[论文解读] Scalable molecular simulation of electrolyte solutions with quantum chemical accuracy

Junji Zhang, Joshua Pagotto+2 more|arXiv (Cornell University)|Oct 19, 2023

Machine Learning in Materials Science被引用 11

一句话总结

论文开发了一套基于 DC-r2SCAN 数据的全原子与 coarse-grained 神经网络势场框架，用以从第一性原理预测水性 LiCl 的性质，揭示 Li-Li 二聚体形成，并实现与实验一致的热力学与动力学输出。

ABSTRACT

Electrolyte solutions play critical role in a vast range of important applications, yet an accurate and scalable method of predicting their properties without fitting to experiment has remained out of reach, despite over a century of effort. Here, we combine state-of-the-art density functional theory and equivariant neural network potentials to demonstrate this capability, reproducing key structural, thermodynamic, and kinetic properties. We show that neural network potentials (NNPs) can be recursively trained on a subset of their own output to enable coarse-grained/continuum-solvent molecular simulations that can access much longer timescales than possible with all atom simulations. We observe the surprising formation of Li cation dimers along with identical anion-anion pairing of chloride and bromide anions. Finally, we reproduce simulate the crystal phase and infinite dilution pairing free energies despite being trained only on moderate concentration solutions. This approach should be scaled to build a greatly expanded database of electrolyte solution properties than currently exists.

研究动机与目标

• 需要准确、基于第一性原理的电解质热力学、动力学和结构预测，超越经验模型。
• 构建一个将 DC-DFT 数据与等变神经网络势结合的可扩展工作流，以实现对电解质的大规模模拟。
• 展示能够从第一性原理预测结构（RDFs）、热力学（活度系数）和动力学（扩散系数）等性质。
• 展示对不同浓度的泛化能力，并实现用于更快模拟的粗粒化。

提出的方法

• 使用 density-corrected DC-DFT (DC-r2SCAN) 生成来自短的 FPMD 轨迹的训练数据。
• 用 655 DC-r2SCAN 的数据训练两个等变神经网络势（NNP1、NNP2），包含 4 Li+, 4 Cl-, 和 80 H2O
• 在训练期间通过屏蔽库伦相互作用来引入长程静电耦合，并在 NNP-MD 模拟中将它们重新加入。
• 进行全原子 NNP-MD，然后建立粗粒度 NNP 学习平均力势以实现更快的模拟。
• 通过 RDFs、扩散系数和 Kirkwood–Buff 活度导数等与实验对比验证。

Figure 1 : Simulation overview Small, short FPMD simulations are run with CP2K. Energies/forces are recomputed with the DC-r 2 SCAN DFA. This data set is then used to train an equivariant NNP which enables much larger simulations with a larger cell. Forces and coordinates of the ions alone are outpu

实验结果

研究问题

• RQ1 一个小型、基于 DC-DFT 的训练集是否能使 NNPs 对电解质溶液具有准确、可迁移性？
• RQ2 等变 NNP 是否能再现 LiCl 水溶液在不同浓度下的结构、热力学和动力学性质？
• RQ3 以全原子数据训练的粗粒度 NNP 能否再现离子-离子和离子-水相关，并预测渗透/活度？
• RQ4 从第一性原理模拟中出现的对 LiCl 溶剂化的新见解（如 Li+ 阳离子二聚体）？

主要发现

• Li cation dimers are observed, with Li–Li separations around 2.68 Å in the first solvation layer.
• NNP 力场 RMSEs are below 10 meV/Å (NNP1: 8.9 meV/Å; NNP2: 9.7 meV/Å).
• All-atom NNP-MD on a 25.26 Å3 cell (512 H2O, 48 ions) yields RDFs and coordination numbers in good agreement with neutron diffraction data (Li–O ~1.97 Å; Li–O coordination ~4.2; Cl–O ~3.17 Å; coordination ~6.9).
• Diffusivities of Li+, Cl−, and H2O in good agreement with experiment; activity derivatives from Kirkwood–Buff theory align with experimental trends.
• A coarse-grained NNP reproduces all-atom RDFs and captures concentration-dependent screening; coarse-grained results converge with tens of CPU hours.
• The infinite-dilution PMF extracted from the coarse-grained model enables osmotic-coefficient estimates via modified Poisson–Boltzmann equations in good agreement with experiment

Figure 2 : Lithium dimer formation Lithium oxygen, and lithium lithium RDFs demonstrate the formation of lithium cation dimers that penetrate into the first hydration layer and an example snapshot from the MD simulation, showing the water molecules and nearest chloride ion surrounding the dimer.