Physics-Informed Machine Learning Methods for Inverse Design of Multi-Phase Materials with Targeted Mechanical Properties

Author

Yunpeng WuFollow

Date of Award

8-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mechanical Engineering

Committee Chair/Advisor

Gang Li

Committee Member

Feng Luo

Committee Member

Huijuan Zhao

Committee Member

Oliver Myers

Abstract

Advances in machine learning algorithms and applications have significantly enhanced engineering inverse design capabilities. This work focuses on the machine learning-based inverse design of material microstructures with targeted linear and nonlinear mechanical properties. It involves developing and applying predictive and generative physics-informed neural networks for both 2D and 3D multiphase materials.

The first investigation aims to develop a machine learning method for the inverse design of 2D multiphase materials, particularly porous materials. We first develop machine learning methods to understand the implicit relationship between a material's microstructure and its mechanical behavior. Specifically, we use ResNet-based models to predict the elastic modulus and stress-strain curves of linear and nonlinear porous materials from their microstructure images. To generate microstructures of porous materials with targeted mechanical behavior, we create variational autoencoder (VAE) based neural networks. These networks generate the microstructure of porous materials from a prescribed elastic modulus or stress-strain curve. In both property prediction and microstructure generation, the stress-strain curves are approximated using cubic polynomials and characterized by their coefficients. To explicitly enforce the mechanics of materials in the generative machine learning models, we devise and incorporate a new condition fusion layer into the traditional VAE architecture. Additionally, a pretrained regression model is introduced to constrain the decoder, ensuring the production of physically meaningful images. The results show that this machine learning approach is capable of ultra-fast prediction of material properties directly from microstructure images, as well as the inverse design of material microstructures to achieve desirable mechanical behaviors.

The second investigation focuses on the inverse design of 3D multiphase materials, specifically considering fiber-reinforced polymer composites (FRPC) as the model system. This research aims to develop physics-informed neural networks for inverse design of such a material system. Compared to 2D porous materials, 3D FRPC involve complex 3D microstructure geometries and require physically feasible topologies, making the inverse design significantly more challenging. To address these challenges, we develop a novel diffusion model for 3D fiber reconstruction and generation. This model includes a forward diffusion process that adds noise to the fiber distribution and a reverse process that denoises to generate desirable fiber distributions. The diffusion model is defined via a stochastic differential equation (SDE), and both the diffusion and reverse processes are modeled as solutions to this SDE. To ensure feasible topology for the generated fiber distributions, non-collision constraints are incorporated into the generative neural networks. The results demonstrate that these new models can generate high-quality 3D FRPC designs with tailored mechanical behaviors while ensuring compliance with physical constraints.

Recommended Citation

Wu, Yunpeng, "Physics-Informed Machine Learning Methods for Inverse Design of Multi-Phase Materials with Targeted Mechanical Properties" (2024). All Dissertations. 3657.
https://open.clemson.edu/all_dissertations/3657