Fractal Semiconductor

Semiconductor fabs don't know what will happen in the process unless they run physical wafers through their machines. That is the reason why yield ramp and root-cause failure analysis take so much time. We have interviewed 100+ engineers at the leading fabs and OEMs, and while they are attempting to use AI, those solutions are inadequate because they are purely empirical, not structured, and thus unreliable in a production environment. We are starting by building differentiable simulations, which don't exist today and will allow AI to "understand" manufacturing from first principles. Our end product is a predictive model: predict what happens on the wafer given machine inputs, and diagnose root cause given wafer output - increasing tool uptime and recovering hundreds of millions in revenue per year.

Semiconductor fabs take 1-5 years to bring a process up to yield and up to months of downtime to root-cause drift/failure in multi-million $ machines. Each unprocessed wafer costs them up to $20K in lost revenue.