
Arthur Kosmala
Data Analytics and Machine Learning
Molecular Modelling Based on Neural Networks with Long-Range Methods from Physics, Chemistry, and Materials Science
Molecular modelling has significantly enhanced our capability to predict application-critical properties of molecules and materials via small-scale insights into their atomistic behaviour. Quantum chemistry calculations from first principles offer superb accuracy but rarely scale beyond systems a few hundred atoms in size. Motivated by this performance gap, recent years experienced a growing number of methods leveraging quantum-chemical reference data to train neural-network surrogate models. This approach holds the promise of retaining first-principles accuracy while accelerating inference by orders of magnitude.
Neural networks that process molecular geometries often perform a set of convolutions over the atomic point cloud or molecular graph, where each convolution is limited by a pre-specified range. This standard practice of truncating interactions is in stark contrast to chemical reality, by which the properties of many diverse and functionally relevant systems are dictated by highly non-local interactions such as electrostatics and dispersion. In my doctoral research, I aim to address this issue by designing neural network architectures directly inspired by rigorous long-range techniques from physics, chemistry, and materials science. My past and present work focuses on methods combining real- and frequency-space representations in a joint architecture – quite analogous to the classic technique of Ewald summation. Given that my primary background is in theoretical physics, I am naturally drawn towards interdisciplinary work and greatly appreciate the MDSI as a collaborative space enabling such discussion.