autofitting_spice

PySpice-based BSIM4 dataset generation and multi-stage calibration workflow for 42 W/L device targets.

Project Structure

.
├── src/                        # Core modules: simulation, optimization, plotting, targets, paths
├── scripts/                    # Preferred executable entry points
├── data/targets/               # Input target CSV files
├── artifacts/                  # Runtime-generated outputs, ignored by git
├── tests/                      # Unit and smoke tests
├── docs/                       # Specs and implementation plans
├── bsim4_dataset.py            # Backward-compatible wrapper for dataset generation
└── run_calibration.py          # Backward-compatible wrapper for calibration

Recommended Commands

Run dataset generation:

MPLCONFIGDIR=/tmp/mplconfig conda run -n spice python scripts/generate_bsim4_dataset.py

Run calibration:

MPLCONFIGDIR=/tmp/mplconfig conda run --no-capture-output -n spice python scripts/run_calibration.py

Run all tests:

MPLCONFIGDIR=/tmp/mplconfig conda run -n spice python -m unittest discover -s tests -v

Key Paths

• Default target CSV: data/targets/virtual_mosfet_metrics_perturbed_5pct.csv
• Calibration outputs: artifacts/calibration_output/
• Dataset outputs: artifacts/dataset_generation/

Notes

• Core path defaults are centralized in src/project_paths.py.
• CLI entry points are centralized in src/cli.py.
• The current calibration pipeline is stable and tested, but the current best worst-case relative error is still about 5.1%, so the <3% target has not yet been reached.