基于神经网络的分子体系势能面构造
详细信息 查看官网全文
- 英文论文题名：Construction of Potential Energy Surfaces using Neural Networks
- 论文作者：陈俊 ; 张东辉
- 英文论文作者：Jun Chen ; D.H.Zhang ; State Key Laboratory of Molecular Reaction Dynamics ; Dalian Institute of Chemical Physics ; Chinese Academy of Sciences
- 年：2016
- 作者机构：中国科学院大连化学物理研究所分子反应动力学国家重点实验室;
- 论文关键词：反应动力学 ; 势能面 ; 神经网络 ; 量子动力学
- 会议召开时间：2016-07-01
- 会议录名称：中国化学会第30届学术年会摘要集-第十九分会：化学中的量子与经典动力学
- 语种：中文
- 分类号：O643.1
- 学会代码：ZGHY
- 会议名称：中国化学会第30届学术年会-第十九分会 ; 化学中的量子与经典动力学
- 会议地点：中国辽宁大连
- 主办单位：中国化学会
- 学会名称：中国化学会
- 页数：1
- 文件大小：139k
- 原文格式：D
- 会议级别：全国
摘要
精确的势能面是化学反应动力学理论研究的基础。分子体系势能面的构造一般从大量分子构型下的从头算出发,使用插值或拟合的手段生成解析形式的函数。而在高维空间中,如何选择合适的构型用于从头算以及后续的拟合,一直是一个难题。我们基于神经网络方法,结合准经典动力学计算以及现有的一些研究成果,建立了一套流程化的分子构型选择方案,能有效地选择出重要的构型,解决了在高维空间中构型选择的难题。利用Levenberg-Marquardt算法,并采用一系列技巧,可以使用神经网络函数将势能面拟合到相当高的精度,最终得到可靠的分子反应势能面~([1])。利用这种方法,成功构造了F/Cl+H_2等三原子体系~([2,3]),OH+H_2/CO等四原子体系~([4,5]),以及H+CH_4等六原子体系~([6])的全维反应势能面。结合量子反应动力学计算,这些势能面相对于从头算数目以及神经网络拟合误差都已经收敛,并且比现有的势能面精确,为精确研究这些反应的动力学行为奠定了基础。
In this presentation, we introduce a schematic method for fast construction of reliable potential energy surfaces, including high efficient method for sampling throughout the whole configuration space, and accurately fitting to the ab initio values by using neural networks. Based on this scheme, a number of high accurate PESs have been constructed, including 3-atoms systems(3D) such as F/Cl + H_2, 4-atoms systems(6D) such as OH + H_2/CO, and 6-atoms systems(12D) such as H + CH_5. Combined with quantum dynamics calculations, these PESs are proved to be fully converged respect to the number of configuration geometries as well as neural network fitting errors, and provide foundations for further reaction dynamics studies.
In this presentation, we introduce a schematic method for fast construction of reliable potential energy surfaces, including high efficient method for sampling throughout the whole configuration space, and accurately fitting to the ab initio values by using neural networks. Based on this scheme, a number of high accurate PESs have been constructed, including 3-atoms systems(3D) such as F/Cl + H_2, 4-atoms systems(6D) such as OH + H_2/CO, and 6-atoms systems(12D) such as H + CH_5. Combined with quantum dynamics calculations, these PESs are proved to be fully converged respect to the number of configuration geometries as well as neural network fitting errors, and provide foundations for further reaction dynamics studies.
引文
[1]陈俊,张东辉.中国科学:化学,2015,45(12):1241
[2]Wang,T.;Chen,J.;Yang,T.G.;Xiao,C.L.;Sun,Z.G.;Huang,L.;Dai,D.X.;Yang,X.M.;Zhang,D.H.Science,2013,342:1499.
[3]Yang,T.G.;Chen,J.;Huang,L.;Wang,T.;Xiao,C.L.;Sun,Z.G.;Dai,D.X.;Yang,X.M.;Zhang,D.H.Science,2015,347:60.
[4]Chen,J;Xu,X.;Xu,X.;Zhang,D.H.J.Chem.Phys.2013,138:154301.
[5]Chen,J;Xu,X.;Xu,X.;Zhang,D.H.J.Chem.Phys.2013,138:221104.
[6]Xu,X.;Chen,J.;Zhang,D.H.Chin.J.Chem.Phys.2014,27:373.
[2]Wang,T.;Chen,J.;Yang,T.G.;Xiao,C.L.;Sun,Z.G.;Huang,L.;Dai,D.X.;Yang,X.M.;Zhang,D.H.Science,2013,342:1499.
[3]Yang,T.G.;Chen,J.;Huang,L.;Wang,T.;Xiao,C.L.;Sun,Z.G.;Dai,D.X.;Yang,X.M.;Zhang,D.H.Science,2015,347:60.
[4]Chen,J;Xu,X.;Xu,X.;Zhang,D.H.J.Chem.Phys.2013,138:154301.
[5]Chen,J;Xu,X.;Xu,X.;Zhang,D.H.J.Chem.Phys.2013,138:221104.
[6]Xu,X.;Chen,J.;Zhang,D.H.Chin.J.Chem.Phys.2014,27:373.