Berlin 2024 – scientific programme

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

DY: Fachverband Dynamik und Statistische Physik

DY 34: Poster: Machine Learning, Data Science, and Reservoir Computing

DY 34.17: Poster

Wednesday, March 20, 2024, 15:00–18:00, Poster C

Balancing short- and long-range interactions in Machine Learning Force Fields — •Tobias Henkes, Igor Poltavsky, and Alexandre Tkatchenko — Université du Luxembourg

Nowadays, highly accurate and data-efficient Machine Learning (ML) Force Fields (FFs) are an established method in computational chemistry and physics. MLFFs can enable simulations of large, complex systems with quantum chemical accuracy by using predefined molecular descriptors or learned representations. However, even state-of-the-art ML models can struggle with handling long- and short-range interactions equally. They often employ a locality assumption, naturally emphasizing short ranges. To overcome this constraint, we embed a coarse-grained description of the global environment into an accurate local atomic representation. Effectively, the ML model now has a balanced picture of local and global features and uses this to simultaneously reproduce long- and short-range effects. We showcase the proposed methodology in the Gradient Domain Machine Learning (GDML) framework. Herein, we utilize a hierarchical descriptor that includes local and global features with adjustable attention weights. The modified GDML approach shows an improved accuracy compared to the default architecture for extended systems such as solvated molecules and liquid water. Ultimately, the developed embedding approach can aid any MLFF model in striking a balance between short- and long-range interactions.

Keywords: Machine Learning; Force Fields; Molecular Dynamics; Long-range interactions