Comparing machine learning potentials for water: Kernel-based regression and Behler-Parrinello neural networks.

In this paper, we investigate the performance of different machine learning potentials (MLPs) in predicting key thermodynamic properties of water using RPBE + D3. Specifically, we scrutinize kernel-based regression and high-dimensional neural networks trained on a highly accurate dataset consisting of about 1500 structures, as well as a smaller dataset, about half the size, obtained using only on-the-fly learning. This study reveals that despite minor differences between the MLPs, their agreement on observables such as the diffusion constant and pair-correlation functions is excellent, especially for the large training dataset. Variations in the predicted density isobars, albeit somewhat larger, are also acceptable, particularly given the errors inherent to approximate density functional theory. Overall, this study emphasizes the relevance of the database over the fitting method. Finally, this study underscores the limitations of root mean square errors and the need for comprehensive testing, advocating the use of multiple MLPs for enhanced certainty, particularly when simulating complex thermodynamic properties that may not be fully captured by simpler tests.

Viscosity and self-diffusion of supercooled and stretched water from molecular dynamics simulations.

Among the numerous anomalies of water, the acceleration of dynamics under pressure is particularly puzzling. Whereas the diffusivity anomaly observed in experiments has been reproduced in several computer studies, the parallel viscosity anomaly has received less attention. Here we simulate viscosity and the self-diffusion coefficient of the TIP4P/2005 water model over a broad temperature and pressure range. We reproduce the experimental behavior and find additional anomalies at negative pressure. The anomalous effect of pressure on dynamic properties becomes more pronounced upon cooling, reaching two orders of magnitude for viscosity at 220 K. We analyze our results with a dynamic extension of a thermodynamic two-state model, an approach which has proved successful in describing experimental data. Water is regarded as a mixture of interconverting species with contrasting dynamic behaviors, one being strong (Arrhenius) and the other fragile (non-Arrhenius). The dynamic parameters of the two-state models are remarkably close between experiment and simulations. The larger pressure range accessible to simulations suggests a modification of the dynamic two-state model, which in turn also improves the agreement with experimental data. Furthermore, our simulations demonstrate the decoupling between viscosity η and self-diffusion coefficient D as a function of temperature T. The Stokes-Einstein relation, which predicts a constant Dη/T, is violated when T is lowered, in connection with the Widom line defined by an equal fraction of the two interconverting species. These results provide a unifying picture of thermodynamics and dynamics in water and call for experiments at negative pressure.

Brownian versus Newtonian devitrification of hard-sphere glasses.

In a recent molecular dynamics simulation work it has been shown that glasses composed of hard spheres crystallize via cooperative, stochastic particle displacements called avalanches [E. Sanz et al., Proc. Natl. Acad. Sci. USA 111, 75 (2014)PNASA60027-842410.1073/pnas.1308338110]. In this Rapid Communication we investigate if such a devitrification mechanism is also present when the dynamics is Brownian rather than Newtonian. The research is motivated in part by the fact that colloidal suspensions, an experimental realization of hard-sphere systems, undergo Brownian motion. We find that Brownian hard-sphere glasses do crystallize via avalanches with very similar characteristics to those found in the Newtonian case. We briefly discuss the implications of these findings for experiments on colloids.