SECOM — no

Weekly fail rate falls 23% → 1–5% over three months. 55–66% of sensors exceed the PSI action threshold in every test segment. And a walk-forward simulation shows no retraining policy beats leaving the model frozen — detecting drift is not the same as being able to fix it.

Bosch — no, seductively

A path's historical fail rate ranks first of 263 features by importance — and still degrades out-of-time ranking. Train-vs-test correlation: Spearman ρ = 0.09 (p = 0.36).

PHM CMP — yes

Error does not grow with tool wear. The most-worn quartile is the most accurate, because consumable usage is an input feature the model already learned.

Problem. 590 sensors per wafer, 6.6% fail. An escaped bad wafer costs ~100× a needless re-inspection. Rank wafers so limited re-inspection capacity lands where failures are.

SPC first. A fab engineer reaches for a control chart, so that is the bar: Hotelling T² on PCA scores catches 8% of failures at 11% false alarms. Failures are not global multivariate outliers — which is what justifies supervised ML.

Threshold from cost, never 0.5. The operating point is the minimum of the expected-cost curve, chosen on a validation window inside training.

I built the sensor correlation network with the standard fMRI functional-connectivity recipe — correlate all pairs, keep the strongest 5% of edges, rank by degree centrality — then tested whether hub sensors make better features than mutual-information top-k.

And against randomly chosen subsets of the same size.

Neither principled method beats random at any k. Both sit inside ±1 sd of the random control everywhere. Nothing beats keeping all 444 features.

Honest label: this is a data-driven correlation network, not a process flow graph. SECOM's sensors are anonymized. The process-flow story is Bosch's.

This sub-project was built around the argument that under drift, a monitor-and-retrain loop matters more than model choice. So I wrote the experiment that tests it: walk forward through the wafer stream, compare four policies at identical windows, refit preprocessing, model and threshold on every retrain.

No policy beat frozen. And that −41% on sliding is not a finding — one window with 5 failures carries 75% of it. Paired window by window: 3–4, p = 0.47.

What it did establish: the PSI trigger fires in 7 of 8 windows, so "retrain when drift is detected" is operationally the same as "retrain always" — the trigger buys no economy here.

Problem. Measuring material removal rate is slow and occupies metrology capacity, and the answer arrives after the run. If the in-process trace predicts it, most wafers skip physical measurement.

The data was supposed to be gone. phmsociety.org now 404s its own download links and the official test labels ship as ?. I recovered the trajectories from the Internet Archive and found the withheld test labels published separately — 424 (WAFER_ID, STAGE) keys matching the official test set exactly. So this is scored on the real competition test set, not a holdout I invented.

The expectation: as the polishing pad and dresser wear, prediction error grows, motivating a retraining trigger on accumulated hours.

The most-worn quartile is the most accurate. Consumable usage is an input feature, so the model learned the wear effect rather than decaying with it.

Same split, same official test set, same metric, same static inputs. Only the trajectory encoding differs: ~200 hand-crafted time and frequency descriptors, versus a network reading the raw signals.

Hand-crafted wins by 2.3×. At ~1,600 training runs, compressing a trajectory into physically-motivated descriptors beats asking a network to rediscover them.

Before claiming ML is needed, run the chart a process engineer reaches for. Hotelling T² compares a unit's measurement vector against an in-control mean and covariance — which presumes every unit is measured the same way.

With 7,927 distinct routes, most measurement slots are empty for most products. Those blanks are stations never visited, not readings that went missing. Imputing them invents in-control behaviour that never happened.

SPC's failure here is structural, not statistical. The plant does not have one process to chart. It has 7,927 — which is exactly the argument for modelling the routing explicitly.

MCC needs a threshold. At the oracle threshold — the best any of them could reach on test — the model without graph features wins, and every threshold-free metric agrees. path_risk still ranked first of 263 features by gain.

Model C injects the structure a completely different way — each product becomes a subgraph of stations it visited, so message passing can combine a reading with its neighbours' along the route. It lands with B, behind A. Two independent routes to the same answer.

Why: a path family's historical fail rate does not carry forward. Across 108 families with ≥200 products in both windows, Pearson r = 0.23, Spearman ρ = 0.09 (p = 0.36), while prevalence drifts 0.66% → 0.40%. The encoding memorised a level, not a property.

Leakage red lines

• No random splitting of timestamped data — all three datasets drift.
• No oversampling outside the CV loop.
• No statistic fit on data the model shouldn't see — imputers, scalers and target encodings are all fit on the training window.
• A wafer's stage A and stage B runs never straddle a split (asserted in code, not assumed).

No accuracy, anywhere

"All pass" scores 93% on SECOM and 99.4% on Bosch. Reported instead: PR-AUC, MCC, expected cost under an asymmetric matrix, and decile capture — because inspection capacity is the real constraint.

No magic features

Public Bosch solutions reach MCC ≈ 0.4+ largely through row-Id ordering, which leaks the response. Excluded. My numbers are what the process data supports.

Scale handled honestly

4.7 GB across ~2,100 sparse columns: read in 100k-row chunks and collapsed to per-station aggregates before anything is held at once. Missing values are routing information ("never visited this station"), preserved as visit counts and left as NaN — never imputed.

Shared, tested foundations

common/ holds metrics, splits, SPC baselines and drift monitoring, used by all three sub-projects. 20 unit tests, including one that catches an int64 overflow in the MCC threshold search at scale.

Actionable now

• Go look at station L3_S32. 4.51% fail rate, 7.8× the floor average, on 24,543 units.
• Re-examine the routing policy that lets products skip the L2 inspection stations — those routes fail at double the rate.
• Deploy CMP virtual metrology on the routine population (MAE ≈ 2% of the value) and keep physical metrology for the tails.
• Do not ship a SECOM model without the monitor. The drift trigger fires in every test segment; a rolling retrain matters more than model choice.

The transferable habit

Every one of these projects has a place where the obvious reading of the numbers was wrong:

Each cost one extra diagnostic. None of them needed a bigger model. The last one cost me my own headline claim, which I rewrote rather than defended. That is the habit I would bring to a yield team.