First principles simulations of the liquid-liquid transition in water using deep neural networks

A metastable liquid-liquid transition (LLT) has been observed in simulations of several molecular models of water, but experimental evidence for or against this phenomenon remains elusive due to rapid ice nucleation under supercooled conditions. As such, attempting to definitively identify water's LLT is an active area of study in simulations and experiments. In this work, we used the Deep Potential Molecular Dynamics (DPMD) technique to generate an ab initio deep neural network (DNN) model for water based on density functional theory calculations with the SCAN exchange correlation functional. We then performed enhanced-sampling molecular simulations in the multithermal-multibaric ensemble to obtain the thermophysical properties of the DNN model over a wide range of temperatures and pressures. The simulation results were suggestive of the existence of a liquid-liquid critical point, and we used a two-state equation of state to estimate its location. These results provide computational evidence, completely from first principles, that is suggestive of the existence of the LLT in real water. We also discuss our ongoing work to apply free energy methods to definitely establish liquid-liquid coexistence using this model and solidify the computational evidence for this phenomenon.