Physics-informed Machine Learning (D2.2)

We will design hybrid physics-guided ML potentials, e.g., to dramatically accelerate atomistic simulations. In addition, we will exploit the concept of chemical connectivity in highly accurate graph neural network models, such as coGN or NequIP, to allow the simulation of large systems (tens of thousands of atoms) over large timescales (up to ns and more), far beyond the limit of electronic-structure-based simulations and far beyond the accuracy of force-field-based methods. We will train graph-neural-network (GNN) models for selected systems developed in Research Area A to better understand nanoscale processes relevant during printing and in applications. These efforts are tightly interlinked with LP2.