Predicting Structure-Dependent Dielectric Responses in High-k Oxides using Physics-Informed Graph Neural Networks

Abstract

As semiconductor scaling extends into the deep nanometer regime, traditional silicon dioxide gate dielectrics encounter fundamental physical limitations imposed by due to severe quantum tunneling leakage currents. High-k gate dielectric materials have therefore become essential for modern nanoelectronics, enabling continued device miniaturization through equivalent oxide thickness scaling and substantially suppressing leakage. However, the discovery and design of optimal high-k gate dielectrics that simultaneously exhibit a high dielectric response, wide bandgaps, and thermodynamic stability remain a formidable challenge. These macroscopic properties originate from intricate interactions among crystal structure, chemical composition, and specific electronic and ionic contributions. While first-principles calculations offer vital insights into these quantum-mechanical mechanisms, conducting systematic, high-throughput explorations of high-k oxide candidates across an expansive chemical and structural space is computationally prohibitive.To address this bottleneck in intelligent materials screening, we develop a physics-informed machine learning framework for predicting the dielectric properties of high-k oxide materials. The framework integrates structure-aware graph neural networks, transfer learning, and hybrid regression models to capture structure-dependent dielectric behavior. The model first undergoes multi-target pretraining on a diverse dataset of inorganic materials, enabling the network to learn general structure-property relationships across a wide range of crystal chemistries and bonding environments. Crystal structures are encoded using a graph neural architecture based on the Atomistic Line Graph Neural Network, which captures atomic connectivity and higher-order geometric correlations associated with bonding environments and angular interactions. Physically motivated descriptors are dynamically injected into the neural representation through a feature-wise linear modulation mechanism, allowing the model to adapt its latent embeddings to material-specific physicochemical properties. A multi-head prediction module simultaneously models the electronic and ionic contributions to the dielectric response, improving representation sharing and task consistency.To further enhance predictive accuracy, the learned graph embeddings, explicit physical descriptors, and intermediate neural outputs are fused into a stacking feature space and processed by gradient-boosted decision tree regressors. This hybrid strategy combines deep structural representation learning with ensemble regression methods. Evaluations on a rigorously separated test set demonstrate strong predictive performance for electronic, ionic, and total dielectric responses. These results indicate that integrating structure-aware graph representations with explicit physical guidance and ensemble learning improves predictive generalization and physical consistency, thereby providing a highly scalable computational framework for accelerating the discovery and design of next-generation high-k gate dielectric materials.