Machine Learning for Predictive Synthesis of Solid-State Materials

The main bottleneck in realizing new materials is the discovery of new multi-stage synthesis routes that can be used to fabricate promising materials. Retrosynthesis is widely used in organic chemistry and self-driving microfluidic laboratories have been demonstrated for high-throughput screening of schedules for liquid-phase molecular and nanoparticle synthesis. There is no comparable scheme for exploring synthesis of condensed phases. As a result, synthesis strategies for promising materials are currently identified by time-consuming trial and error and there are no known predictive schemes for condensed phases and architectures.<br/> <br/>In this talk, I will describe using diffusion models and offline reinforcement learning to predict optimal synthesis strategies for manufacturing functional materials using chemical vapor deposition. These models, trained on molecular dynamics and kinetic Monte Carlo based computational synthesis, learn fabrication parameters including threshold temperatures and chemical potentials for onset of chemical reactions and predicts previously unknown synthesis schedules for heterostructures. The machine-learning driven computational scheme can be extended to predict long-time behavior of reacting systems, far beyond the domain of molecular dynamics simulations, making these predictions directly relevant to experimental synthesis of solid-state materials.