Prediction of new ESIPT Molecules by Machine Learning and Quantum Modelling

Abstract

The discovery of novel compounds exhibiting Excited State Intramolecular Proton Transfer (ESIPT) is critical for advancing technologies in optoelectronics, biosensors, and smart materials. Conventional approaches dependent on exhaustive quantum chemical simulations are inefficient for exploring expansive molecular libraries. We introduce an integrated computational strategy that combines Density Functional Theory (DFT) with supervised machine learning to efficiently classify ESIPT propensity. Using a curated set of 100 organic molecules, we implemented and compared three classifiers: Random Forest (RF), Support Vector Machine (SVM), and a Deep Neural Network (DNN). The RF model delivered the most robust performance, attaining 94.2% accuracy and a 91.1% F1-score under stratified cross validation. Analysis of feature importance identified key electronic descriptors, including absorption wavelength and LUMO energy, as primary predictors. This hybrid DFT-ML pipeline provides a rapid and reliable method for the virtual screening and targeted design of new ESIPT active materials.