DOI: 10.1109/TMTT.2026.3666090   IEEEXplore: https://ieeexplore.ieee.org/document/11415307

A physics-informed neural network (PINN)-based scalable large-signal model for GaN HEMTs.

0.5 – 20.5 GHz (small-signal) & 5.8 – 9.0 GHz (large-signal) demonstrated.

Charge densities, closely associated with DC and RF performance of GaN HEMT devices, are modelled with PINN-based large signal modeling approaches. Self-heating and trapping effects in GaN HEMTs are embedded into the proposed model and contribute to the model’s accuracy and strong scalability.

Highly automated, lower cost, higher accuracy tools for simulation and design of high power, high frequency GaN HEMT-based commercial, satellite and defense communication subsystems.

Higher physical consistency with real devices, higher near-threshold accuracy and no dependence on dataset optimization compared to classical artificial neural network (ANN) techniques. Higher degree of automation is possible, and fewer manual inspection counts per run and fewer number of iterations with shorter workflow times are required when compared to classical and physical ANN approaches. The demonstration in this work required 4 iterations and 1.5 hours of workflow time compared to 9 iterations over 2.5 hours for a classical ANN modeling approach and 69 iterations over 35 hours physical ANN modeling approach demonstrated in prior works.