(521q) Computer Aided Molecular Design Coupled to Deep Learning Techniques As a Less-Expensive Approach to Design Organic Photoredox Catalysts

To date, the prediction of electronic absorption spectra of organic molecules by quantum chemical methods is still limited, with differences of even 100 nm between the theoretical and experimental values. In this work, we present a computationally cheaper approach for the design of organic polyaromatic molecules as potential organic photoredox catalysts. We focus on the first step of predicting their absorbance in the visible region of the electromagnetic spectrum. This problem is important to promote the development and use of solar energy in chemical processes. The design of new chromophoric molecules is approached as a two-step procedure. In the first step, a Mixed-Integer Nonlinear Programming (MINLP) model was formulated for the in-silico design of organic chromophores. A quantitative structure-activity relationship model was developed for the prediction of the difference in energy from the HOMO to the LUMO orbitals of organic dyes. This molecular design approach produced 4 photocatalytic molecules; 3 of these molecules have not been previously reported. In the second step, a large chemical compound database was leveraged to develop a deep learning model aimed at forecasting the absorption of visible light by such and related compounds. We have taken advantage of the deep learning model to verify that the light absorption between this model and the results of the MINLP formulation match with reasonable accuracy. For 3 of the 4 photocatalytic molecules, both results appear similar, verifying that the proposed photocatalytic molecules can absorb light close to the target wavelength. The overarching aim of this research is to extend the use of less expensive computational methods for the prediction of the rest of the important photophysical features that make up a competitive organic photoredox catalyst.