Obtaining transferable chemical insight from solving machine-learning classification problems: Thermodynamical properties prediction, atomic composition as good as Coulomb matrix

Abstract

Machine learning (ML) can be used to construct surrogate models for the fast prediction of a property of interest. ML can thus be applied to chemical projects, where the usual experimentation or calculation techniques can take hours or days for just one sample. In this manner, the most promising candidate samples could be extracted from an extensive database and subjected to further in-depth analysis. Despite their broad applicability, it can be challenging to apply ML methods to a given chemical problem since a multitude of design decisions must be made, such as the molecular descriptor to use or the optimizer to train the model. Here we present a methodology for the meaningful exploration of a given molecular problem through classification experiments. This conceptually simple methodology results in transferable insight on the selected problem and can be used as a platform from which prediction difficulty is estimated, molecular representations are tested and refined, and more precise or ambitious projects can be undertaken. Physicochemical insight can also be obtained. This methodology is illustrated through the use of multiple molecular descriptors for the prediction of enthalpy, Gibbs' free energy, zero-point vibrational energy, and constant-volume calorific capacity of the molecules from the public database QM9 [Ramakrishnan2014] with 133,885 organic molecules. A noteworthy result is that for the classification problem we propose, the low-resolution descriptor `atomic composition' [Tchagang2019] can reach a classification rate almost on par with the high-resolution `sorted Coulomb matrix' [Rupp2012,Montavon2012,Hansen2013] ($>90\%$), provided that an appropriate optimizer is used during training.