Machine Learning Force Fields

In molecular simulation, a force field is the mathematical model that describes how atoms attract and repel each other; it determines every force, every trajectory, and every predicted property. Classical force fields use simple equations fitted to experimental data: fast to evaluate, but approximate. Quantum-mechanical calculations are far more accurate but too expensive for large systems or long simulation times. Machine learning force fields (MLFFs) bridge this gap: neural networks trained on quantum-mechanical data learn to predict interatomic forces with near-quantum accuracy at a small fraction of the computational cost. We develop and apply MLFF architectures (DeePMD, NequIP, MACE) to study systems that are too large for quantum calculations and too sensitive to classical approximations, particularly water in electrolyte solutions and reactive chemistry in extreme environments.

Water diffusion anomalies: Mapping how water molecules move differently near different dissolved ions in salt solutions, effects that classical force fields fail to distinguish but that machine learning models trained on quantum data reproduce accurately

Dielectric relaxation: How liquid water's electrical polarization responds to oscillating electric fields across the microwave-to-infrared spectrum, comparing neural network predictions to classical force field results

Reactive chemistry in supercritical water: Tracking how organic molecules break apart and recombine in water heated above 374 °C and pressurized above 22 MPa, conditions where standard force fields cannot describe bond breaking and formation

Active learning and transferability: Developing strategies for efficiently selecting which quantum-mechanical calculations to run, so that the resulting force field generalizes reliably across temperatures and chemical compositions

We train MLFFs on quantum-mechanical reference data from density functional theory (DFT) calculations using neural network architectures that respect physical symmetries; predictions remain unchanged when the molecule is rotated or translated. Combined with enhanced sampling techniques that accelerate the observation of rare events, these force fields reach nanosecond simulation time scales with near-quantum accuracy.

Yu et al., Science Advances (2024), Water diffusion anomalies via MLFF

Ryu et al., J. Mol. Liquids (2024), Dielectric relaxation of water

Ryu et al., J. Chem. Inf. Model. (2025), Acetic acid in supercritical water