Improving the Accuracy of Machine Learning Force Field Model of Water

The structure and dynamics of water is a longstanding topic of interest in chemical physics. While ab initio wavefunction theories can deal with water molecules quite accurately, applying these theories to bulk phases is yet to be achieved. The advent of machine learning force field now enables simulation of bulk water phases with ab initio accuracy, but mostly at the level of density functional theory. In this work, we employ our newly developed ByteQC package, which allows large scale CCSD(T) wavefunction calculations using efficient quantum embedding schemes, to compute the energy of water clusters containing 32-64 molecules. These accurate data are used to train a new machine learning force field model by employing a stepwise fine-tuning strategy. We test this new water model on various physical properties of different phases of water, demonstrating the potential of resolving the complex structure and dynamics of water with ab initio accuracy.