One of the most transformative applications of machine learning in the physical sciences has been the creation of interatomic potentials that replicate the accuracy of quantum chemical calculations — density functional theory, coupled cluster, or other electronic structure methods — at a fraction of the computational cost. These ML potentials enable molecular dynamics and Monte Carlo simulations of thousands of atoms over nanosecond timescales, accessing regimes of system size and simulation length that were previously intractable without sacrificing chemical accuracy.

This workshop traces the development of ML interatomic potentials from the foundational Behler–Parrinello neural network potentials and their atom-centred symmetry functions, through the ANI family of transferable potentials for organic chemistry and the SchNet architecture with continuous-filter convolutions, to the current state of the art in equivariant architectures — NequIP and MACE — which achieve remarkable data efficiency by encoding the symmetries of three-dimensional space directly into the network architecture.

The session addresses the full workflow of ML potential development: training data generation from electronic structure calculations, active learning strategies for expanding training sets efficiently, uncertainty-aware molec- ular dynamics (FLARE), and validation against experimental and computational benchmarks including radial distribution functions, phonon spectra, melting points, and thermodynamic properties. Participants will gain a thorough understanding of the accuracy–speed trade-off landscape, the practical steps involved in constructing and validating an ML potential for a system of interest, and the current capabilities and limitations of these methods for both molecular and condensed-phase systems.



Extended Key Takeaways & Outputs

• Comprehensive overview of ML potential architectures: Behler–Parrinello, ANI, SchNet, NequIP, MACE

• Understanding of the role of symmetry functions, continuous filters, and equivariant message passing

• Practical knowledge of training data generation, active learning, and uncertainty-aware simulation

• Validation methodology: what to check and what benchmarks reveal

• Assessment of current capabilities and limitations for molecular dynamics and Monte Carlo applications