Understanding Chemical Trends in Generated Databases of Functional Molecules

Understanding Chemical Trends in Generated Databases of Functional Molecules

Chemistry

Francesco Bartucca

Reinhard Maurer, Zsuzsanna Koczor-Benda

This project aims to investigate the output of G-SchNet, a generative machine learning model used by many in the field of computational chemistry. Discrepancies between the distributions of generated and training data are highlighted, in particular a tendency to generate molecules which contain more heavy atoms compared to the training datasets. Databases used for training include a thiols set composed of commercially-available molecules, and the more widely used OE62 and QM9 sets.

Chemistry, Machine Learning, Computational Chemistry, Data Analysis