A Gaussian process regression (GPR) quest to predict HOMO-LUMO energy

Machine learning methods employ statistical algorithms and pattern recognition techniques to learn patterns and make predictions based on statistical patterns. Global reactivity descriptors, such as HOMO-LUMO energy, chemical potential (µ), chemical hardness (η), softness (σ) and electrophilic index (ω) are predicted using Gaussian process regression (GPR) machine learning method. GPR predicted values are in close agreement with the values obtained via ab initio methods. Over 85% prediction accuracy in HOMO energies and reactivity parameters is observed, while LUMO energies were in good range with the DFT evaluations. An appropriate kernel combination with proper tuning of parameters and the selection of quality correlated data can make the GPR model robust and powerful. Machine learning models like GPR could play a pivotal role in assisting and accelerating ab initio calculations and providing insights for highly complex molecular systems.