Zou Wenli, Guo Guina, Suo Bingbing, Liu Wenjian
Shaanxi Key Laboratory for Theoretical Physics Frontiers and Institute of Modern Physics, Northwest University, Xi'an 710127, Shaanxi, P. R. China.
Qingdao Institute for Theoretical and Computational Sciences, Shandong University, Qingdao 266237, Shandong, P. R. China.
J Chem Theory Comput. 2020 Mar 10;16(3):1541-1554. doi: 10.1021/acs.jctc.9b01120. Epub 2020 Feb 10.
The algebraic exact two-component (X2C) relativistic Hamiltonian can be viewed as a matrix functional of the decoupling () and renormalization () matrices. It is precisely their responses to external perturbations that render X2C-based response theories different in form from the nonrelativistic counterparts. However, the situation is not really bad. Sticking to the energy gradients, it can be shown that the nuclear derivatives of and ( and , respectively) can be transformed away to favor transformed, nucleus-independent density matrices, viz., the X2C energy gradients can be written in a form that does not depend explicitly on and . Further combined with the storage of quantities that are already available in the energy calculation, only 35 matrix multiplications are needed to construct the one-electron (relativistic) part of the X2C gradients, thereby rendering the gradient calculations very efficient. More efficiency can be gained by approximating the molecular as the superposition of the atomic ones (denoted as X2C/AXR) and by further approximating the molecular also as the superposition of the atomic ones (denoted as X2C/AU): The numbers of matrix multiplications required for constructing the one-electron (relativistic) parts of the X2C/AXR and X2C/AU gradients are reduced to 18 and 4, respectively. Similar approximations can also be applied to the X2C Hessian. It will be shown numerically that the X2C/AXR gradients and Hessians are extremely accurate (almost indistinguishable from the full X2C ones), whereas the X2C/AU ones do have discernible errors but which are tolerable in view of the dramatic gain in efficiency.
代数精确双分量(X2C)相对论哈密顿量可被视为解耦矩阵()和重整化矩阵()的矩阵泛函。正是它们对外部微扰的响应使得基于X2C的响应理论在形式上不同于非相对论对应理论。然而,情况并非真的很糟。坚持能量梯度,可以证明和(分别为和)的核导数可以被消除,以支持变换后的、与核无关的密度矩阵,即X2C能量梯度可以写成一种不明确依赖于和的形式。进一步结合能量计算中已有的量的存储,构建X2C梯度的单电子(相对论)部分仅需要35次矩阵乘法,从而使梯度计算非常高效。通过将近似为原子的叠加(记为X2C/AXR)以及进一步将近似为原子的叠加(记为X2C/AU),可以获得更高的效率:构建X2C/AXR和X2C/AU梯度的单电子(相对论)部分所需的矩阵乘法次数分别减少到18次和4次。类似的近似也可以应用于X2C海森矩阵。数值结果将表明,X2C/AXR梯度和海森矩阵极其精确(几乎与完整的X2C梯度无法区分),而X2C/AU梯度确实存在可察觉的误差,但鉴于效率的显著提高,这些误差是可以容忍的。