Julija Zavadlav

Multiscale Materials Modeling with Machine Learning Potentials

Multiscale materials modeling is essential for understanding complex phenomena in fields ranging from life sciences to materials engineering. A prominent research area is the development of machine learning potentials (MLPs), particularly those based on Graph Neural Networks (GNNs), which have emerged as a powerful tool for bridging the gap between quantum-mechanical accuracy and classical molecular dynamics efficiency.

In this presentation, I will showcase the significant achievements of both atomistic and coarse-grained MLPs in effectively capturing many-body interactions. I will address the current challenges of MLP development, including the broad and accurate training dataset generation, capturing long-range interactions, and numerical stability. To address these challenges, we propose a range of innovative strategies that encompass novel training objectives, the synergistic integration of diverse data sources, physics-based GNN architectures, and advanced Bayesian methods for uncertainty quantification. Through insightful case studies of various molecular systems, I will demonstrate the practical effectiveness and versatility of our approaches. Lastly, I will introduce our software platform, chemtrain, designed to streamline the training of machine learning potentials with customizable routines and advanced training algorithms, as well as the extension chemtrain-deploy, enabling scalable parallelization across multiple GPUs and million-atom simulations.