Outsmarting quantum chemistry through transfer learning
The transfer learning methods provide a path towards fitting general-purpose atomistic potentials with the accuracy of high level coupled cluster calculations.
Contact
- Justin Smith
The ANI-1ccx potential out performs state-of-the-art force fields and DFT methods at the prediction of torsion profiles for small, drug-like molecules.
Outsmarting quantum chemistry through transfer learning
This work demonstrates empirical potentials based on Deep Neural Networks can surpass the accuracy of Density Functional Theory (DFT) by using Transfer Learning: A potential is first trained to reproduce a large quantity of DFT data, and then fine-tuned by partially retraining to ultra-high-fidelity Coupled-Cluster calculations. The result is the best-to-date empirical model of small organic molecules – far more accurate than traditional Force Fields, and far faster than ab-initio simulation.
