Enhancing Molecular Machine Learning with Quantum-Chemical Descriptors and Hybrid Modeling

Abstract

The integration of quantum chemistry with Machine Learning (ML) offers transformative potential for molecular property prediction, critical for drug design and catalyst development. However, challenges like high computational costs and unreliable predictions in safety-critical applications necessitate robust frameworks. This study aims to develop and evaluate a quantum-enhanced ML framework using a Graph Neural Network (GNN) to predict properties like HOMO-LUMO gaps, binding affinities and reaction energies, focusing on Uncertainty Quantification (UQ), active learning and transfer learning to optimize efficiency and reliability. A GNN model, enhanced with quantum descriptors (MO energy, total electron density) from PySCF, was trained on simulated QM9 subsets and real-world proxies. Techniques included monte carlo dropout and ensemble methods for UQ, active learning to reduce quantum calculations and transfer learning for improved convergence. Performance was evaluated using Mean Absolute Error (MAE) across datasets, with visualization of loss, uncertainty and computational cost. The framework achieved a QM9 MAE of 3.406, outperforming baselines (MAE ~3.5-4.0), but showed higher errors (4.2-4.3) for drug and catalyst proxies due to limited quantum data. UQ provided reliable confidence intervals (e.g., [4.2, 5.4]) and active learning reduced costs by 50%. The framework demonstrates efficacy for QM9-like tasks but requires optimization for real-world applications. Future work should scale to larger datasets, integrate quantum computing and validate with experimental data.