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

Predicting molecular dipole moments is essential for quantum chemistry and materials science applications. In this study, we introduce Q-DFTNet: a Chemistry-Informed Neural Network framework designed to systematically benchmark and interpret graph neural networks (GNNs) for molecular dipole prediction. Seven GNN architectures: GCN, GIN, GraphConv, GATNet, GATConv, SAGEConv, and GIN+EdgeConv, were trained on the QM9 dataset for 100 epochs and evaluated using mean squared error (MSE), mean absolute error (MAE), coefficient of determination (R$^2$), and clustering validity. GraphConv attained the best predictive performance with a test MSE of 0.8514, MAE of 0.6407, and R$^2$ of 0.621, outperforming other models with only 16.5k parameters. Attention-based models (GATNet, GATConv) exhibited faster convergence but lower accuracy, with MSEs exceeding 1.09 and R$^2$ scores below 0.52. Latent spaces were visualized via t-SNE, PCA, and UMAP projections, revealing structural encoding differences across models. Clustering quality, assessed using Silhouette Score (0.4658 for GraphConv), Davies-Bouldin Index (0.7476 for GCN), and Calinski-Harabasz Score (2411.73 for GraphConv), demonstrated how Q-DFTNet captures chemically coherent groupings. A k-means analysis of 1000 molecules in latent space revealed that GIN+EdgeConv formed the most interpretable clusters, with dipole means ranging from 2.407 to 2.778 Debye. Q-DFTNet provides a unified framework for prediction, representation learning, and molecular interpretability, offering valuable guidance for neural architecture selection in quantum machine learning pipelines.