Scaling Foundation Models for Molecular Chemistry

Alexius Wadell, Anoushka Bhutani, Venkatasubramanian Viswanathan

Keywords: foundation model, chemistry, interpretability, neural scaling, compute optimal, property prediction, mixture design

TL;DR: We train a foundation model on 6.14B molecules which achieves SOTA performance across benchmarks, propose Bayesian neural scaling laws enabling compute-optimal molecular foundation model and probe the model to uncover chemical concepts learnt by it.

Confirmation Of Submission Requirements: I submit an abstract. It uses the template provided on the submission page and is no longer than 2 pages.

Submission Number: 233