Artificial Intelligence and Machine Learning

Foundation Models for Discovery and Exploration in Chemical Space Molecular foundation models will hopefully accelerate scientific discovery by making the prediction of molecular properties possible at scale, enabling screening of hundreds of millions of candidate molecules for a given application with a relatively small computational cost. Wadell et al not only demonstrate how to implement a molecular foundation model for battery electrolyte discovery, defining a new ‘molecular surpriseʼ metric to guide the search for novel compounds, but also describe in detail how to build the data set and train the foundation model in a computationally efficient manner.

Screening of 90 million electrolyte solvent molecules using 8 NVIDIA A100 GPUs in 8 hours (estimated cost of $160 to $260 USD on AWS, p4d.24xlarge instance; screened candidates would need further validation using DFT, MD, and experiment). While there is no ground truth for many properties of the evaluated molecules, predicted properties conform to known physical relationships.

(c) 63 candidates are found that push both chemical and electrochemical stability. (b) ‘Surprisingʼ molecules push boundaries, while many reference electrolytes that are aligned with design rules currently used by electrolyte scientists are less surprising.

Molecular foundation models enable rapid and cost-effective screening of 100s of millions of molecules to empower the discovery of new materials. Source: Wadell et al, arXiv, 2025, 2510.18900

Extracted from the PDF; charts and images may read imperfectly.