Unimodal Multitask Deep Learning for Accurate DC and RF Modeling and Prediction of InP HBT

This article presents a deep learning with the unimodal multitask framework for accurate DC and RF modeling and prediction of indium phosphide (InP) heterojunction bipolar transistors (HBTs). By leveraging advanced artificial intelligence (AI) algorithms, the framework captures complex nonlinear device behaviors and cross-domain correlations that are difficult to handle with conventional physics-based models. The proposed model simultaneously predicts DC characteristics with different emitter window width and S-parameters with various frequency ranges. The attention-enhanced residual feature extractor combined with twin neural network branches improves the task-prediction accuracy and feature reuse. The proposed model is verified using InP HBTs with the emitter areas of 0.5 × 10, 0.6 × 10, and 0.7 × 10 &mu;m². Comparison of simulated I–V characteristics and S-parameters against measurement results show that the proposed model can accurately predict the device characteristics from DC to 110 GHz with <0.5% error. The proposed model is expected to retain its predictive accuracy at substantially higher frequencies. The proposed model provides a robust and computationally efficient framework for intelligent device characterization in the next-generation RF design.