Physics-Aware Machine Learning-Driven High-Fidelity Microstructure Generative Design

The Northwestern Polytechnical University team has open-sourced a microstructure generative design framework, which combines Variational Autoencoders (VAE) with physical constraints to achieve high-fidelity generation and inverse optimization of material microstructures. This project is based on the paper Generative design of high-fidelity microstructures using physics-aware machine learning, providing a complete toolchain. The core innovation lies in integrating physical constraints into the machine learning process to ensure the generated results are physically feasible. The project is open-sourced on GitHub, with original authors including Weijie Liao and Ruihao Yuan, and was released in June 2026.

Microstructure is a core factor determining the macroscopic properties of materials. However, traditional design relies on experimental trial-and-error and empirical formulas, which are time-consuming and costly. Machine learning brings new possibilities to this field, but simple generative models lack physical constraints, leading to generated structures that may not be practically feasible or have the expected mechanical properties.

The project is developed using Python 3.8.3 and depends on the PyTorch and pymoo libraries. Its core components include: 1. Variational Autoencoder (VAE) to learn the latent representation of microstructures and generate diverse samples; 2. EBSD image generation module to provide examples of Electron Backscatter Diffraction (EBSD) images; 3. Inverse design optimization to infer microstructures from target properties, using pymoo for multi-objective optimization; 4. Latent space operations to support interpolation and sampling for exploring structure-property relationships.

This toolchain can be applied in: 1. Aerospace material design, quickly screening microstructures that meet strength-to-weight ratio requirements; 2. New energy battery materials, designing electrode structures with optimal porosity and connectivity; 3. Academic research and teaching, providing a cross-learning platform for machine learning and materials science.

The project provides detailed installation guides and Jupyter Notebook demos, configured based on a conda environment with clear dependency versions, making reproduction easy. The demo notebooks can run quickly on ordinary desktops to generate results. Users with a machine learning background can easily extend the code, while users with a materials science background experience reduced 'black box' feeling due to the physical constraints.

The current version mainly focuses on 2D microstructures; 3D expansion is the next direction. Physical constraints are currently based on elasticity mechanics and need to integrate complex behaviors such as plasticity and fracture. The project team provides a contact email (rhyuan@nwpu.edu.cn), maintains the project actively, and is expected to become an important infrastructure in the field of materials informatics.

This project represents a typical application of AI for Science in the materials field, deeply integrating domain knowledge (physical constraints) with generative models, and providing a reference for the development of scientific machine learning.