Regensburg 2025 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 3: Data-driven Materials Science: Big Data and Worksflows

MM 3.5: Talk

Monday, March 17, 2025, 11:15–11:30, H10

Diversity-Driven Active Learning of Interatomic Potentials for Reaction Network Exploration — •Francesco Cannizzaro, King Chun Lai, Patricia Poths, Sebastian Matera, Vanessa J. Bukas, and Karsten Reuter — Fritz-Haber-Institut der MPG, Berlin

We present an automatic workflow for the simultaneous active learning of Machine-Learning Interatomic Potentials (MLIPs) and exploration of complex networks of activated events. This workflow consists of alternating periods of training and the generation of candidate structures for the enrichment of the training set using the recently developed Automatic Process Explorer (APE) [1]. This allows us to determine elementary processes and corresponding barriers without the need of human supervision. From the output of the APE explorations, we identify maximally diverse atomic structures utilizing the DECAF fuzzy classification algorithm [2] and add only these to the training set. We exemplify this strategy on carbon intercalation in Pd, using GAP and MACE as MLIP frameworks. We find that this diversity driven approach outperforms state-of-the-art training set designs based on molecular dynamics for finding activated events and corresponding barriers. Particularly, our workflow performs very well in reducing outliers, which is of utmost importance for activated event dynamics since this is often controlled by only a few barriers.

[1] Lai et al., ChemRxiv, https://doi.org/10.26434/chemrxiv-2024-jbzr7 (2024).

[2] Lai et al., J. Chem. Phys. 159, 024129 (2023).

Keywords: Machine-Learning Interatomic Potentials (MLIPs); Active Learning; Automatic process exploration; Catalytic Restructuring