Constrained Generative AI for Materials Inverse Design

Materials inverse design aims to discover new compounds with targeted structural and functional properties, but the search space is vast and strongly constrained by chemistry and physics. Deep generative models, especially diffusion-based models, have recently become powerful tools for crystal structure generation. However, unconstrained generation often struggles to produce candidates that satisfy specific structural motifs, chemical rules, or stability requirements. In this talk, I will discuss two directions for incorporating hard constraints into generative materials design. First, I will introduce structural motif constraints through diffusion inpainting, where predefined geometric patterns are imposed during generation. This strategy enables motif-guided discovery of candidate frustrated magnets and flat-band materials while preserving the flexibility of the base generative model. Second, I will discuss CrysVCD, a valence-constrained generative framework that enforces charge balance during crystal generation, improving chemical validity and increasing the likelihood of stable generated structures. Together, these examples show how explicit physical constraints can steer broad statistical generation towards targeted exploration of chemically realistic and functionally relevant materials.

Speaker bio

Mouyang Cheng is a PhD student in computational science and engineering & materials science and engineering at MIT, advised by Prof. Mingda Li. He is broadly interested in developing machine-learning methods and computational frameworks for materials discovery and characterization, with a focus on connecting data-driven models to physical insights. His research spans AI-assisted materials design, understanding defects and amorphous systems, as well as the integration of machine learning with simulations and experimental spectroscopy.

Series: MSR New England Generative Modeling & Sampling Seminar