Predicting a new material on paper has become almost routine. Actually making it in the lab is where the trouble starts. Researchers at Lawrence Berkeley National Laboratory have now built an AI model that tackles the harder half of the problem, forecasting how a solid-state reaction unfolds over time rather than just naming the compound you hope to end up with. The work was published this month in Nature Materials and announced by the lab on 3 August.
The model pairs thermodynamics with machine learning to predict the intermediate compounds, the final product, and the impurities that show up along the way. That last part matters more than it sounds. Earlier models could tell you what was stable at the end but ignored kinetics, the plodding movement of atoms through a solid on the way there. "Atoms in solids move more slowly than in liquids, making it harder for them to reach reaction sites," said Kristin Persson, the senior Berkeley Lab scientist and UC Berkeley professor who led the team.
Decades of data, results in minutes
To test it, the group turned the model loose on barium-titanium oxides and checked its predictions against decades of real synthesis records. It reconstructed full reaction pathways in minutes, the kind of thing that normally takes a graduate student months of trial and error at the bench. Persson said the approach can "help accelerate the advancement of solid materials to cost-effective manufacturing and commercialization," which is the step where most promising materials quietly stall.
Toward a general model for solids
The team plans to widen the model beyond oxides and, eventually, build a foundation model that can handle almost any solid-state material. The research was funded by the Department of Energy's Office of Science. It fits a broader shift this year, with AI moving from analysing experiments after the fact to proposing which experiment to run next. The useful discipline here is that humans and instruments still verify every result. A model that guesses a synthesis route is helpful. A model whose routes actually work in a furnace is worth a great deal more, and this one was checked against materials people have already made.
Commentarii · 0