Generative design of high-fidelity microstructures using physics-aware machine learning

Generative models are increasingly being applied in materials science to enable materials design. However, generating out-of-distribution microstructures with high fidelity that satisfy multiple target properties remains a major challenge. Here, we present a physics-aware generative-optimization strategy for the inverse design of microstructures, guided by mutually exclusive properties. The approach integrates interfacial and crystallographic information directly into the loss function of surrogate models, together with introducing a latent-space sampling scheme that balances microstructural diversity with physical realizability. We demonstrate the framework using a small yet diverse alloy dataset comprising well-characterized microstructures and tensile stress-strain responses. By targeting strength-ductility synergy, the generated out-of-distribution microstructures exhibit substantially improved fidelity, particularly in their interfacial features. The produced microstructures are validated by crystal plasticity finite element simulations and experiments. We attribute the enhanced mechanical performance to the emergence of a bimodal grain-size morphology absent from the training data. Our results demonstrate a data-efficient strategy for high-fidelity microstructure design that leverages interfacial physics to achieve multiple properties synergy. This study uses physics-aware AI to generate high-fidelity microstructures beyond existing data distributions, resulting in a bimodal grain structure that enables improved strength and ductility.