Geometric Microstructure Generation And Functionally Gradient Structure Design Based On Deep Learning

Microstructure refers to the very small-scale structures present in materials,which are widely found in both natural and artificially designed biomimetic structures.Its unique design concepts and application prospects have attracted global attention.By designing geometric microstructures that can be densely packed within micro-scale crystalline cells and assembling multiple such microstructures,mechanical metamaterials with spatial variations or functional gradient properties can be created.Therefore,the design and generation of geometric microstructures have become important issues in the field of material design.Existing research work focuses on designing geometric microstructure units by constructing parameterized models.By selecting parameter combinations and adjusting volume fractions,over a thousand different geometric microstructures can be created.Numerical optimization methods are also widely applied to design tiled microstructures with desired properties.These structures can smoothly change material properties within a certain range through interpolation,thereby creating structures with elastic properties that vary with space.However,as the desire for more complex physical properties achievable by microstructures increases,the complexity of the design process has greatly increased.Challenges such as how to design geometric microstructures that meet specified physical properties,how to assemble adjacent units in nonperiodic structures,and how to conduct physical simulation and verification have made it difficult for traditional design methods based on parameterized models and topology optimization to provide solutions.The use of data-driven methods has provided a new approach to the generation and design of geometric microstructures.In this paper,we propose a generative neural network based on VAE-GAN(Variational Autoencoder-Generative Adversarial Network)to generate novel geometric microstructures.By embedding both geometric and manufacturing constraints into the parameters of the generative neural network,we can generate geometric microstructures and perform continuous interpolation on the structures,thereby significantly expanding the design space and property space of geometric microstructures.This framework leverages the capabilities of autoencoder networks in deep learning to extract features from geometric microstructures in different parameter domains.By embedding the shape and properties of the microstructures into a latent space,we obtain a low-dimensional representation of the geometric microstructures.This enables us to design and adjust the topological shape and even the properties of the microstructures in the low-dimensional space,thereby greatly reducing the complexity of the design process.This paper explores the generation capability of neural networks further.It designs a distance metric based on hidden space to explore the distribution of microstructures and constructs an interpolation path to realize the generation of continuous sequences between geometric microstructures with different topologies.The approach obtains geometric microstructure sequences with smooth and excessive properties and analyzes the properties of the generated geometric microstructures using homogenization theory.Compared with the geometric microstructure generated by the parametric design model,this paper can give the structure generated by any parameter model or the microstructure created by manual design.Through the generalization ability of the neural network,the interpolation sequence between any two geometric microstructures can be obtained without the need to know the generation parameters of the model.For structures with different topologies and different parameter models,the interpolation task can be realized.A continuous sequence of geometric microstructures can be obtained for the design of functional gradient structures,making the process of generating designs for geometric microstructures more efficient.