Recent advances in machine learning have resulted in an upsurge of interest in developing a “quantum machine”, a technique of simulating and predicting quantum-chemical properties on the molecular level. This paper explores the development of a large-scale quantum machine in the context of accurately and rapidly classifying molecules to determine photovoltaic efficacy through machine learning. Specifically, this paper proposes several novel representations of molecules that are amenable to learning, in addition to extending and improving existing representations. This paper also proposes and implements extensions to scalable distributed learning algorithms, in order to perform large scale molecular regression. This paper leverages Harvard’s Odyssey supercomputer in order to train various kinds of predictive algorithms over millions of molecules, and assesses cross-validated test performance of these models for predicting photovoltaic efficacy. The study suggests combinations of representations and learning models that may be most desirable in constructing a large-scale system designed to classify molecules by photovoltaic efficacy.
