Gangopadhyay Ananya, Winberg Simon, Naidoo Kevin J
Scientific Computing Research Unit and Department of Chemistry, University of Cape Town, Cape Town, South Africa.
Scientific Computing Research Unit and Department of Electrical Engineering, University of Cape Town, Cape Town, South Africa.
J Comput Chem. 2021 Apr 15;42(10):666-675. doi: 10.1002/jcc.26487. Epub 2021 Feb 6.
A high-speed numerical potential delivering computational performance comparable with complex coarse-grained analytic potentials makes available models that have greater degrees of physical and chemical accuracy. This opens the possibility of increased accuracy in classical molecular dynamics simulations of anisotropic systems. In this work, we report the development of a high-speed lookup table (LUT) of four-dimensional gridded data, that uses cubic B-spline interpolations to derive off grid values and their associated partial derivatives that are located between the known grid data points. The accuracy of the coarse-grained numerical potential using a LUT from uniaxial Gay-Berne (GB) potential produced array of values is within a 3% and a 5% margin of error respectively for the interpolation of the uniaxial GB potential and its partial derivatives. The numerical potential model and partial derivatives speedup is made competitive with the analytical potential by exploiting graphics processing units on board functionality. The capability of the numerical potential is demonstrated by comparing minimizations of a box of 500 naphthalene molecules. The minimizations using a full atomistic (NAMD/CHARMM force field), a biaxial GB and a numerical potential from a LUT using data from the CHARMM pair potential was done. The numerical potential model is significantly more accurate in its approximation of the atomistic local minimum configuration than is the biaxial GB analytical potential function. This demonstrates that using a numerical potential founded on a direct lookup of the atomistic potential landscape significantly improves coarse grain (CG) modeling of complex molecules, possibly paving the way for accurate anisotropic system CG modeling.
一种高速数值势，其计算性能可与复杂的粗粒度解析势相媲美，从而产生了具有更高物理和化学精度的模型。这为提高各向异性系统经典分子动力学模拟的精度开辟了可能性。在这项工作中，我们报告了一种四维网格数据高速查找表（LUT）的开发，该查找表使用三次B样条插值来推导位于已知网格数据点之间的离网格值及其相关偏导数。使用来自单轴Gay-Berne（GB）势的LUT生成的粗粒度数值势的精度，对于单轴GB势及其偏导数的插值，误差范围分别在3%和5%以内。通过利用板载图形处理单元的功能，数值势模型及其偏导数的加速与解析势具有竞争力。通过比较500个萘分子盒子的最小化情况，展示了数值势的能力。分别使用全原子（NAMD/CHARMM力场）、双轴GB和使用来自CHARMM对势数据的LUT的数值势进行了最小化。数值势模型在近似原子局部最小构型方面比双轴GB解析势函数显著更准确。这表明，基于直接查找原子势面建立的数值势显著改善了复杂分子的粗粒度（CG）建模，可能为精确的各向异性系统CG建模铺平道路。
