HPC+AI驱动的第一性原理科学智能计算平台

doi:10.11871/jfdc.issn.2096-742X.2023.03.002

数据与计算发展前沿 ›› 2023, Vol. 5 ›› Issue (3): 13-28.

CSTR: 32002.14.jfdc.CN10-1649/TP.2023.03.002

doi: 10.11871/jfdc.issn.2096-742X.2023.03.002

• 专刊：“人工智能&大数据”科研范式变革专刊（下） • 上一篇下一篇

HPC+AI驱动的第一性原理科学智能计算平台

刘涛¹(),赵曈¹,谭光明^1,²,贾伟乐^1,^2,^*()

1.中国科学院计算技术研究所，处理器全国重点实验室，北京 100190
2.中国科学院大学，北京 100049

收稿日期:2023-05-04 出版日期:2023-06-20 发布日期:2023-06-21
通讯作者: *贾伟乐（E-mail: jiaweile@ict.ac.cn）
作者简介:刘涛，中国科学院计算技术研究所，高级工程师，主要研究方向为高性能计算、机器学习以及科学智能应用。
本文中主要承担工作为科学智能计算平台的整体架构设计。
LIU Tao, is a senior engineer in Institute of Computing Technology, Chinese Ac-ademy of Sciences. His research interests include High Perfor-mance Computing, Machine Learning, and AI for Science app-lications.
In this paper, he is mainly responsible for the overall fram-ework design of the AI for Science computing platform.
E-mail: liutao17@ict.ac.cn|贾伟乐，中国科学院计算技术研究所，副研究员，博士生导师，致力于智能科学计算（HPC+AI）研究，其参与研发的高性能深度学习分子动力学软件，比同类型软件效率提高4个数量级，被广泛应用（软件被用户应用在Nature、Science、PRL上发表文章）。先后获2020年ACM戈登贝尔奖，入选2020年两院院士评选的中国十大科技进展新闻。研究工作也入围了2022年ACM戈登贝尔奖，获得2022年中国超算最佳应用奖。
本文中主要承担工作为文献调研及平台概述。
JIA Weile, is an associate Professor and Ph.D. supervisor in Institute of Computing Technology, Chinese Academy of Sciences. His research focuses in the field of HPC+AI. He is among the key developers of the neural network based molecular dynamics software, and the software is used by users to publish papers in Nature, Science, PRL. He is the recipient of 2020 ACM Gordon Bell Award and was selected as one of the top ten scientific and technological progress news in China in 2020. One recent collaborative work was also selected among the 2022 ACM Gordon Bell Prize Finalists and won the 2022 China Supercomputing Best Application Award.
In this paper, he is mainly responsible for literature research and platform overview.
E-mail: jiaweile@ict.ac.cn
基金资助:
国家重点研发计划高性能计算专项（青年科学家项目）(2021YFB0300600);国家自然科学基金(92270206);国家自然科学基金(T2125013);国家自然科学基金(62032023);国家自然科学基金(61972377);中国科学院稳定支持青年科学家团队(YSBR-005);中国科学院网信专项“大数据+人工智能”科研范式变革应用示范(CAS-WX2021SF-0103)

An AI-for-Science Platform of Molecular Dynamics with Ab initio Accuracy

LIU Tao¹(),ZHAO Tong¹,TAN Guangming^1,²,JIA Weile^1,^2,^*()

1. State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China
2. University of Chinese Academy of Sciences, Beijing 100049, China

Received:2023-05-04 Online:2023-06-20 Published:2023-06-21

摘要/Abstract

摘要：

【目的】 科学智能（AI for Science）方法正在深刻地改变当前科学计算的格局。其融合了物理模型、人工智能与高性能计算，针对传统科学计算中的高维问题，通过数据拟合的方式实现成量级的增加高精度科学计算问题的时间和空间尺度，正在推动一场科研范式的变革。【方法】 本文针对第一性原理精度的分子动力学，提出一种HPC+AI驱动的科学智能计算平台，针对科学智能在工作流上带来的变化与挑战，从科学数据的生成与数据集制备、构型空间探索与训练样本标注、科学智能模型的高效训练及大规模高效推理等四个方面阐述构建科学智能计算平台的关键技术与流程。【结果】 本文所提出的计算平台在整合科学智能计算工作流的基础上，针对HPC+AI驱动的第一性原理精度分子动力学这一典型应用，提出了基于卡尔曼滤波的主动学习策略；改进了拟二阶AI模型训练方法，实现训练时间从天到分钟级的加速；利用五阶多项式AI模型压缩技术实现在同等硬件条件下模型推理的体系规模提高1个数量级，到解时间提高3-9倍。【结论】 通过上述工作的整合，形成一套可用于第一性原理精度分子动力学计算的科学智能计算平台。【局限与展望】 科学智能计算方法与工作流仍处于蓬勃发展阶段，在高精度数据、更通用AI模型和高效的计算方法等方面仍面临巨大的挑战，也将成为本文工作在未来的重要探索方向。

