A processor die in its ceramic package

The AI assistant

An engineering assistant that works through the solver.

Our assistant is Claude, by Anthropic, connected to the platform through tools. It turns a plain-language request into simulations and design searches, then reports what the solver returned.

Tools

The same controls you have.

The assistant can't reach the physics any other way. Each action is a visible tool call.

ToolWhat it does
list_devicesLists device templates and their tunable parameters, with defaults and valid ranges.
run_simulationRuns one simulation and waits for it, returning extracted metrics and a run ID.
optimize_designRuns a constrained Bayesian (TPE) search: an objective, one to three variables, optional constraints.
get_runFetches the parameters, metrics and warnings of an earlier run by ID.

Ground rules

Built not to make things up.

These rules are in the assistant's instructions, not just on this page.

  • Every number comes from a tool result.It never estimates device behaviour itself.
  • It checks before it acts.It looks up parameter names and ranges before launching a simulation.
  • It explains, briefly, with units.Why a change moved Vth, gm or Ron — in mA/mm and mS/mm.
  • It names model limits when they matter.Classical 2DEG, no self-heating, breakdown as an estimate.

Engineer

Find a gate length that maximises transconductance while keeping the breakdown estimate above 100 V.

Assistant · tool calls

list_devices()
optimize_design(device="gan_hemt", objective={metric:"gm_max", direction:"maximize"}, constraints=[{metric:"vbr_est", op:">=", value:100}], variables=[{name:"Lg", …}])
get_run(id="…")

Returns the best feasible design with its run ID, and notes that the breakdown figure is an estimate without avalanche feedback.

Illustrative. The tools, metrics and constraint format are the real ones.

Why Claude

Tool use is the whole design.

The assistant is useful only if it reliably calls the right tools with valid parameters and reports results faithfully. That is what it is built around.

Your data

What gets sent, and when.

When you use the assistant, your messages and the results it works with are sent to Anthropic to generate replies. Without the assistant, nothing leaves the platform. Details are in the privacy policy.

Example requests

Ask in plain language.

"Run the default GaN HEMT and tell me its threshold voltage and peak transconductance."

One simulation, metrics with units, a run ID

"Which gate length maximises gm while keeping the breakdown estimate above 100 V?"

A constrained design search

"Compare SiO₂ and HfO₂ in the dielectric-choice deck at Vg = 1.1 V."

Material choice as a variable

"Why did the threshold shift when I increased the Al mole fraction?"

Explanation grounded in the runs
A chip die on a black background

Boundaries

What the assistant won't do.

  • Invent a number that no simulation returned
  • Change your saved projects without you
  • Hide a partial or failed run
  • Claim accuracy the models haven't earned

It is optional. Every feature of the workbench works without it.

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.