ML Models

XGBoost · CatBoost · Isolation Forest

Yield prediction and equipment anomaly detection trained on 12.5K wafer records and 100K sensor readings from BigQuery.

Yield — XGBoost

MAE0.0196

RMSE0.0247

R²0.7961

Train time0.454s

Yield — CatBoostwinner

MAE0.0189

RMSE0.0238

R²0.8112

Train time1.042s

Anomaly — Isolation Forest

Trained on labeled simulated anomalies — clean separation by design.

Precision0.9730

Recall0.9864

F10.9797

ROC AUC0.9999

Yield — XGBoost

Loading from BigQuery…

Yield — CatBoost

Loading from BigQuery…

Anomaly — Isolation Forest

Loading from BigQuery…

Live Prediction — Yield

Adjust process parameters and get real-time yield prediction from CatBoost or XGBoost.

Feature Importance (CatBoost)

Product node dominates — but that is a process-maturity proxy, not actionable insight. To turn this into root-cause signal, replace product with step-level features (etch time, litho dose, CMP pressure, anneal temp).

Feature Importance (CatBoost)

Loading from BigQuery…

Feature Importance (XGBoost, one-hot encoded)

XGBoost requires one-hot encoding, so categorical levels appear as separate features.

Feature Importance (XGBoost)

Loading from BigQuery…

Detected Anomalies (latest 12)

Isolation Forest flags with ground-truth comparison — demonstrating validation with labeled data.

Equip	Type	Temp °C	Press	Vib g	Score	Truth
EQ-005	Metrology	29.1	0.555	0.156	0.611	match
EQ-018	Etch	237.9	1.408	0.761	0.647	match
EQ-029	Metrology	26.5	0.617	0.168	0.600	match
EQ-022	IonImplant	28.8	0.000	0.559	0.599	match
EQ-013	Lithography	30.3	0.583	0.160	0.598	match
EQ-008	CVD	559.9	2.550	0.641	0.699	match

Detected Anomalies (latest 12)

Loading from BigQuery…

Models trained on simulated data. Production use requires domain-specific calibration and ongoing drift monitoring (training-serving skew, data drift, scheduled retraining).