This project delivers a rigorous, comparison between two complementary approaches for small‑signal modeling of GaN HEMTs:
- Empirical equivalent‑circuit modeling, parameterized RLC networks fitted to measured S‑parameters, providing physical insight and circuit‑level interpretability.
- Data‑driven modeling, feature‑engineered feedforward ANNs that directly predict the real and imaginary parts of S‑parameters across frequency and bias.
We trained both approaches on high‑quality experimental datasets under two bias regimes
- Vd = 10 V, Vg sweep, and
- Vg = −4 V, Vd sweep.
A parallel DC track (Id–Vg and Id–Vd) was also developed and compared. The ANN models consistently improved prediction accuracy (up to ~50% MRE reduction on key metrics such as Re(S21)), reduced inference latency to <0.1 s/sample, and generalized better to unseen bias points, while empirical models retained interpretability and physics guarantees.
This repository contains data preprocessing, model training/evaluation code, plotting tools (Smith charts, |S| dB plots, three‑way experimental/ ANN/ empirical comparisons), and scripts to reproduce DC surrogate predictions and circuit simulations (Keysight ADS / Verilog‑A integration).
- Dual‑track comparison: Side-by-side empirical RLC circuit extraction and Artificial Neural Network (ANN) surrogates for Radio Frequency (RF) S‑parameters and DC drain current (Id) curves.
- Clean preprocessing pipeline: Outlier filtering, noise‑floor trimming, and persistent input/output scalers for deterministic inference.
- Second‑order feature expansion: Automatic polynomial feature generation including squared terms and pairwise cross‑terms of frequency, gate‑source voltage, and drain‑source voltage to expose nonlinear interactions.
- Compact, robust neural network: Four hidden layers with neuron counts 64–128–128–64, Swish activation functions, dropout regularization, Huber loss, and Adam optimizer with learning‑rate scheduling, tuned to generalize across bias conditions.
- Complex‑valued targets (dual output): Networks predict the real and imaginary components simultaneously for each S‑parameter, enabling direct comparison with circuit models and loss evaluation in complex space.
- Direct‑current surrogate tooling: Matrix‑aware functions for drain‑current versus gate‑voltage and drain‑current versus drain‑voltage prediction that parse experimental CSV layouts and generate quality plots on demand.
- Reproducible benchmarking: Scripts to compute mean relative error, mean absolute error, mean squared error, and runtime comparisons between the neural surrogate and the empirical extraction method; model weights, scalers, and fitted circuit parameters are exportable.
- Simulation‑ready outputs: Fitted circuit parameter files and comparison figures prepared for use in Keysight Advanced Design System and Verilog‑A simulation environments.
For a detailed overview and in-depth explanation of the models, you can view the following resources: