MS1H – From Physics to Data – and Back: Trustworthy Machine Learning Potentials for Materials Design
Session Chairs
Event Type
Minisymposium
Chemistry and Materials
Physics
Computational Methods and Applied Mathematics
TimeMonday, June 2913:30 – 15:30 CEST
LocationBldg. 8 – B 102
DescriptionOrganizer(s): Michał Sanocki, Ian Störmer, Julija Zavadlav Koller (Technical University of Munich), and Philip Robin Loche (EPFL)
Machine learning interatomic potentials (MLIPs) are transforming electronic-structure modeling by delivering near–quantum mechanical accuracy at greatly reduced cost. As these models increasingly surrogate explicit electronic-structure calculations, especially for large, complex, or heterogeneous systems, the central challenge becomes establishing trust when direct quantum-mechanical validation is no longer feasible. Balancing data-driven flexibility with physical consistency is essential: symmetry-preserving and physics-informed architectures enhance reliability, while more flexible models demand rigorous verification to ensure they learn the correct physical relationships. Exascale computing reshapes this landscape by enabling large-scale data generation, systematic cross-validation, and interrogation of MLIP behavior under extreme conditions. At the same time, the complexity of training and deploying these models highlights the need for reproducibility, explainability, and robust uncertainty quantification. Emerging approaches aim to embed trust checks – such as enforcing conservation laws, equivariance, and stability-directly into training rather than relying solely on post hoc validation. This minisymposium will gather researchers from chemistry, materials science, physics, and computer science to discuss strategies for developing trusted MLIPs. Topics include physics-informed representations, equivariant architectures, long-range and non-conservative interactions, uncertainty estimation, and automated verification workflows. The session aims to chart pathways toward reliable, scalable, and transparent MLIP frameworks for molecular and materials simulation.
Machine learning interatomic potentials (MLIPs) are transforming electronic-structure modeling by delivering near–quantum mechanical accuracy at greatly reduced cost. As these models increasingly surrogate explicit electronic-structure calculations, especially for large, complex, or heterogeneous systems, the central challenge becomes establishing trust when direct quantum-mechanical validation is no longer feasible. Balancing data-driven flexibility with physical consistency is essential: symmetry-preserving and physics-informed architectures enhance reliability, while more flexible models demand rigorous verification to ensure they learn the correct physical relationships. Exascale computing reshapes this landscape by enabling large-scale data generation, systematic cross-validation, and interrogation of MLIP behavior under extreme conditions. At the same time, the complexity of training and deploying these models highlights the need for reproducibility, explainability, and robust uncertainty quantification. Emerging approaches aim to embed trust checks – such as enforcing conservation laws, equivariance, and stability-directly into training rather than relying solely on post hoc validation. This minisymposium will gather researchers from chemistry, materials science, physics, and computer science to discuss strategies for developing trusted MLIPs. Topics include physics-informed representations, equivariant architectures, long-range and non-conservative interactions, uncertainty estimation, and automated verification workflows. The session aims to chart pathways toward reliable, scalable, and transparent MLIP frameworks for molecular and materials simulation.
Presentations
