Transient gate-driver optimization with fcmaes-core and thevenin

This standalone Rust tutorial optimizes a nanosecond-scale gate-input network without putting Python or a foreign-function call inside the objective. It uses thevenin 0.5.0 for transient simulation and fcmaes-core 0.1.3 for parallel constrained multi-objective search.

It is deliberately separate from the sindr AC circuit-design tutorial. sindr supplies the fast small-signal filter model used there; thevenin supplies explicit transient duration, timestep, and trapezoidal integration needed here.

The model

The example is a lumped MOSFET gate-input network, not a complete switching converter:

10 V pulse ─ Rdrive ─ 8 nH trace ─┬─ gate = 4 nF ─ ground
                                  ├─ Rsnub ─ 2 nF ─ ground
                                  └─ Rleak = 1 MΩ ─ ground

The series-RC branch damps the trace/gate resonance but loads the rising edge. MODE searches two normalized controls, each decoded logarithmically:

The shared gate-driver.cir is the single source for both simulators. Its 0–10 V pulse begins at 5 ns with a 1 ns edge. Every publication evaluation uses trapezoidal integration, a 50 ps maximum timestep, and a 120 ns observation window.

This topology is intentionally small enough that the optimization and the independent validation are auditable. It represents an effective gate capacitance and interconnect resonance. It does not represent nonlinear capacitance, Miller charge, drain switching, magnetic coupling, temperature, or semiconductor loss.

The exact crates.io fcmaes-core = "=0.1.3" pin deliberately validates the published optimizer/simulator pairing. The manifest retains a commented local-path override for testing working-tree optimizer changes before updating the recorded pin.

Objectives and constraints

All values passed to MODE are minimized. The two objectives are:

1. interpolated 10–90% gate rise time in ns;
2. overshoot above 10 V as a percentage of 10 V.

The two constraints are feasible at <= 0:

peak driver current - 5 A <= 0
2% settling time - 75 ns <= 0

The current is reconstructed from (Vdrive - Vtrace) / Rdrive. The 10% and 90% crossing times are linearly interpolated between adjacent transient samples. Settling is the final sample outside 9.8–10.2 V; a waveform still outside the band at 120 ns therefore receives the maximum recorded settling time and is infeasible.

This measurement layer matters as much as the simulator. Using a grid index for a crossing would create the same staircase objective that the companion AC tutorial removes from peak-frequency extraction.

Validation before optimization

The tutorial was admitted only after two numerical gates passed.

Independent ngspice comparison

The fixed validation set is the Cartesian product of seven inclusive values for each normalized control: 49 designs including every bound. Rust writes their decoded values and thevenin measurements. The independent ngspice_reference.py harness loads libngspice directly, replays the identical template synchronously, and extracts separate transient vectors. It is never called by MODE.

Every maximum passes. The checked summary.json contains the unrounded medians, 95th percentiles, maxima, limits, and pass status.

Timestep refinement

Three designs—strongly ringing, intermediate, and heavily damped—were replayed at maximum timesteps of 100, 50, and 25 ps. Comparing the publication 50 ps setting with 25 ps gives these worst differences:

MODE result

The seed-42 publication run used 4,096 candidate evaluations, population 128, and 16 candidate workers. It completed in 1.151 s on the recorded Ryzen 9 9950X machine and retained 128 feasible nondominated designs. In this smooth two-objective problem the final population resolves one continuous trade-off curve, so a fully nondominated population is expected.

The plot marks the two objective extremes and a compromise selected by the sum of independently range-normalized objectives:

Parallel candidate evaluation

The fixed 512-design benchmark was repeated five times at each width. All 7,680 transient evaluations succeeded:

thevenin transient evaluation is serial for each small circuit; fcmaes-core owns parallelism across independent candidates.

Reproduce

The pure-Rust optimization, timestep study, grid, and scaling experiment need only Cargo:

cargo run --release -- \
  --mode all --preset publication --workers 16

The independent reference step additionally needs the ngspice shared library. On Debian or Ubuntu:

sudo apt install libngspice0 libngspice0-dev
python validation/ngspice_reference.py \
  --candidates results/publication/validation/candidates.csv \
  --output results/publication/validation/ngspice.csv
python validation/compare_results.py
python plot_results.py --write
python plot_results.py --check

The exact recorded environment is in environment.json. Dependency licensing provenance is documented separately in DEPENDENCY_NOTICE.md.

Artifacts

results/publication/
  environment.json
  mo/{run.json,pareto.csv,convergence.csv,waveforms.csv}
  validation/
    {candidates.csv,thevenin.csv,ngspice.csv,comparison.csv,summary.json}
    {timestep.csv,scaling.csv}

Native Rust writes the optimizer, waveform, timestep, and scaling evidence. The ngspice harness writes its independent raw measurements. The comparison script writes the versioned publication decision, and plot_results.py checks the complete bundle before rendering deterministic SVGs.

Test

cargo fmt --all -- --check
cargo clippy --all-targets -- -D warnings
cargo test --release
cargo doc --no-deps
cargo run --release -- \
  --mode all --preset smoke --workers 2 --no-output
python -m py_compile validation/*.py plot_results.py
python plot_results.py --check

The Rust tests cover exact decoding bounds, interpolated crossings, finite transient replay, constrained MODE smoke search, and the 50-to-25 ps convergence gates. The Python checks enforce the independent 49-design comparison and byte-for-byte figure reproduction.