Physics first,
data second.
In crystal growth you cannot wait for data. A run takes weeks to months. The failure modes you most want to predict are rare by design, because the process has already been tuned to avoid them. A facility with twenty years of history may still have only a few hundred runs on record, and the useful labels are thinner than that.
Learning from observation alone does not work at that sample size. So we invert the order. We build the simulation first and use real runs to correct it.
Build the twin from physics and geometry
We start from the furnace itself. Hot zone geometry, heater configuration, insulation, crucible dimensions, material properties. That gives a thermal model of your furnace that produces sensible answers on day one, before it has seen a single one of your runs.
Generate training data from the twin
The twin can be run thousands of times across the parameter ranges you care about. Gradient, rotation, withdrawal rate, power profile. That generated set is large enough and varied enough to train a model, including regions of the parameter space you would never deliberately enter on real hardware.
Correct against real runs
Your furnace does not behave exactly like its model. As real runs accumulate, we use them to correct the twin and the model trained on it. Each run moves the simulation closer to your specific hardware. Ten runs are useful here, because they are correcting a model that already works rather than teaching one from nothing.
Test changes before committing a run
Once the twin tracks your process, you can ask it questions. What happens to the gradient if we change this heater ratio. Where does this profile put the interface shape. You get an answer in hours instead of committing months of furnace time to find out.
What this does not do
- It does not control your furnace. We do not take a control action on your hardware. The output is information you act on.
- It does not remove the need to run the furnace. The twin narrows the set of experiments worth running. Real runs still decide the answer.
- It does not arrive already knowing your process. The first version is built from physics. Accuracy against your specific hardware improves as your runs correct it.