Signatures of a liquid–liquid transition in an ab initio deep neural network model for water

• DOI: 10.1073/pnas.2015440117
• Publication Date: 2020-10-03T00:25:38Z
• Authors: Thomas E. Gartner; Linfeng Zhang; Pablo M. Piaggi; Roberto Car; Athanassios Z. Pa...; Pablo G. Debenedetti
• Keywords: Metastability; Supercooling; Liquid water; Potential energy surface

The possible existence of a metastable liquid-liquid transition (LLT) and corresponding critical point (LLCP) in supercooled liquid water remains topic much debate. An LLT has been rigorously proved three empirically parametrized molecular models water, evidence consistent with an reported for several other such models. In contrast, experimental proof this phenomenon elusive due to rapid ice nucleation under deeply conditions. work, we combined density functional theory (DFT), machine learning, simulations shed additional light on the water. We trained deep neural network (DNN) model represent ab initio potential energy surface from DFT calculations using Strongly Constrained Appropriately Normed (SCAN) functional. then used advanced sampling multithermal-multibaric ensemble efficiently explore thermophysical properties DNN model. simulation results are LLCP, although they do not constitute rigorous thereof. fit data two-state equation state provide estimate LLCP's location. These results-obtained purely first-principles approach no empirical parameters-are strongly suggestive LLT, bolstering hypothesis that can separate into two distinct forms.