Support portal Careers Contact Investors
EN
EN · English JA · 日本語 KO · 한국어 ZH · 简体中文
Request a Demo
Request a Demo
Request a Demo
Support portal Careers Contact Investors
Language EN JA KO ZH

Surrogate modeling: fast predictive models of process and device behavior

You need process and device answers faster than a full simulation can return them, so Silvaco gives you surrogate models, built from validated multi-physics simulation, that answer in real time.

Talk to our Technical Experts Read the documentation

What surrogate modeling changes for your team

Surrogate modeling trains a fast model on results an expensive simulation has already produced. Ask it a process or device question and it answers in real time, inside the range it learned. The TCAD runs still happen. They happen up front, to build the model, rather than every time someone needs an answer.

That changes who gets to ask. If you're a process, integration or design engineer, you can see how a process choice moves a structure or a device. You don't run TCAD yourself, or wait for someone who can. A surrogate is never better than the physics it learned from. That's why a Silvaco surrogate starts from TCAD results your team has already validated.

Where Silvaco fits

Building a surrogate and using one are usually two different people's jobs.

Using a surrogate

For process, integration and design engineers who don't run TCAD. A finished model needs no FTCO workflow, TCAD license or TCAD expertise to query. Your TCAD team hands it over as an XML model card.

Building a surrogate

Inside the FTCO workflow, run by your TCAD team. Victory DoE plans which simulations to run so they cover your parameter space. Victory Process and Victory Device run them. Victory Analytics trains the model on the results, with neural networks or Gaussian process regression, and SHAP-based feature selection keeps the inputs that matter. Fab measurements can go into the training data alongside simulation, so the model reflects your line.

See how FTCO builds a surrogate

Three kinds of surrogate model

Each one answers a different kind of question.

  • Structure surrogate

    Built from process simulation results. Change a process parameter and the device geometry redraws as you watch, so you follow how the structure evolves without running the process simulation again.

  • Parametric surrogate

    Models device metrics against the process parameters that drive them. Use it to find which parameters matter, or turn the question around and ask which process values reach your target.

  • Curve surrogate

    Reproduces complete I–V curves in real time, so the whole characteristic comes back with every change.

    For circuit-level surrogates, go to Utmost IV and compact model extraction.

  • Creation versus use

    Creation is specialist work, done once for each question. Use is open to anyone, as often as they need it. The line between the two is also the model's limit. Inside the range the TCAD runs sampled, you can rely on the answer. Outside it, the question goes back to your TCAD team for a new set of runs.

Applications

What engineers use a finished surrogate for.

Evidence and resources

Want to see a surrogate built from start to finish? Start with the August webinar.

More on surrogate modeling

Every surrogate starts with a question someone keeps asking

Which process or device question does your team ask every week, and who waits for the answer? Our engineers will tell you whether a surrogate fits it and what building one would take.

Talk to our Technical Experts Documentation Support and Downloads

Questions about surrogate modeling

A surrogate model is a fast approximation of an expensive simulation, trained on results that simulation has already produced. In semiconductor work, you adjust a process parameter and it returns the structure, device metrics or full I–V curve the simulation would have given. The answer comes back in real time, but only within the range it was trained on.

A digital twin is a physics-based counterpart of your process, trained on simulation and fab data, that holds what your team knows about that line. A surrogate is narrower: a fast model of one simulated behavior, such as a structure or an I–V curve. The two work together, and a finished surrogate can also be used on its own.

A TCAD engineer creates it by planning a set of simulations, running them and fitting a model to what comes back. Process, integration and design engineers use it. They query the finished model to explore process parameters and find the values that hit a target. No simulation runs on their side, and nobody waits on a specialist.

Inside the range its training runs covered, a surrogate tracks the simulation it learned from. Outside that range it still returns answers, but they're extrapolations, and any systematic error in the training data carries straight through. Check it against measured silicon before you rely on it.

No. Creating a surrogate takes TCAD and a trained specialist. Using a finished one takes neither. Your TCAD team builds and validates the model, then shares it as a model card, an XML file that engineers without TCAD can query in real time.

A designed set of simulations that samples the parameter range you care about, with the geometric and electrical results of every run. Coverage matters more than the fitting method. Adding measured fab data, and calibrating the simulations to silicon first, is what lets you trust the model against real devices.

Related

All solutions