Machine learning–driven design of π-conjugated systems in isomer space with target absorption spectra: a case study of Effenberger dye

Abstract

This work presents a proof-of-concept study exploring the isomer space of the Effenberger dye to evaluate approaches for predicting optical properties aimed at the design of new compounds. A diverse library of isomers was generated, differing from the reference dye, and quantum chemical parameters were calculated using the Tight Binding Density Functional Theory (DFTB) method, a computationally efficient approximation to TD-DFT. This allowed the evaluation of a large number of molecules with reasonable accuracy. For each isomer, a broad set of molecular descriptors was calculated, including geometric, electronic, and topological parameters. The Boruta algorithm was applied to select the most relevant descriptors, reducing dimensionality and avoiding overfitting. Regression analysis was used as a simple baseline to correlate descriptors with long-wavelength transition energies. An artificial neural network (ANN) was then constructed using TensorFlow/Keras, with three hidden layers, ReLU activations, a linear output, and regularization methods. The model was trained on a representative subset, while the remaining molecules (30%) formed the test set. The ANN achieved high predictive performance, showing good agreement between predicted and reference TD-DFT values. This approach demonstrates that even a relatively small but diverse set of isomers can be used to build effective predictive models. The combination of DFTB, molecular descriptors, Boruta-based feature selection, and ANN modeling provides a versatile framework for designing new π-conjugated materials with tailored optical properties.