NEP-MB-pol: A unified machine-learned framework for fast and accurate prediction of water's thermodynamic and transport properties
Ke Xu,Ting Liang,Nan Xu,Penghua Ying,Shunda Chen,Ning Wei,Jianbin Xu,Zheyong Fan
2024-11-20
Abstract:Water's unique hydrogen-bonding network and anomalous properties pose significant challenges for accurately modeling its structural, thermodynamic, and transport behavior across varied conditions. Although machine-learned potentials have advanced the prediction of individual properties, a unified computational framework capable of simultaneously capturing water's complex and subtle properties with high accuracy has remained elusive. Here, we address this challenge by introducing NEP-MB-pol, a highly accurate and efficient neuroevolution potential (NEP) trained on extensive many-body polarization (MB-pol) reference data approaching coupled-cluster-level accuracy, combined with path-integral molecular dynamics and quantum-correction techniques to incorporate nuclear quantum effects. This NEP-MB-pol framework reproduces experimentally measured structural, thermodynamic, and transport properties of water across a broad temperature range, achieving simultaneous, fast, and accurate prediction of self-diffusion coefficient, viscosity, and thermal conductivity. Our approach provides a unified and robust tool for exploring thermodynamic and transport properties of water under diverse conditions, with significant potential for broader applications across research fields.
Chemical Physics,Materials Science,Soft Condensed Matter,Statistical Mechanics