Accelerating Excited-States Calculations using Active Learning Configuration Interaction for Cyclic Organic Molecules

Understanding the electronic excited states of organic materials is key to the design of optoelectronic, photocatalytic and photovoltaic devices. For molecular systems, multiconfigurational methods can be used to compute the excitation states of such molecules as electronic correlations are included. A conceptually simple and robust approach among the multiconfigurational methods is the selected configuration interaction (SCI) theory, which utilizes only the energetically important electronic configurations for Configuration Interaction (CI) calculations. Compared to the Full Configuration Interaction (FCI) theory, SCI calculations can save a significant amount of computational cost, allowing to tackle much larger systems of interest. However, the success of the SCI calculations depends on how important configurations are selected, so different approaches have been introduced. To address this challenge, we developed the Active Space Configuration Interaction (ALCI) method [1] to predict the lowest singlet–singlet vertical excitation of several polycyclic aromatic hydrocarbons (i.e., acenes and pyrene), inspired by the Machine Learning Configuration Interaction (MLCI) approach [2] used for ground state calculations of small molecules. In the ALCI approach, a binary classification machine learning model is employed to identify candidate important configurations for subsequent SCI calculation during iterative CI calculations. It has been shown that our approach can accurately predict the excitation energies compared to CASCI results for relatively small- and medium-sized systems (i.e., active spaces up to (16e, 16o)), and capture the general trend of excitation energies of the systems up to (26e, 26o), for which a complete active space CI (CASCI) calculation is not yet affordable. In this talk, we present an updated scheme of the ALCI approach to predict excited states of various cyclic organic molecules in the QUEST database [3]. The updates include several features: First, unimportant configurations identified near the importance metric in previous cycles are recycled to obtain more accurate results. Second, any transition, not just for the lowest singlet-singlet transition, can be selected for calculation. Third, GPU-supported ML algorithms (i.e., LightGBM, XGBoost, Gaussian Process, ANNs) are implemented to speed up the ML training/prediction steps. We will also analyze the identified important/unimportant configurations to discuss what importance metric would be appropriate.<br/><br/>References<br/>[1] W. Jeong, C. A. Gaggioli, L. Gagliardi. <i>J. Chem. Theory Comput.</i>, <b>2021</b>, <i>17</i>, 7518–7530.<br/>[2] J. P. Coe. <i>J. Chem. Theory Comput.</i> <b>2018</b>, <i>14</i>, 5739–5749.<br/>[3] M. Véril, A. Scemama, M. Caffarel, F. Lipparini, M. Boggio-Pasqua, D. Jacquemin, P.-F. Loos. <i>WIREs Comput. Mol. Sci.</i> <b>2021</b>; e1517