Accelerated Material Phonon Property Prediction in Limited Data Regime Through a Robust Context-Gated GCN Hybrid Architecture

ABSTRACT Prediction of lattice vibration frequencies (phonon density of states peaks) is a fundamental parameter in the design of thermal and electronic materials. Ab-initio simulation methods such as Density Functional Theory (DFT) offer high accuracy but are hampered by high computational costs, especially for large-scale material screening. Graph-based Deep Learning approaches (Graph Neural Networks/GNNs) offer fast solutions, but often suffer from convergence instability (gradient explosion) and low accuracy when trained on small, highly variable datasets. This study proposes a hybrid architecture: Context-Gated Robust GCN. This method integrates (1) a Graph Convolutional Network stabilized by target normalization and gradient pruning, (2) an adaptive Gating Mechanism that modulates crystal features based on environmental context, and (3) Gradient Boosting (XGBoost) as the final predictor. Evaluation on the Matbench Phonons dataset shows that this hybrid method achieves stable convergence with an R2 score of around 0.88 (depending on the data split), significantly outperforming a complex GNN model without stabilization. Residual analysis shows a normal error distribution, validating the model's reliability as a fast material pre-selection tool. Keywords: Materials Informatics, Graph Neural Networks, XGBoost, Phonons, Hybrid AI, Small-Data Regime.