WashU Researchers Use Machine Learning to Automate Chemical Material Discovery

Bridging the Gap Between Molecular Prediction and Laboratory Reality

Researchers at Washington University in St. Louis are developing machine learning methods to model the synthesis of chemical compounds rather than just predicting new structures.

While artificial intelligence models excel at identifying potential molecules from large training datasets, bench chemists and material scientists need practical instructions to manufacture those materials in a laboratory setting.

Zhiling Zheng, an assistant professor of chemistry in Arts & Sciences at Washington University in St. Louis, noted in a prize-winning essay for the journal Science that artificial intelligence systems must learn to read chemistry like a human chemist. According to Zheng, AI is already powerful at predicting new structures, but the primary interest for scientists is making the materials themselves.

Curating a Century of Chemical Recipes

At the Washington University McKelvey School of Engineering, Christopher Cooper published a paper in the journal Matter detailing how to curate large amounts of data for polymer synthesis. Data curation serves as the first major step in the process, requiring researchers to collect and convert information into a format that machine learning models can digest.

According to Cooper and Zheng, existing machine learning models are trained on the known rules of chemistry to make predictions. However, the models lack the application of those rules to run simulations on how to manufacture specific molecules. This requires collecting the recipes of chemical synthesis—instructions buried across a century of journals, textbooks, and footnotes—and translating them for the AI systems.

You want a model to be able to understand those instructions, mash them together and say, ‘this is higher likelihood of being successful.’

Christopher Cooper, Washington University McKelvey School of Engineering

Unlocking Endless Possibilities in Metal Organic Frameworks

Zheng studies metal organic frameworks, known as MOFs, which feature metal ions as nodes and organic blocks as linkers, creating cube-shaped structures with endless potential uses. Because the design possibilities for MOFs number in the millions, it is impossible for human researchers to run every experiment manually.

To address this open-ended challenge, Zheng proposed a new methodology utilizing large-language models to train AI agents much like a graduate student. Zheng and his team trained language models on a dataset of approximately 4,000 linker transformations of MOFs derived from scientific literature.

Accelerating Discovery Through Autonomous Simulation

Using computational simulations, an AI agent filtered the 4,000 possibilities based on chemical constraints and identified 10 new viable materials. These newly discovered materials demonstrate stronger water-harvesting performances than state-of-the-art aluminum-based adsorbents, bypassing the labor-intensive brute-force linker screening process.

Cooper is currently leading a similar project focused on polymer design.