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

Fab Technology Co-Optimization™: your process as a digital twin

FTCO takes the output of your process and device simulations, trains it against your fab's measured data, and gives fab engineers a model they can question in real time.

Talk to an FTCO expert

What fab technology co-optimization means

  • One model for process, device and circuit

    Process, device and circuit usually get tuned one after another. Each team hands the next a fixed result, so a trade-off that crosses those lines shows up late. Fab technology co-optimization puts all three in one model. An implant dose that helps the transistor but costs you at the circuit level surfaces while you can still change it.

  • What the digital twin is

    It's a trained, non-linear model of your manufacturing process. What it knows comes from simulation runs and measurements from your fab, fed in over repeated rounds through Victory DoE and Victory Analytics. Put device and circuit simulations into the training as well, and it maps how process parameters move device and circuit results without anyone writing those links by hand.

  • What it does that a simulation run can't

    Ask it a question and the answer comes back while you're still at the tool. A TCAD deck answers one question per run and needs someone who writes decks. The twin answers in real time, through an interface built for engineers who aren't simulation specialists. Fewer process decisions wait on a physical wafer learning cycle.

    And it doesn't replace TCAD. Every answer comes from physics your TCAD engineers already validated.

The FTCO workflow, from simulation to digital twin

Simulation and fab measurements become training data. You design and run an experiment set, train a model on the results, then start asking it questions.

  1. Simulate and measure Process, device and circuit runs from the Victory TCAD simulators, next to measured data from your fab.
  2. Design the experiment Victory DoE makes any TCAD input a variable and lays out the experiment set.
  3. Run and collect The set runs in simulation, not on wafers, spread across your cluster through LSF if you have one. Results land in a single table.
  4. Train the model Victory Analytics fits the model and scores how well it holds.
  5. Query the twin Query it for design targets and yield, in real time.

Calibrating the digital twin against measured silicon

Simulation gives you a physically consistent answer. Your fab gives you the true one. Calibration is how the twin learns the gap between them.

  • Fitting simulation to silicon

    FTCO trains on simulated and measured process data together. The fit follows what your process actually did, not only what the simulator expected. Victory Analytics cleans the measured side first, flagging outliers and filling in missing values.

  • What you need to bring

    Production and measured data from the process you want to model, with the process conditions behind it, plus the TCAD simulations of that same flow. How much you need depends on how many parameters you vary and how non-linear the response is, so bring what you have to an FTCO expert for an answer sized to your process.

Process optimization with the digital twin, in real time

Four ways to put a trained twin to work. Each answers in real time, where the same question asked of a simulator means another run.

  • Design targeting

    Start from the design value you need. The twin returns the process parameter values that hit it, so you don't sweep the parameter space until something lands.

    • Need a particular threshold voltage? The twin works back to the doping doses that produce it.

  • Process capability analysis

    See what your real variation does to yield before the line shows you. Monte Carlo and Cp/Cpk studies run against the twin, not on production wafers.

    • Oxide thickness drifting from lot to lot? The twin shows the spread that drift puts into your process and device metrics.

  • Virtual experimentation

    Find out which process parameters actually move the design target you care about. The twin screens process and device variables and highlights the ones that matter.

    • On one transistor, the screen picks out implant dose as a strong lever on threshold voltage and a negligible one on breakdown voltage.

  • Interactive structure modeling

    Change a process parameter and the structure view redraws right away. What you see comes from physics-based training data, not interpolated pictures.

    • The etch profile, for instance, reshapes as you move the etch parameters.

How a surrogate model is built, and where it stops

A surrogate is the trained model on its own: a fast stand-in for a simulation you no longer need to rerun.

  • Built from a designed experiment

    Victory DoE samples the parameter space with a designed set of simulation runs and gathers the results in one table. Victory Analytics then fits a neural network or a Gaussian process regression model, whichever suits the data.

