Multiproperty Deep Learning of the Correlation Energy of Electrons and the Physicochemical Properties of Molecules.

Abstract

The density-based descriptors from the information-theoretic approach (ITA) are used as features for multiproperty deep learning (DL), predicting the correlation energy and physicochemical properties of molecules. In addition to response properties (molecular polarizability α_iso and NMR shielding constant σ_iso) where ITA has been shown to work well before, we consider four conceptually distinct but practically related concepts: electron correlation, redox potential, octanol-water partition coefficient (logK_ow), and the wavelength of maximum absorption (λ_max). The DL-predicted results are in good agreement with either the calculated or experimental counterparts, indicative of the model's robustness. We verified the transferability of redox potentials of phenazine derivatives. Generalizability is observed for the λ_max data: small chromophores are used for training/validation but the test set has sizable molecules. The trained DL model outperforms the conventional TD-DFT method in terms of accuracy and efficiency. We also showcase that the isotropic quadrupole moment (Θ_iso) is a good predictor of logK_ow. This establishes that versatile density-based ITA quantities can be used to make accurate, low-cost predictions of both extensive and intensive properties, suggesting that this ITA-DL protocol has the potential for closed-loop chemistry automation. Implication of this work is straightforward, that a universal framework should be possible based on the ITA-based DL models.