关键词: 科学智能, 第一性原理计算, 分子动力学, 主动学习, 卡尔曼滤波, 模型压缩

Abstract:

[Objective] AI for Science is changing the landscape of the traditional scientific computing by combining physical models, artificial intelligence, and high-performance computing to address challenging problems such as molecular dynamics with ab initio accuracy. This approach adapts neural networks to fitting high-dimensional functions, achieving orders of magnitude increases in the temporal and spatial scales, leading to a paradigm shift in scientific research. [Methods] This paper proposes an HPC+AI-driven computing platform for molecular dynamics with ab initio accuracy. Aiming at the changes and challenges brought by the workflow, the key technologies and processes for building an AI for Science computing platform are described from four aspects: generating scientific data and preparing datasets, exploring configuration space and labeling training samples, efficiently training AI for Science models, and performing large scale efficient inference (MD simulation). [Results] Based on the computational platform proposed in this paper and AI for Science computing workflows, this paper proposes an active learning strategy based on Kalman filtering for the typical application of HPC+AI-driven first-principles accuracy molecular dynamics. The training method for the quasi-second-order AI model is improved, achieving a training time acceleration from days to minutes. A fifth-order polynomial model compression technology increases the system scale by one order of magnitude for model inference and accelerates time-to-solution by 3-9 times. [Conclusions] All of the above work is combined to form an AI for Science computing platform for first-principles accuracy molecular dynamics calculations. [Limitations and Prospects] The AI for Science computing approach and workflows are still in a vigorous stage of development and facing significant challenges in high-precision data, more general AI models, and efficient computing methods. These challenges will also be important directions for future exploration in this work.

Key words: AI for Science, first-principles calculation, molecular dynamics, active learning, Kalman Filtering, Model compression

刘涛, 赵曈, 谭光明, 贾伟乐. HPC+AI驱动的第一性原理科学智能计算平台[J]. 数据与计算发展前沿, 2023, 5(3): 13-28.

LIU Tao, ZHAO Tong, TAN Guangming, JIA Weile. An AI-for-Science Platform of Molecular Dynamics with Ab initio Accuracy[J]. Frontiers of Data and Computing, 2023, 5(3): 13-28, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2023.03.002.

