(66d) Liquid-Liquid Phase Transition in Water from a Machine-Learning First Principles Force Field | AIChE

(66d) Liquid-Liquid Phase Transition in Water from a Machine-Learning First Principles Force Field


Gartner, T. E. III - Presenter, University of Delaware
Zhang, L., Princeton University
Piaggi, P. M., Princeton University
Car, R., Princeton University
Debenedetti, P., Princeton University
Panagiotopoulos, A., Princeton University
Finding an explanation for water's anomalous thermophysical properties (e.g., liquid water's observed maximum in density, and sharp increases in compressibility, heat capacity, and the magnitude of the thermal expansion coefficient upon cooling) is an area of vigorous research activity and debate. One theory is that under strongly supercooled conditions and high pressures, water undergoes a liquid-liquid phase separation into metastable high-density and low-density liquid states, and the thermodynamic implications of this transition can be traced to explain some of water's anomalies. Computer simulation techniques have led the charge in studying this phenomenon, as they are not in general affected by rapid crystallization, the occurrence of which in experiments places the temperatures and pressures where the liquid-liquid transition is theorized to occur outside the range accessible by current experimental techniques. As such, the search for ever more advanced computational approaches to definitively establish (or disprove) water's liquid-liquid transition is an active area of research. Herein, we use the Deep Potential Molecular Dynamics (DPMD) technique to generate a neural network-based model of water obtained from high-accuracy density functional theory calculations. We then use biased DPMD simulations in the multithermal-multibaric ensemble to efficiently obtain the thermophysical properties of this model over a wide range of temperatures and pressures. We analyze these results in the context of a phenomenological two-state equation of state, and we discuss how our findings provide ab-initio evidence consistent with a liquid-liquid transition. Lastly, we describe ongoing efforts to apply free energy techniques to definitely establish liquid-liquid coexistence using this DPMD model. This work brings state-of-the-art machine learning, ab-initio, and statistical mechanical methods to significantly deepen our understanding of the sources of water's unusual properties, bolstering the hypothesis according to which this most ubiquitous yet unusual substance possesses two critical points.