Characterizing Microelectronic Devices via Scalable, Confinement-Aware Equivariant Networks

Modern microelectronic devices are comprised of materials with critical dimensions of a few nanometers. At these sizes, material properties change in nontrivial ways due to quantum confinement and atomic-level variability, creating a multi-scale modeling challenge that requires atomistic simulations for accurate prediction. However, such simulations are often prohibitively slow or intractable, making highly expensive iterative rounds of experimentation the default option. To address this issue, we introduce EBFormer, a geometry-aware equivariant neural network that predicts electronic properties of nanostructures by jointly capturing atomistic interactions and geometric effects, achieving orders of magnitude speed-up over state-of-the-art physical simulators while preserving high accuracy. This is accomplished through the introduction of a boundary cross-attention mechanism, a scalable approach to augment local graph convolution with information of the nanostructure geometry. We validate EBFormer on nanowire and nanosheet transistors, representing advanced modern microelectronic architectures, and show superior in-distribution and out-of-distribution performance on both material property inference and downstream device characteristics compared to leading architectures. Combined with superior asymptotic scalability and data- and parameter-efficiency, our work paves a pathway to atomistic, automated, high-throughput and predictive nanoscale design that is otherwise not available today.