摘要
对于研究高分子溶液的动力学行为来说,计算分子动力学(MD)采用的隐式溶剂模型中的流体动力学相互作用(HI)是极其重要的.对于计算HI来说,Cholesky方法是最常用到的分解扩散张量的方法.然而,当体系包含粒子的数目较大时,该方法所需的计算量往往大到难以承受.因此,开发计算HI的高效率技术非常必要.近来,机器学习,特别是深度神经网络(DNN),已经在计算多体相互作用方面得到了显著提升.该技术可以显著减少MD模拟所需的计算时间,因此受到了越来越多的关注.在本项研究中,通过应用机器学习方法,我们发展了一种用于加速计算HI的DNN模型,其相互作用训练集通过针对高分子和溶剂的MD模拟来获取.该DNN模型则通过Deepmd-kit软件包来训练.我们选取表现最好的模型,用以取代对扩散张量的分解.计算得到的高分子链的回转半径和扩散系数用来验证模型的精确度.测试结果表明,相较于显式溶剂模型,我们的DNN模型在计算单链的回转半径和扩散系数方面具有优秀的计算精度.此外,该方法还具有两个显著优势.首先,设计训练集只需要特定链长的一条链,其结果却可以应用于其他链长却不影响计算精度.这就大大降低了训练需求.其次,运行时间与链长线性相关,这就使得计算HI的计算成本也大幅缩减.总之,我们开发的方法提供了一种针对长期悬而未决的计算高分子溶液中的HI的成本过于高昂的解决思路.
Abstract
Calculating hydrodynamic interactions(HI)in the implicit solvent model of molecular dynamics(MD)is crucial to capture physically realistic dynamics in polymer solutions.The Cholesky method is the most commonly used to decompose the diffusion tensor when calculating the HI.However,the computa-tional cost of this method is often overwhelming when the particle system has a large number of particles.Therefore,it has become essential that efficient techniques are proposed to accelerate the calculation of HI.Recently,machine learning,especially deep neural networks(DNN),has rapidly progressed in com-puting multi-body interactions.This technology can significantly reduce the time required for MD simula-tions and has gained increasing attention in recent years.In this work,using the machine learning method,we develop a DNN model for accelerated calculation of HI.We obtain the interaction dataset by performing MD simulations on polymer and solvents.The DNN model is trained by the Deepmd-kit package.We se-lect the best-performing model and replace the decomposition of the diffusion tensor.The radius of gyra-tion and the diffusion coefficient of polymer chains are calculated to verify the model's accuracy.The test results show that our DNN model has excellent computational accuracy in calculating the radius of gyration and diffusion coefficients of single chains in comparison with explicit solvent(ES).Moreover,the method has two significant advantages.Firstly,only one chain length is used to design the training set,and the same applies to other chain lengths with less impact on the computational accuracy,which can significantly reduce the need for training.Second,the running time is linearly related to the chain length,which can significantly reduce the cost required to compute HI.As a result,our method provides a solution to the long-standing problem surrounding the computational costliness of HI in polymer solutions.
基金项目
National Natural Science Foundation of China(22373025)
NETL
NSTL
万方数据