Farahvash Ardavan, Lee Chee-Kong, Sun Qiming, Shi Liang, Willard Adam P
Department of Chemistry, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA.
Tencent America, Palo Alto, California 94306, USA.
J Chem Phys. 2020 Aug 21;153(7):074111. doi: 10.1063/5.0016009.
In this manuscript, we develop multiple machine learning (ML) models to accelerate a scheme for parameterizing site-based models of exciton dynamics from all-atom configurations of condensed phase sexithiophene systems. This scheme encodes the details of a system's specific molecular morphology in the correlated distributions of model parameters through the analysis of many single-molecule excited-state electronic-structure calculations. These calculations yield excitation energies for each molecule in the system and the network of pair-wise intermolecular electronic couplings. Here, we demonstrate that the excitation energies can be accurately predicted using a kernel ridge regression (KRR) model with Coulomb matrix featurization. We present two ML models for predicting intermolecular couplings. The first one utilizes a deep neural network and bi-molecular featurization to predict the coupling directly, which we find to perform poorly. The second one utilizes a KRR model to predict unimolecular transition densities, which can subsequently be analyzed to compute the coupling. We find that the latter approach performs excellently, indicating that an effective, generalizable strategy for predicting simple bimolecular properties is through the indirect application of ML to predict higher-order unimolecular properties. Such an approach necessitates a much smaller feature space and can incorporate the insight of well-established molecular physics.
在本手稿中，我们开发了多个机器学习（ML）模型，以加速一种方案，该方案用于从凝聚相六噻吩系统的全原子构型对基于位点的激子动力学模型进行参数化。该方案通过分析许多单分子激发态电子结构计算，在模型参数的相关分布中编码系统特定分子形态的细节。这些计算得出系统中每个分子的激发能以及成对分子间电子耦合网络。在此，我们证明使用具有库仑矩阵特征化的核岭回归（KRR）模型可以准确预测激发能。我们提出了两个用于预测分子间耦合的ML模型。第一个利用深度神经网络和双分子特征化直接预测耦合，我们发现其性能不佳。第二个利用KRR模型预测单分子跃迁密度，随后可对其进行分析以计算耦合。我们发现后一种方法表现出色，这表明预测简单双分子性质的一种有效、可推广的策略是通过间接应用ML来预测高阶单分子性质。这种方法需要的特征空间要小得多，并且可以纳入成熟分子物理学的见解。
