Accelerating Low-k Dielectric Material Discovery: From Graph Machine Learning to Synthesis

The exploration of dielectric materials with a specific permittivity remains a significant challenge. Dielectric materials with low permittivity are widely used in semiconductor interlayers and communication substrates for reducing parasitic capacitance and latency transmission. However, discovering novel dielectric materials often relies on a trial-and-error strategy, which is inefficient and time-consuming. Herein, a graph neural network (Res-GCN) is proposed to directly predict permittivity from the connections of atoms, after which a material searching pipeline is developed based on pattern recognition approaches. The model improves relative accuracy by ∼267% over the classical Clausius-Mossotti model, ∼38% over deep neural networks, and ∼17% over crystal convolution neural networks and achieves a root-mean-squared error of ∼1.788 over different crystal symmetries. The high-throughput screening is conducted over 6000 material entries within minutes, identifying promising low-permittivity candidates, and two novel dielectric ceramics with expected properties were discovered and synthesized through only eight targeted experiments. Such machine learning approaches provide an efficient and scalable framework for accelerating the discovery of dielectric materials, significantly shortening the research cycle, and enhancing the synergy between prediction and experiment.