图/表 12

图1

图2

图3

图4

图5

图6

图7

图8

图9

表1

图10

图11

参考文献 48

[1]	HAMMOND B L, LESTER W A, REYNOLDS P J. Mon-te Carlo methods in ab initio quantum chemistry[M]. Sing-apore: World Scientific, 1994:1-41.
[2]	ERCOLESSI F, ADAMS J B. Interatomic Potentials fr-om First-Principles Calculations: The Force-Matching Method[J]. Europhysics Letters (EPL), 1994, 26(8): 583-588. doi: 10.1209/0295-5075/26/8/005
[3]	BLANK T B, BROWN S D, CALHOUN A W, et al. Ne-ural network models of potential energy surfaces[J]. The Journal of Chemical Physics, 1995, 103(10): 4129-4137. doi: 10.1063/1.469597
[4]	LU D, WANG H, CHEN M, et al. 86 PFLOPS Deep Pot-ential Molecular Dynamics simulation of 100 million atoms with ab initio accuracy[J]. Computer Physics Com-munications, 2021, 259: 107624.
[5]	ZHANG L, HAN J, WANG H, et al. Deep Potential Mo-lecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics[J]. Physical Review Letters, 2018, 120(14): 143001. doi: 10.1103/PhysRevLett.120.143001
[6]	WANG H, ZHANG L, HAN J, et al. DeePMD-kit: A de-ep learning package for many-body potential energy rep-resentation and molecular dynamics[J]. Computer Phy-sics Communications, 2018, 228: 178-184.
[7]	ZHANG D, BI H, DAI F Z, et al. DPA-1: Pretraining of Attention-based Deep Potential Model for Molecular Simu-lation[EB/OL]. arXiv, 2022[2023-05-03]. http://arxiv.org/abs/2208.08236.
[8]	ZHANG Y, WANG H, CHEN W, et al. DP-GEN: A conc-urrent learning platform for the generation of reliable deep learning based potential energy models[J]. Com-puter Physics Communications, 2020, 253: 107206.
[9]	HU S, ZHANG W, SHA Q, et al. RLEKF: An Optimizer for Deep Potential with Ab Initio Accuracy[EB/OL]. arXiv, 2022[2023-05-03]. http://arxiv.org/abs/2212.06989.
[10]	GUO Z, LU D, YAN Y, et al. Extending the limit of molecular dynamics with ab initio accuracy to 10 billion atoms[C]// Proceedings of the 27th ACM SIGPLAN Sym-posium on Principles and Practice of Parallel Programm-ing, Seoul Republic of Korea: ACM, 2022: 205-218.
[11]	BEHLER J, PARRINELLO M. Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces[J]. Physical Review Letters, 2007, 98 (14): 146401. doi: 10.1103/PhysRevLett.98.146401
[12]	SCHÜTT K, KINDERMANS P J, SAUCEDA FELIX H E, et al. SchNet: A continuous-filter convolutional ne-ural network for modeling quantum interactions[C]// Advances in Neural Information Processing Systems:Vol 30, Curran Associates, Inc., 2017:991-1001.
[13]	ZHANG L, HAN J, WANG H, et al. End-to-end Symm-etry Preserving Inter-atomic Potential Energy Model for Finite and Extended Systems[C]// Advances in Neural In-formation Processing Systems:Vol 31, Curran Associates, Inc., 2018:4436-4446.
[14]	GASTEIGER J, GIRI S, MARGRAF J T, et al. Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules[EB/OL]. arXiv, 2022[2023-05-03]. http://arxiv.org/abs/2011.14115.
[15]	MADHAWA K, ISHIGURO K, NAKAGO K, et al. Gra-phNVP: An Invertible Flow Model for Generating Mol-ecular Graphs[EB/OL]. arXiv, 2019[2023-05-03]. http://arxiv.org/abs/1905.11600.
[16]	XIE T, GROSSMAN J C. Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Pred-iction of Material Properties[J]. Physical Review Letters, 2018, 120(14): 145301. doi: 10.1103/PhysRevLett.120.145301
[17]	VASWANI A, SHAZEER N, PARMAR N, et al. Attenti-on is All you Need[C]// Advances in Neural Information Processing Systems:Vol 30, Curran Associates, Inc., 2017:5998-6008.
[18]	YING C, CAI T, LUO S, et al. Do Transformers Really Perform Badly for Graph Representation?[C]// Advances in Neural Information Processing Systems:Vol 34, Curr-an Associates, Inc., 2021: 28877-28888.
[19]	Atomly[DB/OL]. [2023-05-03]. https://www.atomly.net.
[20]	AIS Square[DB/OL]. [2023-05-03]. https://www.aissq-uare.com.
[21]	ScienceDB[DB/OL]. [2023-05-03]. https://www.scidb.cn.
[22]	FREDERIKSEN S L, JACOBSEN K W, BROWN K S, et al. Bayesian Ensemble Approach to Error Estimation of Interatomic Potentials[J]. Physical Review Letters, 2004, 93(16): 165501. doi: 10.1103/PhysRevLett.93.165501
[23]	BOTU V, RAMPRASAD R. Adaptive machine learning framework to accelerate ab initio molecular dynamics[J]. International Journal of Quantum Chemistry, 2015, 115(16): 1074-1083. doi: 10.1002/qua.24836
[24]	SZLACHTA W J, BARTÓK A P, CSÁNYI G. Accuracy and transferability of Gaussian approximation potential models for tungsten[J]. Physical Review B, 2014, 90(10): 104108. doi: 10.1103/PhysRevB.90.104108
[25]	GLIELMO A, SOLLICH P, DE VITA A. Accurate inter-atomic force fields via machine learning with covariant kernels[J]. Physical Review B, 2017, 95(21): 21-4302.
[26]	BARTÓK A P, PAYNE M C, KONDOR R, et al. Gau-ssian Approximation Potentials: The Accuracy of Quan-tum Mechanics, without the Electrons[J]. Physical Rev-iew Letters, 2010, 104(13): 136403.