    Put the effort into the experiment set. Good coverage makes an ordinary fit useful, and no fitting method recovers a region you never sampled.

  • What you do with it afterward

    Evaluating a surrogate costs a function call, not a TCAD run. That makes real-time design targeting and Monte Carlo yield studies practical. Both need far more evaluations than anyone would queue as simulations. And a finished surrogate reaches engineers who don't run simulations themselves.

  • Where its limits are

    A surrogate is dependable inside the region its experiment set covered. Outside it, the answers lose their footing. Knowing the range you sampled is part of owning the model.

    It also inherits any systematic error in its training data. Calibrate against measured silicon before you lean on it.

Surrogate modeling, in depth

The tools behind FTCO

Two tools do the machine-learning work, and the Victory TCAD simulators supply the physics they learn from. All of it sits in Silvaco's AI-Driven Manufacturing portfolio.

  • Victory Analytics

    Machine-learning analysis built for semiconductor research, development and manufacturing. It trains the twin and runs the optimization on it.

    Explore Victory Analytics

  • Victory DoE

    Designs the experiments and keeps every simulation project's results in one place, ready for Victory Analytics.

    Explore Victory DoE

  • The Victory TCAD simulators

    Victory Process and Victory Device generate the process and device data the twin learns from, and Victory Visual lets you inspect every result in 2D and 3D.

    Explore Silvaco TCAD Explore Victory Visual

Built with Micron Technology for memory devices

FTCO took shape through a collaboration with Micron Technology, Inc., around the process questions memory development raises.

  • Memory Digital twin April 2024

    Micron Technology

    Micron's engineers use an FTCO digital twin to speed design and pathfinding on memory products. The twin learns from Micron's production data and from physics-based simulation of etching, deposition and mechanical stress. Engineers outside the TCAD team get process, design and yield optimization in their own hands.

    “Silvaco’s AI and digital twin solution is enabling us to accelerate our groundbreaking advancements in memory and storage.”

    Dr. Gurtej Sandhu, Principal Fellow of Technology Pathfinding, Micron Technology, Inc.

    Read the Micron partnership announcement

FTCO resources

Conference talks and webinars on fab technology co-optimization, plus a SEMI interview on where it came from.

See all webinars Browse the SURGE 2025 archive

Talk to an FTCO expert

Tell us which process you want a digital twin of and what data you already have on it. Measurements from split conditions and the parameters you're trading off are a good place to start. An FTCO expert will get back to you on what it takes for your process.

Use your work email address.

* Required

Send to an FTCO expert

Questions about FTCO

FTCO stands for Fab Technology Co-Optimization. It trains a non-linear model of your fab's process on simulation results and measured data until the model works as a digital twin. Fab and process engineers then query it in real time instead of waiting for a TCAD specialist to run a new simulation.

It's a trained model that stands in for a real process. It learns from simulation and fab measurements, then predicts how a process change affects the structure, the device and yield. Engineers use it to try a change in software before any physical wafer goes through it.

Victory DoE runs a designed set of simulations across the parameter space. Victory Analytics fits a neural network or Gaussian process regression to the results, then scores the fit. Accuracy depends on how well that experiment set covers the region you ask about, and on calibration against measured data.

The twin is trained on measured fab data alongside simulation results, not on simulation alone. That way its predictions track what your line actually produced. Victory Analytics screens the measured data before training, catching outliers and imputing gaps so a few bad readings don't skew the fit.

A TCAD engineer typically sets up the simulations the digital twin learns from, using Victory DoE and Victory Analytics, and trains it against your fab data. Once it's trained, process and fab engineers who don't run TCAD themselves query it directly, through an interface built for non-specialists.

FTCO runs on Victory Analytics for training and optimization and on Victory DoE for experiment design. Victory Visual displays the results. Victory TCAD simulators such as Victory Process and Victory Device supply the physics the digital twin learns from, and your own fab measurements supply the calibration.

Related