Inventing new materials can be costly and time-consuming, with millions of variables involved. Finding materials with specific arrangements of atoms that define their crystal structures and engineering properties is incredibly valuable within material science.

Toyota Research Institute (TRI) has developed an innovative tool, Computational Autonomy for Materials Discovery, CAMD, a cloud-computing platform designed to improve the discovery of new materials. CAMD uses artificial intelligence to predict which test simulations to run. The system reduces the potential number of simulations from millions to thousands, providing a playground to design, test, and optimize agents to shorten computational experiment time and lower costs.

CAMD, an open-source software package, taps into material modeling available through machine learning and density functional theory (DFT) to determine which hypothetical materials to probe from potentially billions. This significantly reduces the expense of cloud-computing-based experimentation, which can cost one hundred dollars per material. CAMD has discovered roughly 30,000 new compounds that are likely synthesizable, and its simulations provide invaluable data on which element combinations to try in the lab.

The next step is the difficult task of producing the physical materials through real-world experimentation. Python Inorganic Reaction Organizer (Piro) is an open-source application that builds on CAMD’s discovery powers. It combines machine learning with a new physical model to make recommendations on synthesizing materials in a real-world laboratory. Machine learning helps solve this by predicting which experiments most closely follow rules that pair similar materials with similar ingredients for synthesis, and physics helps predict how quickly different combinations of ingredients will result in the desired material. Although the synthesis process remains largely unexplored, predicting outcomes using old and new techniques is still possible.

As the name suggests, Piro compares different sets of reactants – or precursors – that might result in a particular crystal when mixed and heated to high temperatures. It then predicts which reactions are more likely to form the target crystalline compound.

Ultimately, predictions from these tools must be tested through physical lab experiments where compounds are synthesized, characterized, and tested in actual devices, such as lithium-ion batteries, fuel cells, and solar cells. We have begun this journey and shown that computer-predicted materials can be realized in the lab, but the rate of physical verification remains slow. More exciting work in this space is underway.

Find more detailed information at these GitHub links:

• https://github.com/TRI-AMDD/CAMD
• https://github.com/TRI-AMDD/piro

Montoya J, Grimley C, Aykol M, Ophus C, Sternlicht H, Savitzky BH, et al. Computer-assisted discovery and rational synthesis of ternary oxides. ChemRxiv. 2023. This content is a preprint and has not been peer-reviewed.

Ye, W., Lei, X., Aykol, M., Montoya, J.H. Novel inorganic crystal structures predicted using autonomous simulation agents. Scientific Data 2022, 9 (302).

Aykol, Muratahan, Joseph H. Montoya and Jens Hummelshoej. “Rational Solid-State Synthesis Routes for Inorganic Materials.” American Chemical Society Publications, June 2021.

Montoya, JH, Winther, K, Flores, RA, Bligaard, T, Hummelshoj, JS, Aykol M. "Autonomous intelligent agents for accelerated materials discovery." Chemical Science, 2020.