The coding screen is the easy part. The domain round is where semiconductor DS offers are decided, and no standard prep covers it.
YieldOps is the proving ground for applied machine learning in manufacturing. Built for the round that comes after LeetCode.
You do not know which ones break on fab data and why. XGBoost with a global StandardScaler destroys the PM-cycle signal. Random Forest with MDI importance overrates Chamber ID by design. K-fold on time-series fab data hides the covariate shift that kills models at deployment.
Closed by: Mission Track + Wafer Journey
You have not debugged a SECS/GEM timestamp desync. An exact JOIN on FDC and MES data returns zero rows. A join without a 2-second tolerance can silently match today's etch step to a sensor reading from three days ago, corrupting your training data at the rows that matter most.
Closed by: Mission Track + Round 1 Gauntlet
Your model runs the night shift. When a CUSUM alarm fires at 2 AM, it is a process engineer who gets the page, not you. But it is your threshold logic that decided whether to fire it. The academic answer to a drift signal is to retrain. The hired answer is to route the alarm to maintenance, because drift in a fab almost always means a consumable is degrading, and retraining teaches the model that a worn-out chamber is the new normal.
Closed by: Wafer Journey + Simulations
You have never set a threshold based on wafer cost. In a fab where yield is 99%, a model that predicts Pass on every wafer achieves 99% accuracy. The threshold is not 0.5. It comes from a cost curve: a false negative ships a $50,000 killer defect to a customer. A false positive costs 30 minutes of metrology time.
Closed by: Interview Prep + Mission Track