发表机构
University of Cambridge; Max Planck Institute for Polymer Research; Max Planck Initiative for Liquid Research; Universitat Politècnica de Catalunya（剑桥大学; 马克斯·普朗克聚合物研究所; 马克斯·普朗克液体研究倡议; 加泰罗尼亚理工大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文通过对比DFT与RPA在石墨烯-水界面的表现，提出多观测评分框架，识别出六个可靠泛函，并揭示动力学和光谱性质更难重现，为固液界面模拟提供基准策略。
AI 中文摘要
固体界面处水的模拟在基本性质上（从接触角到摩擦）经常因所用模型（经验力场或密度泛函理论（DFT）交换关联（XC）近似）的不同而产生分歧。然而，此前缺乏精确的计算参考来判断哪种模型（如果有的话）是可靠的。在此，我们针对石墨烯-水界面，评估了跨越Jacob阶梯各梯级的XC近似与随机相位近似（RPA）的对比。利用基于DFT和RPA数据训练的机器学习势（MLPs），我们实现了表征界面结构、润湿性、摩擦和振动和频产生光谱所需的大量采样。我们发现XC近似与RPA以及彼此之间存在显著分歧。为理解这一点，我们引入了一个可迁移的多观测评分框架，将每个泛函与RPA的一致性提炼为单一度量。六个泛函被确定为研究石墨烯-水及其他色散主导界面的可靠起点，即B3LYP-D3(0)、revPBE-D3(0)、revPBE-D3(BJ)、B97M-rV、r$^{2}$SCAN和revPBE0-D3(0)。值得注意的是，这一排名并不遵循基于泛函家族或Jacob阶梯梯级的简单规则。跨泛函而言，动力学和光谱性质往往比结构性质更难用DFT重现，这对泛函选择有直接影响。我们的框架还为任何原子级模型（包括新一代基础MLPs）在固液界面上与高精度参考进行基准测试提供了一种通用策略。超越模拟层面，这项工作确定了哪些模型可被信赖，以将测量的接触角、摩擦和振动光谱与背后的分子结构联系起来。
英文摘要
Simulations of water at solid interfaces routinely disagree on basic properties, from contact angle to friction, depending on the model used, whether an empirical force field or a density functional theory (DFT) exchange-correlation (XC) approximation. Yet no accurate computational reference has been available to determine which, if any, is reliable. Here we assess XC approximations spanning the rungs of Jacob's Ladder against the random-phase approximation (RPA) at the graphene-water interface. Using machine-learned potentials (MLPs) trained on both DFT and RPA data, we achieve the extensive sampling required to characterize interfacial structure, wettability, friction, and vibrational sum-frequency generation spectra. We find that XC approximations disagree substantially with RPA and with each other. To make sense of this, we introduce a transferable, multi-observable scoring framework that distills each functional's agreement with RPA into a single measure. Six functionals emerge as reliable starting points for studies of graphene-water and other dispersion-dominated interfaces, namely B3LYP-D3(0), revPBE-D3(0), revPBE-D3(BJ), B97M-rV, r$^{2}$SCAN, and revPBE0-D3(0). Notably, this ranking follows no simple rule based on functional family or rung of Jacob's Ladder. What does hold across functionals is that dynamical and spectroscopic properties tend to be harder for DFT to reproduce than structural ones, with direct consequences for functional selection. Our framework also offers a general strategy for benchmarking any atomistic model, including the new generation of foundational MLPs, against high-accuracy references at solid--liquid interfaces. Beyond simulation, this work identifies which models can be trusted to connect measured contact angles, friction, and vibrational spectra to the molecular structure behind them.
