Researchers at Lawrence Berkeley National Laboratory have developed an artificial intelligence (AI) approach that accurately predicts how chemical reactions between solid powders unfold when heated. The method is the first to account for the solid-state reactions through the materials during these processes. These are chemical changes that occur when solid powders are mixed and heated together rather than dissolved in liquids. The predictions include intermediate compounds that appear along the way, the final products, and any impurities, which are unwanted extra substances that form alongside the desired material.
Synthesizing useful solid materials for technologies such as batteries, sensors, and medical devices has long been a slow process. Scientists can now use computers to find materials with promising properties, yet turning those ideas into real samples often requires weeks or years of trial-and-error mixing and heating. Existing computer models relied mainly on thermodynamic principles that indicate which reactions are preferred because they lead to more stable products. However, these models frequently failed because they ignored kinetics, the rates at which atoms actually travel through the solids to reach the places where reactions can occur.
Kinetics at work
Atoms move more slowly in solids than in liquids, so even preferred reactions may not take place unless the right number of atoms can reach the reaction sites. The new approach adds a machine learning component trained to estimate these travel rates. It assumes that the boundary between reacting solids is highly disordered, similar to a crowded exit after a large event. Users supply the starting materials, their amounts, and the heating schedule. Within minutes the model simulates the entire sequence of events.
The approach was tested on barium-titanium oxides, a group of compounds used in electronics whose formation is strongly controlled by atom movement rates. Simulations of different mixtures and temperatures matched decades of laboratory results, correctly identifying intermediate compounds, final products, and impurities. The same methods can now be extended to other classes of solid materials, with the goal of creating a broadly useful tool that shortens the path from discovery to practical manufacturing.
This research was published in Nature Materials.