2-D Dynamic-Kernel Convolutional Neural Network for Extrapolative Nonlinear RF Device Modeling

This article presents an enhanced 2-D dynamic-kernel convolutional neural network (2-D DK-CNN) framework for nonlinear radio frequency (RF) device frequency domain behavioral modeling. The model features a unique 2-D convolutional structure that jointly processes behavior model coefficients, enabling the network to directly capture harmonic interactions under strong nonlinear conditions. This structure also provides a unified feature space that supports robust extrapolation across frequencies, bias conditions, and device sizes. As a result, it addresses a key limitation of conventional behavioral models, which rely on extensive and costly measurements while offering limited extrapolation capability. To support this framework, several algorithmic strategies are introduced, including adaptive loss control for stable convergence, coefficient conditioning for explicit operating-condition encoding, multitask learning (MTL) for simultaneous harmonic response extraction, and normalization schemes tailored to hierarchical convolutional layers. The proposed model is validated on gallium nitride (GaN) high electron mobility transistor (HEMT) devices across diverse operating conditions. The results demonstrate its accuracy, robustness, and scalability. This technique promises to be a powerful and practically integrable tool for computer-aided design (CAD)-based nonlinear RF device modeling and design.