Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 22: Data-driven Materials Science: Big Data and Workflows III

MM 22.4: Vortrag

Mittwoch, 11. März 2026, 11:00–11:15, SCH/A216

Uncertainty Propagation in Machine-learned Interatomic Potentials — •Haitham Gaafer, Jan Janssen, and Jörg Neugebauer — Computational Materials Design, Max-Planck-Institute for Sustainable Materials, Düsseldorf

Accurate multiscale materials modeling requires that uncertainties be quantified and propagated consistently from the electronic-structure level to macroscopic property predictions. Machine-learned interatomic potentials (MLIPs), trained on density-functional theory (DFT) reference data, now routinely reach near-DFT accuracy at dramatically reduced computational cost. Yet the connection between fitting errors in an MLIP and uncertainties in derived physical properties, such as bulk moduli or phase stabilities, remains insufficiently understood. We present a data-driven pyiron workflow designed to analyze how uncertainties originating in MLIP training propagate into thermomechanical property predictions. As a case study, we construct diverse DFT training sets for Cu, Ag, and Au using the Automated Small SYmmetric Structure Training (ASSYST) workflow, and fit computationally efficient atomic cluster expansion (ACE) potentials employing a minimal basis optimized to reach a target root-mean-square error. These potentials are subsequently used to determine equations of state and to quantify uncertainties in key properties, including the equilibrium lattice constant, bulk modulus, and its pressure derivative. Our results provide a transparent link between MLIP fitting quality and property reliability, offering a systematic route for uncertainty-aware atomistic modeling.

Keywords: Machine-learning; Interatomic; Potentials; Uncertainty; Propagation