深度勢能分子動力學在甲烷中的應用

Abstract

本研究主要探討四面體甲烷二聚體分子之量子化學計算並將第一性原理分子動力學(Ab Initio Molecular Dynamics, AIMD)結合DeePMD-kit深度學習建構出勢能面模型並進行後續分子動力學模擬及其熱力學性質分析。 首先，在量子計算方面以密度泛函理論(Density Functional Theory, DFT)方法做計算，並針對甲烷分子間作用力類型篩選出了八種描述甲烷較理想的DFT交換相關泛函(Exchange-Correlation functional, XC functional)。而為了更貼近實際計算的情況以及瞭解有無使用DFT色散校正(Dispersion correction)的差別性與重要性，本此研究使用了AIMD贋勢基底(Pseudopotential basis set)在經由量子力學優化後的甲烷D3d二聚體構型上計算能量曲線，計算時中心原子距離從1.4 Å至18.5 Å共取14個點，並與CCSD(T)/CBS結果相比較，可較直觀地瞭解交換相關泛函和色散校正在贋勢基底選用上的差異。 而在分子動力學模擬方面利用CP2K軟體計算出各DFT交換相關泛函產生之AIMD軌跡數據作為DeePMD-kit深度學習的訓練資料。隨後透過深度學習建構出勢能面模型，並將此模型提取到LAMMPS軟體進行後續的分子動力學模擬，藉此得到了甲烷之徑向分佈函數(Radial Distribution Function, RDF)、均方位移(Mean Square Displacement, MSD)、速度自相關函數(Velocity AutoCorrelation Function, VACF)、擴散係數(Diffusion Coefficient, DC)以及剪力黏滯係數(Shear Viscosity Coefficient, SVC)等相關熱力學性質。最後，將結果與實驗值和經驗力場相互比較。 本研究結果表明，以AIMD和深度學習相結合後產生之力場模型，同時具有DFT之高精度，也具有基於經驗勢之高效率，為分子動力學模擬在精度與效率兩難之間，提供了一個全新的見解。

This study focuses on the quantum chemical calculation of tetrahedral methane dimer molecules and combines Ab Initio Molecular Dynamics (AIMD) with DeePMD-kit deep learning to construct a potential energy surface model for subsequent molecular dynamics simulations and thermal properties analysis. Firstly, the DFT method is used in the quantum computation, and eight DFT Exchange-Correlation functionals are selected for the type of intermolecular forces to describe methane. In order to be closer to the actual calculation situation and to understand the difference and importance of using DFT dispersion correction or not, the AIMD Pseudopotential basis set was used to calculate the energy curves on the quantum mechanics optimized methane D3d dimer configuration, and the distance between the central atoms was taken from 1.4 Å to 18.5 Å at a total of 14 points. Comparing with the results of the CCSD(T)/CBS, the exchange-correlation functional, dispersion correction set can be understood more intuitively. In the molecular dynamics simulation, the AIMD trajectory data generated by DFT was calculated by CP2K software as the training data for DeePMD-kit deep learning. Then the force field model was constructed by deep learning, and this force field model was extracted to LAMMPS software for subsequent molecular dynamics simulation, so as to obtain the relevant thermodynamic properties of methane, such as RDF, MSD, VACF, DC, SVC, etc. Finally, the results are compared with experimental values and empirical force fields. The results of this study show that the force field model generated by combining AIMD and deep learning has both the high accuracy of DFT and the high efficiency based on the empirical potentials. All in all, deep learning has brought new insights to addressing the accuracy versus efficiency dilemma in molecular simulations.