Predicting that a material should exist is now almost routine; working out how to make it remains the bottleneck. MIT researchers led by PhD candidate Elton Pan, with faculty including Elsa Olivetti and Rafael Gomez-Bombarelli, have attacked that bottleneck with DiffSyn, a generative model described in Nature Computational Science that proposes synthesis recipes for a target structure, specifying reaction temperatures, times and precursor ratios.
The team trained the diffusion model on more than 23,000 synthesis recipes extracted from 50 years of scientific literature. Where earlier approaches tied each structure to a single canonical route, DiffSyn maps each structure to many plausible ones, reflecting how synthesis actually works in practice. Benchmarked on zeolites, a notoriously complex family of microporous materials used in catalysis and ion exchange, the model achieved state-of-the-art prediction accuracy and can sample 1,000 candidate recipes in under a minute.
The proof lies in the furnace: following the model's suggestions, the researchers synthesised a new zeolite with improved thermal stability. Since zeolite crystallisation can take days or weeks per attempt, a tool that ranks promising conditions before anyone touches a crucible could compress discovery timelines substantially. The group plans to extend the approach to other materials classes, including metal-organic frameworks.