机器学习加速氧化还原电位和酸度常数计算

doi:10.13208/j.electrochem.2307181

电化学（中英文） ›› 2024, Vol. 30 ›› Issue (2): 2307181. doi: 10.13208/j.electrochem.2307181

所属专题： “教程”专题文章

• 教程 • 上一篇

机器学习加速氧化还原电位和酸度常数计算

王锋^a, 程俊^a^,^b^,^c^,^*()

^a厦门大学化学化工学院，固体表面物理化学国家重点实验室，能源材料化学协同创新中心，福建厦门 361005
^b福建省能源材料科学与技术创新实验室（IKKEM），福建厦门 361005
^c厦门大学人工智能研究院，福建厦门 361005

收稿日期:2023-07-18 修回日期:2023-10-04 接受日期:2023-11-06 发布日期:2023-11-06 出版日期:2024-02-28
通讯作者: *程俊，Tel: (86-592)2181570；E-mail: juncheng@xmu.edu.cn
基金资助:
国家自然科学基金项目(22225302);国家自然科学基金项目(21991151);国家自然科学基金项目(21991150);国家自然科学基金项目(22021001);国家自然科学基金项目(92161113);中央高校基本科研业务费专项资金(20720220009);AI4EC联合实验室基金(RD2023100101);AI4EC联合实验室基金(RD2022070501)

Automated Workflow for Redox Potentials and Acidity Constants Calculations from Machine Learning Molecular Dynamics

Feng Wang^a, Jun Cheng^a^,^b^,^c^,^*()

^aState Key Laboratory of Physical Chemistry of Solid Surfaces, Collaborative Innovation Center of Chemistry for Energy Materials (iChEM), College of Chemistry and Chemical Engineering, Xiamen University, Xiamen 361005, China
^bIInnovation Laboratory for Sciences and Technologies of Energy Materials of Fujian Province (IKKEM), Xiamen 361005, China
^cInstitute of Artificial Intelligence, Xiamen University, Xiamen 361005, China

Received:2023-07-18 Revised:2023-10-04 Accepted:2023-11-06 Online:2023-11-06 Published:2024-02-28

摘要/Abstract

摘要：

氧化还原电位和酸度常数作为重要的物理化学性质被应用于分析能源材料重要指标值。为了实现能源材料的计算设计，发展计算电化学的方法，在复杂电化学环境下计算这些性质至关重要。近年来，利用计算电化学方法计算氧化还原电位和酸度常数已经受到了广泛的关注。然而，常用的计算方法如基于隐式溶剂化模型的小分子自由能计算，对于复杂溶剂化环境的处理非常有限。因此，基于第一性原理分子动力学（AIMD）的自由能计算被引入来描述复杂溶剂化环境中的溶质-溶剂相互作用。同时，基于AIMD的自由能计算方法已经被证实可以准确预测这些物理化学性质。然而，由于AIMD计算效率低且计算资源需求大，需要引入机器学习分子动力学（MLMD）加速计算。MLMD通过机器学习方法，构建模拟体系结构到第一性原理计算结果的一对一映射，可以在低成本下实现长时间尺度的AIMD。对于氧化还原电位和酸度常数计算，如何构建训练机器学习势函数模型所需的数据集至关重要。本文介绍了如何通过自动化工作流实现自由能计算势函数的自动化构建，通过机器学习分子动力学计算自由能并转化为对应的物理化学性质。

关键词: 机器学习分子动力学, 自动化工作流, 复杂体系, 自由能计算

Abstract:

