Prototype · private beta

Device simulation that shows its work.

TCAD Studio is a browser workbench for 2D simulation of GaN HEMTs and silicon MOSFETs: a readable input deck, a structure view, results you can compare, a target search, and a Claude-powered assistant that only reports numbers the solver actually produced.

DEVSIM physics · Bayesian design search · Claude with tool use

nmos_dielectric_choice.deck2D · DC
title "NMOS with selectable gate dielectric"# Target-search demo: which dielectric and gate voltage give a required drain current?set DIEL=SiO2 options=SiO2,Si3N4,Al2O3,HfO2 label="Gate dielectric"set tox=0.005 min=0.002 max=0.02 unit=um label="Dielectric thickness"set VG=1.0 min=0 max=3 unit=V label="Gate bias" region name=diel material=$DIEL x=0.4,0.9 y=-tox,0region name=si   material=Si    x=0,1.3   y=0,0.6 electrode name=gate   type=gate wf=4.1 x=0.4,0.9 y=-tox,-toxelectrode name=source type=ohmic       x=0,0.3   y=0,0electrode name=drain  type=ohmic       x=1.0,1.3 y=0,0electrode name=body   type=ohmic       x=0,1.3   y=0.6,0.6 doping uniform region=si acceptor=3e17doping gaussian donor=1e20 x=0,0.4   y=0,0 char=0.0332 lateral=0.7doping gaussian donor=1e20 x=0.9,1.3 y=0,0 char=0.0332 lateral=0.7 models ith=0.1solve op vg=VG vd=1solve idvg vd=1 vg=-0.5:2:0.1
88.6mA/mmVg = 1.0 V
116.8mA/mmVg = 1.1 V
147.6mA/mmVg = 1.2 V
308mS/mmpeak gm
Short transfer sweep at Vd = 1 V from a local verification run — not a full-range characterisation.

02 · In the workbench

Built for the engineer at the bench.

TCAD Studio results view showing extracted metrics and transfer and transconductance plots for an NMOS device
FIG. 1Results view: extracted metrics, transfer and transconductance curves, and the files behind them. Captured during local verification with test projects.

03 · The assistant

AI is not the physics engine. It makes the physics engine easier to use.

The assistant is Claude, working through the same tools you do. It sets up runs and design searches from a plain-language request, and reports back with run IDs you can open.

  • Every number comes from a run.Figures it reports must come from a simulation result in the conversation, never an estimate.
  • It states model limits.When a simplification matters — no self-heating, classical 2DEG — it says so.

How the assistant works →

Engineer

Which gate length gives the highest transconductance on the GaN HEMT?

Assistant · tool calls

list_devices()
optimize_design(device="gan_hemt", objective={metric:"gm_max", direction:"maximize"}, variables=[{name:"Lg", …}])

Reports the best gate length found, its gm and the run IDs — and notes that self-heating isn't modelled.

Illustrative. The tool and metric names are the real ones.

04 · Physics

Established physics, stated plainly.

2D drift-diffusion on the open-source DEVSIM engine, with the model limits published rather than hidden.

Device

AlGaN/GaN HEMT

Polarisation-induced 2DEG, transfer and output sweeps, off-state breakdown estimate.

Device

Silicon NMOS

Threshold, subthreshold swing, on/off current and transconductance.

Device

Custom 2D deck

Rectangular regions, boundary contacts, doping profiles and DC sweeps, defined in text.

Models, metrics and known limits →

05 · How we make claims

Trust before scale.

Numbers come from runs

Results, the assistant's answers and downloadable files all trace back to a simulation you can reopen.

Partial stays partial

If a sweep fails to converge, completed points are kept and the run is marked partial — it can never become an optimiser's best design.

No unearned comparisons

We don't claim accuracy or speed advantages over established TCAD tools before published benchmarks against measured devices.

84backend tests passing in CI
34frontend regression tests
2device templates plus custom decks
4runs overlaid in one comparison

Working on GaN or silicon devices?

We're opening the private beta to a small number of labs and teams. Tell us what you simulate.