[27]	LI Z, KERMODE J R, DE VITA A. Molecular Dynamics with On-the-Fly Machine Learning of Quantum-Mech-anical Forces[J]. Physical Review Letters, 2015, 114(9): 096405. doi: 10.1103/PhysRevLett.114.096405
[28]	PODRYABINKIN E V, SHAPEEV A V. Active learning of linearly parametrized interatomic potentials[J]. Com-putational Materials Science, 2017, 140: 171-180.
[29]	ZHANG L, LIN D Y, WANG H, et al. Active learning of uniformly accurate interatomic potentials for materials simulation[J]. Physical Review Materials, 2019, 3(2): 023804. doi: 10.1103/PhysRevMaterials.3.023804
[30]	ZENG J, ZHANG L, WANG H, et al. Exploring the Chemical Space of Linear Alkane Pyrolysis via Deep Potential GENerator[J]. Energy & Fuels, 2021, 35(1): 762-769. doi: 10.1021/acs.energyfuels.0c03211
[31]	ZHANG L. Machine Learning for Multi-Scale Molecular Modeling:Theories, Algorithms, and Applications[D]. USA: Princeton University, 2020.
[32]	ZENG J, CAO L, XU M, et al. Neural Network Based in Silico Simulation of Combustion Reactions[J]. Nature Communications, 2020, 11(1): 5713. doi: 10.1038/s41467-020-19497-z
[33]	WEN T, WANG R, ZHU L, et al. Specialising neural network potentials for accurate properties and appli-cation to the mechanical response of titanium[J]. npj Comp-utational Materials, 2021, 7(1): 206.
[34]	WANG Y, ZHANG L, XU B, et al. A generalizable machine learning potential of Ag-Au nanoalloys and its application to surface reconstruction, segregation and diffusion[J]. Modelling and Simulation in Materials Science and Engineering, 2022, 30(2): 025003. doi: 10.1088/1361-651X/ac4002
[35]	KINGMA D P, BA J. Adam: A Method for Stochastic Optimization[C/OL]// 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. arXiv, 2015[2023-05-03]. http://arxiv.org/abs/1412.6980.
[36]	RUDER S. An overview of gradient descent optimization algorithms[EB/OL]. arXiv, 2017[2023-05-03]. http://arxiv.org/abs/1609.04747.
[37]	JIA W, FU J, CAO Z, et al. Fast plane wave density func-tional theory molecular dynamics calculations on multi-GPU machines[J]. Journal of Computational Physics, 2013, 251: 102-115.
[38]	JIA W, CAO Z, WANG L, et al. The analysis of a plane wave pseudopotential density functional theory code on a GPU machine[J]. Computer Physics Communications, 2013, 184(1): 9-18. doi: 10.1016/j.cpc.2012.08.002
[39]	HU W, LIN L, YANG C. DGDFT: A massively parallel method for large scale density functional theory calcu-lations[J]. The Journal of Chemical Physics, 2015, 143 (12): 124110. doi: 10.1063/1.4931732
[40]	LIU W, HONG G, DAI D, et al. The Beijing fourcomp-onent density functional program package (BDF) and its application to EuO, EuS, YbO and YbS[J]. Theoretical Ch-emistry Accounts: Theory, Computation, and Modeling (Theoretica Chimica Acta), 1997, 96(2): 75-83.
[41]	HU W, LIN L, YANG C. Interpolative Separable Density Fitting Decomposition for Accelerating Hybrid Density Functional Calculations with Applications to Defects in Silicon[J]. Journal of Chemical Theory and Computation, 2017, 13(11): 5420-5431. doi: 10.1021/acs.jctc.7b00807 pmid: 28960982
[42]	HU W, LIN L, BANERJEE A S, et al. Adaptively Com-pressed Exchange Operator for Large-Scale Hybrid Den-sity Functional Calculations with Applications to the Ad-sorption of Water on Silicene[J]. Journal of Chemical Theory and Computation, 2017, 13(3): 1188-1198. doi: 10.1021/acs.jctc.6b01184
[43]	KALMAN R E. A New Approach to Linear Filtering and Prediction Problems[J]. Journal of Basic Engineering, 1960, 82(1): 35-45. doi: 10.1115/1.3662552
[44]	SMITH G L, SCHMIDT S F, MCGEE L A. Application of statistical filter theory to the optimal estimation of position and velocity on board a circumlunar vehicle[M]. Washington, D. C: National Aeronautics and Space Administration, 1962:1-26.
[45]	SINGHAL S, WU L. Training Multilayer Perceptrons with the Extended Kalman Algorithm[C]// Advances in Neural Information Processing Systems:Vol 1, Morgan-Kaufmann, 1988:133-140.
[46]	BEHLER J. Atom-centered symmetry functions for constructing high-dimensional neural network pote-ntials[J]. The Journal of Chemical Physics, 2011, 134(7): 074106. doi: 10.1063/1.3553717
[47]	MURTUZA S, CHORIAN S F. Node decoupled extended Kalman filter based learning algorithm for neural net-works[C]// Proceedings of 1994 9th IEEE International Symposium on Intelligent Control, 1994: 364-369.
[48]	HEIMES F. Extended Kalman filter neural network training: experimental results and algorithm impro-vements[C]// SMC’98 Conference Proceedings, 1998 IEEE International Conference on Systems, Man, and Cybernetics (Cat. No.98CH36218): Vol 2, San Diego, CA, USA: IEEE, 1998: 1639-1644

物理量	$\mathrm{V}_{0}\left(Å^{3} / \text { atom }\right)$	E_Vf(eV)	a₀(Å)
DFT	41.6892	0.0002	3.4674
ALKPU	41.2217	0.0016	3.4544

HPC+AI驱动的第一性原理科学智能计算平台

An AI-for-Science Platform of Molecular Dynamics with Ab initio Accuracy

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 48

相关文章 0

编辑推荐

Metrics

本文评价