(162a) The Outlook for Physically Informed Machine Learning in Computational Molecular Science

We are witnessing an explosion of interest and activity in the application of artificial intelligence and machine learning areas related to chemical engineering, from determining force fields to materials discovery. We will provide some examples of these applications to look at some studies from “early adopters” to provide a perhaps eclectic snapshot of the current state-of-the-art. This will include some of our own work to apply novel Bayesian Optimization techniques that we have developed to tackle particularly challenging materials discovery problems. These problems include solvent engineering for the solubilization of lead salts in complex organic-inorganic cationic solutions, polymorph predictions in organic semiconductors, and an accelerated route to defect engineering in inorganic electronic materials. These applications are strikingly successful and they inspire confidence in the usefulness of machine learning tools to make dramatic and potentially disruptive changes to materials discovery. However, these studies also uncover some important challenges to their expansion and exploitation. This can occur in the form of traditional and scope-defining chemical engineering concerns, such as scale-up. But the studies also highlight the enduring role of physical insight and intuition and the need to codify these very human traits. The use of these techniques also raises metaphysical questions related to our concept of understanding physical processes and our willingness to expand our trust in machine learning-derived predictions that exceed our ability to test them either experimentally or computationally.