Teaching machines physics and chemistry
Quantum-mechanical calculations tell us how atoms push and pull on each other, but they are far too slow for the system sizes and timescales that decide whether a material survives in a real device. I train machine-learned interatomic potentials on DFT data, so that molecular dynamics can run at close to first-principles accuracy for a fraction of the cost.
For CALF-20, a benchmark MOF for industrial-scale CO₂ capture, this approach revealed negative thermal expansion, negative area compression, a two-step elastic response and a fracture strain of up to 27%, all properties that standard force fields could not have predicted. I release the potentials and training data openly, and I use the same data-driven mindset elsewhere: screening candidate frameworks, and clustering large simulation datasets to find the descriptors that matter.
Nature Communications 2024 · ACS Appl. Eng. Mater. 2024 · ML force field (Zenodo)