Combining fingerprint-based deep neural networks with machine learning potentials for the study of eumelanin...

Melanin is a promising, versatile biomaterial, but its heterogeneous, complex structure prevents the development of melanin-based applications. Previously we introduced a deep neural network (DNN) model, trained on DFT, to estimate the thermodynamic and vertical absorption (S₁) of minimal eumelanin components (5,6-dihydroxyindole, DHI, and dimers). The model uses purpose-made structural fingerprints (FP) that represent the regio- and stereochemical diversity and the different oxidation states of the dimers with minimal extension. Here we extend the modeling to the three lowest excited states and oscillator strengths and add an input node describing interfragment bonding to improve the prediction of the optical absorption. The resulting descriptor set is superior in terms of parsimony and predictive efficiency to general FP descriptors. Good fittings are obtained with errors of 3% for thermodynamic stability, with an accuracy similar to or better than that of state-of-the-art machine learning (ML) potentials. Good performance is also obtained for the vertical absorption energies and oscillator strengths, with errors of 5-6% and 3-7%, respectively. For the oscillator strength, the models are also trained on the logarithm of this quantity to improve the predictions for medium or weakly absorbing states. To reduce the computational cost associated with training thermodynamic stability, DFT can be replaced by AIQM2 maintaining the predictive capacity. Our results show that FP-based descriptors are convenient to describe materials where diversity arises from combining repeated units with small variations, like DHI with different oxidation states and bonding sites in the present case. The fact that predictions of optical properties are improved by adding the interfragment bond descriptor uncovers the importance of this feature for optical absorption. Finally, the FP-based approach can be made more efficient by training the models on ML potential data rather than DFT.