Data-Efficient Machine Learning Potentials from Transfer Learning of Periodic Correlated Electronic Structure Methods: Liquid Water at AFQMC, CCSD, and CCSD(T) Accuracy

Abstract

Obtaining the atomistic structure and dynamics of disordered condensed-phase systems from first-principles remains one of the forefront challenges of chemical theory. Here we exploit recent advances in periodic electronic structure and provide a dataefficient approach to obtain machine-learned condensed-phase potential energy surfaces using AFQMC, CCSD, and CCSD(T) from a very small number (≤200) of energies by leveraging a transfer learning scheme starting from lower-tier electronic structure methods. We demonstrate the effectiveness of this approach for liquid water by performing both classical and path integral molecular dynamics simulations on these machine-learned potential energy surfaces. By doing this, we uncover the interplay of dynamical electron correlation and nuclear quantum effects across the entire liquid range of water while providing a general strategy for efficiently utilizing periodic correlated electronic structure methods to explore disordered condensed-phase systems.

Transfer learning for chemically accurate interatomic neural network potentials

“…[29][30][31][32] In this work, we propose an alternative approach to ref. [38][39][40][41][42][43][44][45] and investigate the performance of interatomic NN models, which have been trained only on energy labels during the fine-tuning step, on atomic forces.…”

“…This behavior is unexpected and has not been observed previously. [38][39][40][41][42][43][44][45] As a possible explanation, we hypothesize that the large number of iterations during pre-training on a large pre-training data set trains the NN to be able to overfit some details of the data more easily, which enables it to overfit the fine-tuning data set more easily. This would suggest that decreasing the number of pre-training epochs when pretraining on very large data sets may allow to circumvent this phenomenon.…”

“…In summary, we extend the observations of ref. [38][39][40][41][42][43][44][45] regarding the improvement in sample-efficiency by also studying smaller data sets, differences between force-and-energy and energy-only fine-tuning, the effect of the pre-training set size, and the behavior in molecular dynamics simulations.…”

“…While in ref. 38–44 some hidden layers have been fine-tuned, the approach proposed in ref. 45 employs linear probing, i.e.…”

“…In ref. 44, the developed models were applied to simulate bulk liquid water at various ab initio levels of theory. Ref.…”