AARMS Scientific Machine Learning Seminar: Stijn De Baerdemacker (UNB)

The task of quantum chemistry is to compute chemical properties of molecular compounds from the fundamental laws of quantum mechanics.  This poses a tremendous challenge as the most accurate methods also come with a perniciously scaling computational cost, limiting systematic and meaningful calculations mostly to medium-small molecules.  This is an undesirable situation as there is a high potential for computationally driven discovery of chemical compounds. For this reason, there have been significant advances in the design of Machine Learning (ML) algorithms to interpolate chemical observables from precomputed data, rather than engaging into costly computations.  At present, the ML methods of choice in (quantum) chemistry are based upon Graph Convolutional Neural Networks (GCNN), as molecular compounds are easily reinterpreted into graph-based structures.

In this presentation, I will discuss recent results from the QuNB group on the analysis of the feature space of SchNet, a popular GCNN for quantum chemistry with high predictive power.  I will show that chemical information from chemical functional groups is learned and encoded in the feature space of the GCNN, giving a strong quality hallmark of interpretability to the GCNN ML algorithm.

Joint work from the QuNB group (UNB) with Guillaume Acke (Ghent University)

Webex link: https://mun.webex.com/mun/j.php?MTID=m8b04b29275dfeb0c3617be2958e577ff