Q-DFTNet: A Chemistry-Informed Neural Network Framework for Predicting Molecular Dipole Moments via DFT-Driven QM9...

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Mechanics

Background:

• Predicting molecular properties like dipole moments is crucial in computational chemistry and materials science.
• Graph Neural Networks (GNNs) show promise for molecular property prediction but require careful benchmarking.
• Existing models may lack interpretability or computational efficiency for large-scale applications.

Purpose of the Study:

• To present Q-DFTNet, a chemistry-informed neural network (ChINN) framework for benchmarking GNNs.
• To evaluate seven GNN architectures for dipole moment prediction on the QM9 dataset.
• To identify GNNs offering an optimal balance of accuracy, interpretability, and computational efficiency.

Main Methods:

• Trained seven GNN architectures (GCN, GIN, GraphConv, GATConv, GATNet, SAGEConv, GIN+EdgeConv) for 100 epochs on the QM9 dataset.
• Evaluated models using performance metrics (MSE, MAE, R^2) and interpretability analyses (t-SNE, PCA, UMAP, residual plots).
• Assessed accuracy-complexity trade-offs based on trainable parameters and predictive performance.

Main Results:

• GraphConv achieved the lowest test MSE (0.7054) and MAE (0.6196) with minimal parameters (16.5k), indicating superior accuracy-complexity.
• GIN+EdgeConv showed strong performance, leveraging edge-awareness for enhanced expressivity.
• Attention-based models (GATConv, GATNet) underperformed despite higher complexity, while latent space analysis revealed better cluster separability for GraphConv, GIN+EdgeConv, and GCN.

Conclusions:

• GraphConv offers the best accuracy-complexity trade-off for dipole moment prediction within the Q-DFTNet framework.
• Q-DFTNet provides a chemically grounded baseline for GNN deployment in quantum chemistry and materials discovery.
• The framework highlights the importance of model architecture selection for achieving interpretable and efficient molecular property prediction.