Bayesian Analysis of Theoretical Rotational Constants from Low-Cost Electronic Structure Methods
Bayesian Analysis of Theoretical Rotational Constants from Low-Cost Electronic Structure Methods
复制标题
DOI：
10.1021/acs.jpca.9b09982
复制
发表时间：
2020-02-06
影响因子：
2.9
通讯作者：
McCarthy, Michael
中科院分区：
文献类型：
--
作者：
Lee, Kin Long Kelvin;McCarthy, Michael
随着微波光谱学界对电子结构程序的使用越来越多，越来越需要评估常用的低成本量子化学方法的性能，特别是关于转动常数，因为这些量是指导实验的核心。在这里，我们系统地基准测试了几种低级从头算和密度泛函与旋转光谱文献中常用的各种基组相结合所提供的预测能力。在我们的分析中的数据集由6916个优化的几何形状的76个代表性的物种，高分辨率的实验气相旋转常数。我们采用了贝叶斯方法来分析每种方法和基组组合的性能，采用Hamilton Monte Carlo采样来确定旋转常数和偶极矩的理论预测的不确定性。我们的分析建立了一个准确度和不确定性的层次结构，旋转光谱文献中常用的方法，如B3 LYP和MP2，比新一代泛函（如明尼苏达家族和omega B 97 X-D）产生更低的准确度和更高的不确定性，当与中等大小的6-31+G（d）基配对时，提供了相对于计算成本的最佳性能。此外，我们提供了统计比例因子，可用于经验校正振动-旋转效应，作为进一步提高从这些相对低成本的理论方法预测的旋转常数的准确性的一种手段。作为其中的一部分，我们证明了不确定性可以用于旋转光谱的模拟，以与宽带光谱相互关联，这种方法可以用于快速有效地调查新分子的实验光谱。
With an ever-increasing usage of electronic structure programs by the microwave spectroscopy community, there is a growing need to assess the performance of commonly used, low-cost quantum chemical methods, particularly with respect to rotational constants because these quantities are central in guiding experiments. Here, we systematically benchmark the predictive power afforded by several low-level ab initio and density functionals combined with a variety of basis sets that are commonly employed in the rotational spectroscopy literature. The data set in our analysis consists of 6916 optimized geometries of 76 representative species where high-resolution experimental gas-phase rotational constants are available. We adopted a Bayesian approach for analyzing the performance of each method and basis set combination, employing Hamiltonian Monte Carlo sampling to determine the uncertainty in theoretical predictions of rotational constants and dipole moments. Our analysis establishes a hierarchy of accuracy and uncertainty, with commonly used methods in the rotational spectroscopy literature such as B3LYP and MP2 yielding lower accuracy and higher uncertainty than newer-generation functionals such as those from the Minnesota family, and omega B97X-D, which, when paired with a modestly sized 6-31+G(d) basis, provides optimal performance with respect to computational cost. Additionally, we provide statistical scaling factors that can be used to empirically correct for vibration-rotation effects, as a means to further improve the accuracy of rotational constants predicted from these relatively low-cost theoretical methods. As part of this, we demonstrate that the uncertainties can be used in simulations of rotational spectra to cross-correlate with broadband spectra, a methodology that could be used to quickly and efficiently survey experimental spectra for new molecules.