How a wafer fab’s device engineers hand a full TCAD round to Micius

A round of TCAD device simulation means building the structure step by step, moving between tools, waiting for the solver and reading the results on experience. Change one parameter and the round starts again. In a project between Zhiman and a wafer fab, a device engineer can describe the round in one sentence and Micius runs it, then says plainly what the results are and aren’t good for.

At a wafer fab, a handful of engineers who know the tools well usually run device simulation with TCAD (technology computer-aided design). Take one evaluating a new superjunction power device, which depends on charge balance between alternating P and N pillars: under reverse bias they deplete each other and spread the field evenly across the drift region. The engineer builds the structure, meshes it, picks the physical models, submits the job and waits for the solver to return the potential and field. The results often go to another tool for plots, and the transfer and output curves need their own sweep across bias points. If the peak field sits in the wrong place or a curve looks off, it’s back to the parameters and the whole round starts again, for every doping level and pillar width worth trying.

Knowing whether a round can be trusted is harder still: did the solve converge, do the structure and doping make sense, does the field look right? That judgement, and the choice of what to change next, has rested with experienced engineers.

In this project the round goes to Micius, which knows device physics and how TCAD tools work. It can drive the simulation software itself: building the structure, submitting the solve and reading the results.

The engineer writes one sentence covering the device, the structural requirements, the solves to run and the bias points to sweep. Micius turns it into a parameterised structure generated directly from dimensions and doping, with no process simulation in this round, and the cross-section appears in the workspace. It calls the solver for the 2D potential and field, then runs a self-check: is the charge balanced, did the solve converge? After the check, it sweeps the transfer and output (I–V, current–voltage) curves at the requested bias points and switch the view to them. Each step and its result is written into the conversation as it happens, so the engineer can see where the round has got to.

For the superjunction MOSFET, the request asked for the blocking-state potential and field first and, if those looked right, the curves. Structure, solve, self-check and sweep followed in that order, and Micius reported round 1 done.

The results stay on the 2D section, and clicking anywhere on it reads the doping, potential and field at that point, taken from this round’s solve, with no export to another tool. Mid-way down an N pillar, the read-out also names the region; equipotential lines over the section show how the voltage spreads between the pillars. A 3D view lets the engineer turn the structure and see how the regions sit, though it shows geometry only, since the simulation is 2D.

At the end of the round, Micius summarises the device, the solve, the self-check and the current at key bias points, and states the limits without being asked. The peak field is an electrostatic screening value, not a breakdown criterion. Nothing has been calibrated against process or measured data, so the results can’t be used to sign off breakdown voltage or on-resistance. Avalanche breakdown wasn’t modelled, so breakdown voltage isn’t in the results at all, and the structure came from dimensions and doping rather than process steps. Because uncalibrated TCAD numbers can travel downstream looking like specs, these limits sit directly under the curves: the curves are for reading trends and comparing directions.

After the summary, Micius proposes directions for the next round, each stating what to change and what to look at: break the charge balance on purpose to see how sensitive the field is, run a small DOE across a range of values, or make the gate oxide thinner or thicker to see how turn-on shifts. These are the next steps that used to depend on an experienced engineer thinking of them.

The engineer picks one or writes their own request: change the structure, add a check, sweep more bias points, switch the view. If the engineer picks the first and changes only the N-pillar doping, round 2 carries every other setting over from round 1 and runs the same build, solve, self-check and sweep.

Each round’s request, structure, results, checks and limits stay together in one conversation. An engineer who comes back days later finds the last round’s conclusions and proposals where they left them, and decides from there what runs next.