Adaptive structural fusion for dielectric constant prediction using physics-guided graph neural networks

Abstract Accurate prediction of dielectric response is essential for improving the discovery of materials for electronic, energy-storage, radio-frequency, sensing, and optoelectronic applications. However, dielectric datasets are typically dominated by materials with low to moderate dielectric response, while high-dielectric materials are severely under-represented, making reliable prediction across the full property range challenging. This work proposes a physics-guided extension of the Crystal Graph Convolutional Neural Network (CGCNN) for total dielectric constant prediction through Adaptive Structural Fusion (ASF). This refines learned atomic representations using five physically motivated local descriptors: coordination number, mean bond length, bond-angle variation, local electronegativity difference, and atomic polarizability. A context-dependent Knowledge-Guided Graph Aggregation (KGGA) module is subsequently evaluated both independently and in combination with ASF to investigate whether adaptive graph aggregation depends on the quality of the underlying node representations. Four controlled architectures -the baseline CGCNN, CGCNN with KGGA, CGCNN with ASF, and the proposed model CGCNN with both ASF and KGGA- were evaluated on an identical frozen split of a curated benchmark containing 5,874 inorganic crystalline materials. Relative to the baseline CGCNN, the proposed framework – CGCNN with both ASF and KGGA- reduced the mean absolute error (MAE) from 11.139 to 8.828, decreased the root mean square error (RMSE) from 53.315 to 42.275, and increased the coefficient of determination (R²) from 0.214 to 0.506, corresponding to approximately 21% reductions in MAE and RMSE. Paired material-level statistical analysis further demonstrated a mean absolute-error reduction of 2.311 (95% bootstrap confidence interval: 0.960–3.736) and lower prediction error for 68.4% of the 588 held-out test materials. Regime-wise evaluation showed that the proposed model achieved the lowest MAE in five of six predefined dielectric regimes, a descriptive observation that should be interpreted cautiously given the sparse high-dielectric test samples. Representative material analysis further illustrated substantial error reductions for selected chemically and structurally diverse inorganic compounds. These results indicate that physics-guided representation refinement and adaptive graph aggregation are complementary strategies for improving dielectric property prediction while highlighting the importance of regime-aware evaluation for long-tailed materials datasets.