framework of a forward model, taking the material types and material distribution vectors of cube cells as input, taking a stress-strain curve as output, and outputting stress-strain conditions of materials subjected to pressure in X, Y, Z in an output layer after the materials are processed by a hidden layer; Step S32, the hidden layer is provided with two full-connection layers, namely a ReLU activation layer and a normalization layer, so that a data set obtained based on finite element simulation is fully utilized, the relationship between the types and the distribution of materials and the stress-strain curve of the materials is found, a forward prediction model is trained better, and the prediction precision of mechanical response is ensured; S33, after a forward model is built, performing multiple rounds of training; Step S34, during each round of training, the data of the data set is firstly randomly divided according to the proportion of 70% training set and 30% verification set, and then a callback function is configured to realize self-adaptive adjustment of learning rate, and each round of model is stored; step S35, in the training process, the mean square error is used as a loss function to evaluate the error between the predicted curve coordinate and the real curve coordinate, and a random gradient descent method is used as an optimizer to improve the convergence efficiency; Step S36, detecting training conditions in the training process, and storing a model with optimal convergence after the training is completed for auxiliary training of the reverse model.
5. The machine learning based method driven multi-material structure reverse design method of claim 1 wherein step S4 comprises direct training of a reverse model and fusing of a forward model aided multi-training model.
6. The machine learning method driven multi-material structure reverse design method according to claim 1, wherein the step S5 comprises the steps of: s51, after the reverse model training is completed, inputting 21 characteristic values corresponding to a stress-strain curve of a target into a model, and outputting a binary matrix of 1 multiplied by 27 for representing the multi-material structural design parameters by the reverse model; step S52, after obtaining the predicted distribution, substituting the predicted distribution into a forward model, and outputting corresponding mechanical characteristic values again to restore the predicted distribution into a stress-strain curve to obtain the performance of the predicted multi-material structure; S53, printing a multi-material entity by using a1 multiplied by 27 binary matrix output by a reverse model as an optimal design parameter and utilizing FDM, removing a supporting structure, trimming corners, and carrying out compression experiments on the supporting structure in three directions X, Y, Z to obtain an actual mechanical property curve of the material; And S54, comparing the mechanical property curve with a target curve, and verifying the matching degree of the design parameters and the target mechanical requirement.

Description

Multi-material structure reverse design method based on machine learning method driving Technical Field The invention belongs to the field of multi-material structure reverse design methods, and particularly relates to a multi-material structure reverse design method driven by a machine learning method. Background The multi-material structure refers to a novel functional structure formed by combining two or more base materials with different physical and mechanical properties according to a specific spatial distribution rule through manual design. The structure has the core characteristics that the macroscopic mechanical property of the structure not only depends on the intrinsic characteristics of a single material, but also realizes functional customization (such as anisotropic response, gradient hardness, local reinforcement and the like) which is difficult to be completed by the single material through the variety selection, spatial layout optimization and proportion collocation of different materials. In addition, compared with the traditional single-material structure, the multi-material structure has the advantages that the performance of the multi-material structure depends on the intrinsic characteristics of the single material, the type selection of the materials, the spatial distribution rule, the interface combination state and the proportion collocation of the components. Therefore, the design core of the multi-material structure is to establish a precise mapping relationship between the combination of material types and the influence of the spatial distribution structure on the macroscopic target performance, and the current research is mainly in this direction. At present, the design method of the multi-material structure is still carried out along the conventional technical path. The traditional path mainly comprises two types of theoretical analysis and numerical simulation, but the theoretical analysis method cannot accurately describe the nonlinear relation of mechanical response through a simplified model, and can only be suitable for basic scenes with simple structures and fewer material components. The numerical simulation method needs to verify performance through a large number of iterative simulations, has extremely high calculation cost, and is difficult to cover all potential optimization schemes. In addition, the existing design method has extremely low efficiency of an optimization mode of manually adjusting parameters and is difficult to meet the requirement of multi-target performance when facing the characteristics of multiple materials, multiple types of materials, complex structural performance relationship and non-negligible interface effect, and meanwhile, according to different target performance or application scenes, a theoretical model is required to be reconstructed or core parameters of numerical simulation are required to be adjusted, so that the universality is poor, the research and development period is long, and the cost is high. Aiming at the limitations of the existing method, the invention provides a multi-material structure reverse design method driven by a machine learning method, which builds a high-performance prediction and design model by training a large amount of material structure performance data in a machine learning driving mode, provides a new way for the efficient design of multi-materials, and realizes the accurate customization of the target performance-oriented material type selection and spatial distribution structure. Disclosure of Invention In order to solve the technical problems, the invention provides a multi-material structure reverse design method driven by a machine learning method, which solves the problems that the existing method has extremely high calculation cost, is difficult to cover all potential optimization schemes, has extremely low efficiency of an optimization mode of manually adjusting parameters when facing the characteristics of multiple material types, complex structure performance relationship and non-negligible interface effect, is difficult to consider multi-target performance requirements, needs to reconstruct a theoretical model or adjust core parameters of numerical simulation aiming at different target performances or application scenes, has poor universality, and causes long research and development period and high cost. The technical scheme of the invention is as follows: A multi-material structure reverse design method based on machine learning method driving comprises the following steps: Step S1, defining input and output of a model; s2, constructing a data set according to the input and output of the model; s3, constructing a forward model by utilizing the data set; S4, constructing and training a reverse model on the basis of the forward model; and S5, after the reverse model training is completed, manufacturing and performance verifying the multi-material entity. Preferably, step S1 is specifically: w