计算机科学 ›› 2025, Vol. 52 ›› Issue (5): 50-57.doi: 10.11896/jsjkx.241100176
李恩吉, 胡思宇, 谭光明, 贾伟乐
LI Enji, HU Siyu, TAN Guangming, JIA Weile
摘要： 分子动力学模拟是一种广泛应用于多个学科(如材料科学、计算化学等)的关键研究方法。近年来,随着计算能力的提升、神经网络模型的发展以及第一性原理数据的增加,神经网络力场模型已经展现出高精度的预测能力。目前存在多种神经网络力场模型的训练算法,而神经网络力场模型处于一个快速迭代的阶段,当前仍然缺乏神经网络力场模型及与之适配的优化器的指导建议。选取3种有代表性的神经网络力场模型和目前3种用于神经网络力场模型上的优化算法,在4个真实数据集上进行测试和评估,分析影响其收敛性的原因。设计实验对其进行全方位的评估,包括模型参数量对优化器的影响,神经网络宽度对收敛性的影响,以及模型训练时间与优化器的关联等。文中工作可以针对神经网络力场模型,给出优化器算法的建议。
中图分类号:
| [1]ENGEL E.Density functional theory[M].Springer,2011. [2]ZEPEDA-NÚÑEZ L,CHEN Y,ZHANG J,et al.Deep density:circumventing thekohn-sham equations via symmetry preserving neural networks[J].arXiv:1912.00775,2019. [3]HAFNER J.Ab-initio simulations of materials usingvasp:Densi-tyfunctional theory and beyond[J].Journal of Computational Chemistry,2008,29(13):2044-2078. [4]GIANNOZZI P,ANDREUSSI O,BRUMME T,et al.Advanced capabilities for materials modelling with quantum espresso[J].Journal of Physics:Condensed Matter,2017,29(46):465901. [5]JIA W,FU J,CAO Z,et al.Fast plane wave density functional theory molecular dynamics calculations on multi-gpu machines[J].Journal of Computational Physics,2013,251:102-115. [6]KOURA K,MATSUMOTO H.Variable soft sphere molecular model for inverse-power-law orlennard-jones potential[J].Phy-sics of Fluids A:Fluid Dynamics,1991,3(10):2459-2465. [7]FOILES S,BASKES M,DAW M S.Embedded-atom-methodfunctions for thefcc metals cu,ag,au,ni,pd,pt,and their alloys[J].Physical Review B,1986,33(12):7983. [8]SENFTLE T P,HONG S,ISLAM M M,et al.The reaxff reactive forcefield:development,applications and future directions[J].Computational Materials,2016,2(1):1-14. [9]NI B,LEE K H,SINNOTT S B.A reactive empirical bond order(REBO) potential for hydrocarbon-oxygen interactions[J].Journal of Physics:Condensed Matter,2004,16(41):7261. [10]NGUYEN T D.Gpu-accelerated tersoff potentials for massively parallel molecular dynamics simulations[J].Computer Physics Communications,2017,212:113-122. [11]HUANG Y,KANG J,GODDARD III W A,et al.Density functional theory based neural network force fields from energy decompositions[J].Physical Review B,2019,99(6):064103. [12]THOMPSON A,SWILER L,TROTT C,et al.Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials[J].Journal of Computational Phy-sics,2015,285:316-330. [13]LEE K,YOO D,JEONG W,et al.Simple-nn:An efficient package for training and executing neural-network interatomic potentials[J].Computer Physics Communications,2019,242:95-103. [14]BEHLER J.Representing potential energy surfaces by high-dimensional neural network potentials[J].Journal of Physics:Condensed Matter,2014,26(18):183001. [15]WANG H,ZHANG L,HAN J,et al.Deepmd-kit:A deep learning package for many-body potential energy representation and molecular dynamics[J].Computer Physics Communications,2018,228:178-184. [16]FAN Z,WANG Y,YING P,et al.GPUMD:A package for constructing accurate machine-learned potentials and performing highly efficient atomistic simulations[J].The Journal of Chemical Physics,2022,157(11):114801. [17]DAI H,MACBETH C.Effects of learning parameters on lear-ning procedure and performance of abpnn[J].Neural networks,1997,10(8):1505-1521. [18]SMITH J S,ISAYEV O,ROITBERG A E.Ani-1:an extensible neural network potential withdft accuracy at force field computational cost[J].Chemical science,2017,8(4):3192-3203. [19]SCHÜTT K,UNKE O,GASTEGGER M.Equivariant message passing for the prediction of tensorial properties and molecularspectra[C]//International Conference on Machine Learning.PMLR,2021:9377-9388. [20]BATZNER S,MUSAELIAN A,SUN L,et al.E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials[J].Nature Communications,2022,13(1):2453. [21]HAGHIGHATLARI M,LI J,GUAN X,et al.Newtonnet:anewtonian message passing network for deep learning of inte-ratomic potentials and forces[J].Digital Discovery,2022,1(3):333-343. [22]JIA W,WANG H,CHEN M,et al.Pushing the limit of molecular dynamics with ab initio accuracy to 100 million atoms with machine learning[C]//SC20:International Conference for High Performance Computing,Networking,Storage and Analysis.IEEE,2020:1-14. [23]KINGMA D,BA J.Adam:A method for stochastic optimization[J].arXiv.1412.6980,2014. [24]HU S,ZHANG W,SHA Q,et al.Rlekf:An optimizer for deep potential with ab initio accuracy[C]//Proceedings of the AAAI Conference on Artificial Intelligence:volume 37.2023:7910-7918. [25]HU S,ZHAO T,SHA Q,et al.Training onedeepmd model in minutes:a step towards online learning[C]//PPoPP '24:Proceedings of the 29th ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming.New York,NY,USA:Association for Computing Machinery,2024:257-269. [26]SCHAUL T,GLASMACHERS T,SCHMIDHUBER J.Highdimensions and heavy tails for natural evolution strategies[C]//Proceedings of the 13th Annual Conference on Genetic and Evolutionary Computation(GECCO'11).New York,NY,USA:Association for Computing Machinery,2011:845-852. [27]LAMBORA A,GUPTA K,CHOPRA K.Genetic algorithm- a literature review[C]//2019 International Conference on Machine Learning,Big Data,Cloud and Parallel Computing(COMITCon).2019:380-384. [28]PIEPER A,KREUTZER M,ALVERMANN A,et al.High-performance implementation ofchebyshev filter diagonalization for interior eigenvalue computations[J].Journal of Computational Physics,2016,325:226-243. [29]ZAVERKIN V,HOLZMÜLLER D,STEINWART I,et al.Fast and sample-efficient interatomic neural network potentials for molecules and materials based on gaussian moments[J].Journal of Chemical Theory and Computation,2021,17(10):6658-6670. [30]NISHIJIMA T.Universal approximation theorem for neuralnetworks[J].arXiv:2102.10993,2021. |
|
||
首页