Usage#

Everything runs through make; each target is a thin wrapper around the prismo CLI.

make validate-gradient-containers  # composed adjoint vs central finite differences
make run-containers                # the optimization; outputs/ gets figures + checkpoint.json
make animate                       # rebuild doping_evolution.{gif,mp4} from the checkpoint
make probe-objective-containers RUN_ARGS="--design outputs/checkpoint.json"

Knobs are passed through RUN_ARGS (prismo run --help lists them all):

Option

Meaning

--loss-weight w

Optimize Δneff − w·α_mode (w in n_eff per dB/cm). Default 0: loss is reported, not penalized.

--seed u|lateral|vertical

Initial junction topology. Default u, the seed of the headline run; the MMA optimum is local, so this picks the basin.

--contact-offset, --domain-width

Geometry of the shared mesh [µm].

--mesh-size

Characteristic mesh size [µm] of the shared mesh (coarser = faster).

--mode-index k

Which guided mode the eigensolve tracks (0 = fundamental).

--r-min, --move-limit, --max-iter, --ftol-rel

Density-filter radius [µm] and MMA controls.

Outputs#

Every run prints Δn_eff (warm and cold re-solve), VπLπ, modal loss and VπLπ·α, and writes to outputs/:

  • convergence.pdf — Δn_eff and VπLπ per iteration

  • doping_field.pdf — initial vs optimized signed design field

  • mode_field.pdf — |E| of the tracked mode

  • depletion_field.pdf — carriers swept out between 0 V and −5 V under the mode

  • bias_sweep.pdf — Δn_eff, α and VπLπ·α vs bias, seed vs optimized (--bias-sweep-points 0 skips it)

  • gradient_validation.pdf — relative error of the adjoint vs step size

  • loss_convergence.pdf, tradeoff.pdf — loss-aware runs

  • doping_evolution.{gif,mp4} — net doping at every evaluation

  • checkpoint.json — best design + full history (prismo animate replays it)

The vocabulary (design field, doping map, move limit, cold re-evaluation, …) is defined once in the Glossary; the reasoning behind the shared mesh, the pivoting eigensolve and the loss model is in Implementation choices.