Quantum (Bio) Molecular Simulations with Machine Learning Force Fields
Presenter
DescriptionMachine learning force fields (MLFFs) promise to bridge the gap between quantum-mechanical accuracy and the computational efficiency needed to simulate realistic (bio)molecular systems [1]. Yet their predictive power is often limited by the quality and coverage of training data, as well as by locality assumptions that miss the long-range effects governing molecular structure and dynamics. In this talk, I will present two contributions aimed at addressing these limitations. First, I will present SO3LR [2], a pretrained MLFF that couples an SO(3)-equivariant neural network with universal pairwise potentials for long-range electrostatics and dispersion. Second, I will introduce QCell [3], a quantum-mechanical dataset of ~0.5M diverse molecular fragments extending chemical space coverage of cellular components, designed to provide the breadth of data needed to train truly general-purpose models. Selected examples will illustrate how these advances enable simulations of complex (bio)molecular systems with near-ab initio accuracy. I will conclude with a discussion of current limitations and future directions.
[1] Unke et al., Chem. Rev. 2021, 121, 16, 10142; https://doi.org/10.1021/acs.chemrev.0c01111
[2] Kabylda et al., J. Am. Chem. Soc. 2025, 147, 37, 33723; https://doi.org/10.1021/jacs.5c09558
[3] Kabylda et al., arXiv:2510.09939 2026; https://doi.org/10.48550/arXiv.2510.09939
[1] Unke et al., Chem. Rev. 2021, 121, 16, 10142; https://doi.org/10.1021/acs.chemrev.0c01111
[2] Kabylda et al., J. Am. Chem. Soc. 2025, 147, 37, 33723; https://doi.org/10.1021/jacs.5c09558
[3] Kabylda et al., arXiv:2510.09939 2026; https://doi.org/10.48550/arXiv.2510.09939
TimeMonday, June 2914:00 – 14:30 CEST
LocationBldg. 8 – B 102
SessionMS1H – From Physics to Data – and Back: Trustworthy Machine Learning Potentials for Materials Design
Session Chairs
Event Type
Minisymposium
