Presentation Information
[9p-B21-8]Materials Design with Generative Models and Accelerated Atomistic Simulation
〇Ryota Tomioka1 (1.MSR)
Keywords:
functional materials design,generative model,atomistic simulation
Designing functional materials with desired properties is a central challenge in areas such as energy storage, catalysis, and carbon capture. In this talk, I will discuss two complementary machine-learning approaches for materials design: generative models and accelerated atomistic simulation. MatterGen, developed at Microsoft Research, is a diffusion-based foundation model for inorganic crystal structure generation. When fine-tuned on property datasets, it can generate previously unknown crystal structures with targeted characteristics, highlighting the potential of generative models to extrapolate beyond the training data. MatterSim is a transformer-based foundation model for machine-learning interatomic potentials. It enables fast, accurate atomistic simulations and, with fine-tuning, high-accuracy prediction of material properties. I will describe the principles and applications of both models, and also discuss recent advances including inference acceleration, phase-diagram prediction, and multi-task extensions. Together, these approaches illustrate how generation and prediction can be combined to advance materials design.