Ab initio potentials for atomistic simulations via deep learning

Computational chemists face a critical challenge in calculating molecular systems' potential energy at the quantum level. While density functional theory provides a solution, it demands substantial computational resources. To address this limitation, researchers have developed machine learning potentials that deliver near-quantum accuracy while requiring only molecular mechanics-level computational power. In this work of thesis we present OBIWAN, a novel feed-forward neural network with unique structural features that includes a new type of general-purpose neural network layer. A key advantage of OBIWAN is its efficient scaling when incorporating unseen atomic species. This design allows users to add new atom types without modifying the network's architecture. As a consequence, OBIWAN can build upon previous training when working with new datasets, leading to rapid convergence. This ability to avoid starting from scratch aligns also with sustainable computing practices by reducing overall computational requirements.