ANN for MOSFET SPICE Model Extraction

This project explores how an artificial neural network can learn MOSFET electrical behaviour from simulated I-V data and act as a first step toward SPICE model extraction.

The current implementation is intentionally lightweight:

• Generates MOSFET drain current sweeps using a compact Shichman-Hodges style model with process/device variation.
• Trains a PyTorch neural network surrogate to predict log10(|Id| + eps) from bias and geometry inputs.
• Evaluates generalisation on held-out operating points.
• Tests robustness to imperfect input data by injecting bias noise.

Project Layout

mosfet_ann/
  data.py          Synthetic MOSFET data generation
  model.py         ANN training, prediction, evaluation, persistence
  robustness.py    Noise-sensitivity experiment
scripts/
  generate_data.py
  train_ann.py
  predict_current.py
  evaluate_robustness.py
docs/
  report_outline.md
  literature_notes.md
tests/
  test_pipeline.py

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Quick Start

Generate a dataset:

python scripts/generate_data.py --samples 8000 --output data/mosfet_iv.csv

Train the ANN:

python scripts/train_ann.py --data data/mosfet_iv.csv --model artifacts/mosfet_ann.pt

Predict drain current for one MOSFET bias/device point:

python scripts/predict_current.py \
  --model artifacts/mosfet_ann.pt \
  --vgs 0.8 \
  --vds 0.8 \
  --width-um 1.0 \
  --length-um 0.18

Evaluate robustness to noisy bias inputs:

python scripts/evaluate_robustness.py \
  --data data/mosfet_iv.csv \
  --model artifacts/mosfet_ann.pt \
  --output artifacts/robustness.csv

Current Model Inputs

The ANN uses:

• vgs: gate-source voltage
• vds: drain-source voltage
• vbs: body-source voltage
• width_um: MOSFET width in micrometres
• length_um: MOSFET length in micrometres
• temperature_c: temperature in Celsius
• vth0: nominal threshold voltage
• kp: transconductance parameter
• lambda_: channel-length modulation
• gamma: body-effect coefficient
• phi: surface potential

The target is log10(abs(id_a) + 1e-15), which is much easier for an ANN to learn than raw drain current across many orders of magnitude.

Next Technical Steps

1. Replace or complement the analytic generator with ngspice simulations.
2. Add C-V data targets, not just I-V.
3. Compare architectures: MLP, residual MLP, monotonic networks, and physics-informed loss terms.
4. Export a SPICE-compatible behavioural source or Verilog-A approximation.
5. Validate against measured MOSFET characterisation data when available.