QMGBP-DL: a deep learning and machine learning approach for quantum molecular graph band-gap prediction

Area of Science:

• Computational chemistry and materials science.
• Application of deep learning in predicting quantum material properties.

Background:

• Accurate prediction of molecular and quantum material properties, particularly band gap, is essential for advancing drug design and material science.
• While graph neural networks and probabilistic encoders are used in molecular data analysis, their specific application for band gap prediction is an ongoing research area.

Purpose of the Study:

• To introduce QMGBP-DL, a novel deep learning framework designed to enhance the accuracy of molecular and material band gap energy predictions.
• To evaluate the effectiveness of integrating graph convolutional networks (GCNs) for molecular encoding with various machine learning models for property prediction.

Main Methods:

• QMGBP-DL utilizes a graph convolutional network (GCN) encoder to generate latent representations of chemical structures from SMILES strings.
• These latent representations are optimized using Kullback-Leibler divergence loss and subsequently used to train diverse machine learning models.
• The approach was validated on the QM9, PCQM4M, and OPV datasets, with performance compared against established methods like DenseGNN, MEGNet, and ALIGNN.

Main Results:

• QMGBP-DL demonstrated significant improvements in predicting molecular and material properties, especially band gap energy.
• The Random Forest model, when combined with GCN-derived latent spaces, showed particularly strong performance.
• Comparative analysis indicated that QMGBP-DL achieved notably lower Mean Absolute Error (MAE) values for predicting HOMO, LUMO, and band gap compared to existing methods.

Conclusions:

• The integration of GCN-based molecular encoding with traditional machine learning models, particularly Random Forest, offers a powerful and effective strategy for band gap prediction.
• QMGBP-DL facilitates accelerated discovery and design of novel materials by providing accurate property predictions.
• The study underscores the efficacy of combining advanced graph-based feature extraction with robust machine learning algorithms for complex chemical property prediction.