Like many other fields, there has been a recent and overwhelming wave of machine learning and artificial intelligence methods being employed in the chemical sciences. While these methods have the undoubted ability to drive innovation and capabilities, their application to chemical sciences requires a nuanced understanding of molecular representations and structure-property relationships.

In this talk, I will discuss the role of molecular featurization – how we transform atoms and molecules into mathematical signals appropriate for machine-learning thermodynamic quantities – and unsupervised analyses that allow us to easily understand and assess these so-called “featurizations” in the context of complex machine learning tasks. In doing so, I will demonstrate how linear methods – that constitute the simplest, most robust, and most transparent approaches to automatically processing large amounts of data – can be leveraged to understand molecular interactions in both unsupervised and supervised tasks.

I will end by outlining recent work focused on the issue of machine-learning extrapolation – using data and machine-learning methods to predict the thermodynamics of systems outside of our initial training set, a problem in the macromolecular community, where high-quality simulation data is difficult to obtain.