Characterization of Quantum Derived Electronic Properties of Molecules: A Computational Intelligence Approach
Julio J. Valdés, Alain Tchagang · Lecture notes in computer science · 2019
Abstract The availability of BIG molecular databases derived from quantum mechanics computations represent an opportunity for computational intelligence practitioners to develop new tools with same accuracy but much lower computational complexity compared to the costly Schrödinger equation. In this study, unsupervised and supervised learning methods are applied to investigate the internal structure of the data and to learn the mapping between the atomic coordinates of molecules and their properties. Low dimensional spaces revealed a well defined clustering structure as defined by the measures used for comparing molecules based their atom distributions and chemical composition. Supervised learning techniques were applied on the original predictor variables, as well as on a subset of selected variables found using evolutionary algorithms guided by residual variance analysis (Gamma Test). Black and white box modeling approaches were used (random forests, neural networks and model trees and adaptive regression respectively). All of them delivered good performance, error and correlation-wise, with neural networks producing the best results. In particular white box techniques obtained explicit functional dependencies, some of them achieving considerably reduction of the feature set and expressed as simple models.