Machine Learning-Based Prediction of Electronic and Thermodynamic Properties of Small Organic Molecules Using the QM9 Database

Abstract

The ability to accurately predict the electronic and thermodynamic properties of molecules is crucial in computational chemistry, molecular screening and materials discovery. Conventional quantum-chemical calculations provide reliable estimates but require substantial computational resources, creating a need for efficient and interpretable predictive approaches. This study developed and compared descriptor-based machine-learning models for predicting HOMO energy, HOMO–LUMO gap, free energy at 298.15 K, and heat capacity at 298.15 K using the QM9 database. A curated dataset comprising 133,885 small organic molecules described by 20 numerical molecular descriptors was analysed. Ridge Regression, Random Forest, and Extra Trees models were trained using an 80:20 train–test split and evaluated through mean absolute error, root mean square error, and the coefficient of determination (R²). Feature-importance analysis and residual evaluation were performed to assess model interpretability and prediction stability. Random Forest achieved the highest accuracy for HOMO energy (R² = 0.8227) and HOMO–LUMO gap (R² = 0.8992), Ridge Regression produced near-perfect prediction of free energy at 298.15 K (R² = 1.0000), and Extra Trees yielded the best performance for heat capacity (R² = 0.9748). Exact molecular weight, molecular weight, oxygen atoms, carbon atoms, and hydrogen-bond acceptors emerged as the most influential descriptors, while residual distributions centred near zero indicated minimal systematic prediction bias. The proposed framework offers a reliable, interpretable, and computationally fast solution for initial prediction of molecular properties and high-throughput molecular screening, but further testing with a more chemically diverse set of molecules is encouraged to expand its applicability.