Redox potentials and acidity constants are key properties for evaluating the performance of energy materials. To achieve computational design of new generation of energy materials with higher performances, computing redox potentials and acidity constants with computational chemistry have attracted lots of attention. However, many works are done by using implicit solvation models, which is difficult to be applied to complex solvation environments due to hard parameterization. Recently, ab initio molecular dynamics (AIMD) has been applied to investigate real electrolytes with complex solvation. Furthermore, AIMD based free energy calculation methods have been established to calculate these physical chemical properties accurately. However, due to the low efficiency of ab initio calculations and the high computational costs, AIMD based free energy calculations are limited to systems with less than 1000 atoms. To solve the dilemma, machine learning molecular dynamics (MLMD) is introduced to accelerate the calculations. By using machine learning method to construct one-to-one mapping from structures to computed potential energies and atomic forces, molecular dynamics can be carried out with much low costs under ab initio accuracy. In order to achieve the MLMD based free energy calculation, a new scheme for machine learning potential (MLP) should be introduced to collect training datasets. By combining the free energy perturbation sampling method and concurrent learning scheme, the training datasets can be collected along the reaction’s pathway (insertion of an electron or a proton) with high efficiency and the free energy calculations based on MLMD show good accuracy in comparison with AIMD simulation. This paper describes how to constructing machine learning potential for free energy calculation through the automated workflow, and how to use MLMD to compute accurate free energy differences and corresponding physical chemical properties.

Key words: Machine learning molecular dynamics, Automated workflow, Complex systems, Free energy calculation

王锋, 程俊. 机器学习加速氧化还原电位和酸度常数计算[J]. 电化学（中英文）, 2024, 30(2): 2307181.

Feng Wang, Jun Cheng. Automated Workflow for Redox Potentials and Acidity Constants Calculations from Machine Learning Molecular Dynamics[J]. Journal of Electrochemistry, 2024, 30(2): 2307181.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://electrochem.xmu.edu.cn/CN/10.13208/j.electrochem.2307181

https://electrochem.xmu.edu.cn/CN/Y2024/V30/I2/2307181

图/表 9

图1

图2

表1

通过自动化工作流获得的数据集大小和势函数对训练集的预测误差[31]

	$N i n i, N f i n$	$R M S E E$ $(m e V / a t o m)$	$R M S E F$ $(m e V / A ̊)$
$H C l → C l -$	800, 800	0.553, 0.615	42.8, 44.2
$H 3 O + → H 2 O$	6016, 5234	0.653, 0.809	47.5, 50.8
$H 2 S → H S -$	835, 835	0.580, 0.700	44.0, 45.7
$H 2 O → O H -$	11048, 10648	0.556, 0.579	44.0, 43.4
$H S - → S 2 -$	850, 827	0.544, 0.551	43.4, 45.5
$C l - → C l ∙$	875, 873	0.532, 0.586	44.7, 48.1
$O H - → O H ∙$	1996, 1996	0.797, 0.706	51.6, 49.5
$H S - → H S ∙$	887, 887	0.722, 0.604	47.7, 47.8
$O 2 - → O 2$	5499, 5499	0.819, 0.753	53.5, 52.1
$C O 2 - → C O 2$	5791, 5791	0.733, 0.603	57.1, 47.7

表1

图3

表2

通过基于MLMD模拟得到的酸度常数[31]

	AIMD	MLMD	MLMD*	exp.
HCl	-7.0	-8.2 (-8.8)	-6.7 (-7.3)	-7.0
$H 3 O +$	-3.2	-3.2 (-3.2)	-1.7 (-1.7)	-1.7
$H 2 S$	7.8	5.2 (5.0)	6.7 (6.5)	7.0
$H 2 O$	13.5	11.9 (11.6)	13.4 (13.1)	15.6
$H 2 S$	16.2	14.2 (14.5)	15.7 (16.0)	17.0

表2

图4

表3

通过基于MLMD模拟得到的氧化还原电位。MLMD计算使用的 Δ d p A H 3 O +为MLMD计算出的15.512 eV[31]。

	AIMD	MLMD (3P)	MLMD	exp.
$C l ∙ / C l -$	1.5	1.58	1.84	2.41
$O H ∙ / O H -$	1.3	1.49 (1.51)	1.68 (1.68)	1.90
$H S ∙ / H S -$	0.5	0.69	0.76	1.08
$O 2 / O 2 -$	-0.5	-0.35	-0.22	-0.16
$C O 2 / C O 2 -$	-2.07	-1.98	-2.06	-1.90

表3

图5

图6

机器学习加速氧化还原电位和酸度常数计算

Automated Workflow for Redox Potentials and Acidity Constants Calculations from Machine Learning Molecular Dynamics

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 9

参考文献

相关文章 1

Metrics