Atomistic Potential Development

We aim to develop computational methods that can accurately predict and model nanoscale materials and chemical reactions, potentially containing thousands of atoms.

---

After John's sabbatical with Alexandre Tkatchenko at the University of Luxembourg we have been investigating symmetrized gradient-domain machine learning (sGDML) for solvated reaction mechanisms. sGDML force fields have demonstrated molecular dynamics with CCSD(T) accuracy. ML force fields, however, are limited to the system they are trained on. To sidestep this issue, we apply the sGDML approach to learn fundamental interactions of monomer, dimer, trimer, and tetramer solvent clusters. We have found that with only hundreds of training data points up to 3-body interactions we can achieve reasonably accurate and transferable molecular dynamics simulations on larger solvent clusters of arbitrary size.

In collaboration with John Kitchin's (Carnegie Mellon University-ChemE) group, we have studied how ReaxFF and Behler–Parrinello neural network (BPNN) atomistic potentials should be trained to be accurate and tractable across multiple structural regimes of gold as a representative example of a single-component material. We trained these potentials using subsets of 9,972 Kohn-Sham density functional theory calculations and then validated their predictions against the untrained data. Our best ReaxFF potential was trained from 848 data points and could reliably predict surface and bulk data; however, it was substantially less accurate for molecular clusters of 126 atoms or fewer. Training the ReaxFF potential to more data also resulted in overfitting and lower accuracy. In contrast, BPNN could be fit to 9,734 calculations, and this potential performed comparably or better than ReaxFF across all regimes. However, the BPNN potential in this implementation brings significantly higher computational cost.