Model Insights
Methodology and validation for the corrosion-rate prediction models. Trained 2026-08-22 on a synthetic, physics-informed pressure-vessel dataset — presented here for transparency into how a production model would be evaluated before deployment on real inspection data.
Predicted vs. Actual Corrosion Rate
Held-out test set — fleet ensemble model
Feature Importance
Fleet ensemble model
Methodology Notes
- Corrosion-rate ground truth is synthetic, generated from an API 571 damage-mechanism model (mechanism base rate x material resistance x temperature window x H2S/chloride/velocity/pH modifiers) with lognormal measurement noise — this dataset is for demonstration, not calibrated to a real Phillips 66 asset.
- t-min is computed from ASME BPVC Section VIII, Div. 1, UG-27(c)(1) thin cylindrical shell formula using per-material allowable stress with a simplified high-temperature derate.
- Remaining life follows API 510 practice: (current thickness − t-min) / long-term measured corrosion rate.
- Recommended reinspection interval follows API 510 half-life logic, capped at a 10-year internal interval; actual interval-setting should follow a documented RBI program per API 580/581.
Model Architecture
Fleet Model (offline, batch-scored)
Gradient Boosted Trees (scikit-learn GradientBoostingRegressor), 200 estimators, max depth 3. Trained on the full inspection-history feature set and used to score every vessel in the registry. Best suited for periodic fleet-wide reassessment (e.g. ahead of a turnaround planning cycle or annual RBI refresh).
What-If Model (live, client-side)
Ridge Regression (standardized features) on standardized features. Exported as plain coefficients so an engineer can drag a temperature, H2S, or velocity slider on a vessel and see an instant re-prediction — no round trip to a scoring service. Used for interactive exploration, not as the system of record.