From Reactive to Predictive: Applying ML to Silicon Yield Data

For years, post-silicon validation worked the same way: run production, collect data, wait for something to go wrong, then scramble to find the root cause. It is an expensive cycle. By the time you detect a yield excursion, you have already shipped bad lots, wasted test time, and burned through engineering hours on firefighting. I wanted to flip that model. Instead of reacting to failures, what if we could predict them? The Dataset Nobody Was Using Our UltraFlex test systems generate enormous parametric datasets. Every device that passes through test gets hundreds of measurements: leakage currents, timing margins, voltage thresholds, frequency response. Most of this data was being used for simple pass/fail binning and then archived. But hidden in those measurements are early warning signals. A parametric value might technically pass spec but drift toward the limit over time. A subtle correlation between two seemingly unrelated parameters might indicate a process shift at the fab. These patterns are invisible in standard yield reports but obvious to a well-trained model. Building the Feature Set The first challenge was feature engineering. Raw parametric data is noisy and high-dimensional. I worked with the test engineers to identify which parameters were physically meaningful indicators of silicon health versus measurement noise. We built aggregated features at the wafer-lot level: distribution statistics (mean, sigma, skew, kurtosis) for key parametric groups, inter-parameter correlation coefficients, and trend features capturing how each parameter evolved over recent production lots. The Model We chose gradient-boosted trees (XGBoost) for their ability to handle mixed feature types and capture non-linear interactions without extensive preprocessing. The target variable was binary: whether a lot would experience a yield drop greater than 3 sigma within the next 5 production lots. Training on 18 months of historical data, the model achieved 87% recall on yield excursion prediction with a false positive rate under 12%. That means we catch the vast majority of incoming problems with a manageable number of false alarms. The Impact The model now runs as a batch job after each production lot completes. When it flags a potential excursion, the validation team gets an automated alert with the specific parameters driving the prediction. They can then investigate proactively, adjusting test limits or escalating to the fab before the problem compounds. Detection time went from days (waiting for enough failed lots to trigger a statistical alarm) to hours (the model flags risk after the first anomalous lot). That translates directly to fewer scrapped devices and lower cost of quality. The key insight: the data was always there. We just were not asking the right questions of it.