Predicting HOMO-LUMO Gaps Using Hartree-Fock Calculated Data and Machine Learning Models.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning Applications

Background:

• Quantum mechanics (QM) calculations for Highest Occupied Molecular Orbital-Lowest Unoccupied Molecular Orbital (HOMO-LUMO) gaps are computationally expensive.
• Experimental determination of HOMO-LUMO gaps is time-consuming and costly.
• Machine Learning (ML) presents a cost-effective and rapid alternative for predicting these electronic properties.

Purpose of the Study:

• To develop and evaluate ML models for predicting HOMO-LUMO energy gaps in organic molecules.
• To explore the use of molecular descriptors derived from Simplified Molecular Input Line Entry System (SMILES) for ML model training.
• To create an ensemble ML model for enhanced prediction accuracy and applicability to diverse molecular structures.

Main Methods:

• Utilized a dataset of 46,717 small molecules with HOMO-LUMO gap values from Hartree-Fock (HF) calculations.
• Generated molecular descriptors using RDKit from SMILES representations.
• Trained and compared various regression-based ML models including LightGBM, Bidirectional LSTM, CatBoost, and Multilayer Perceptron (MLP).

Main Results:

• LightGBM, Bidirectional LSTM, CatBoost, and MLP models achieved a mean absolute error (MAE) below 0.25 eV.
• A weighted ensemble model combining LightGBM, Bidirectional LSTM, and MLP achieved an MAE of 0.1660 eV.
• SHAP analysis identified 20 critical molecular descriptors, and models were adapted for experimental HOMO-LUMO gap estimation up to carbon number 50.

Conclusions:

• ML models, particularly ensemble approaches, offer efficient and accurate prediction of HOMO-LUMO gaps.
• The developed models demonstrate versatility for both small and large molecules, aiding in high-throughput screening.
• The study highlights the potential of ML in accelerating chemical discovery by reducing reliance on intensive computations and experiments.