Project VulcanOps: From Research Discovery to Production ML

This project extends the findings of Project NOSTRADAMUS. Read that case study first for the scientific background.

The Challenge

NOSTRADAMUS proved the concept: trace element interactions in lead-free alloys can be predicted before casting, and a pre-production chemistry gate can stop defects at the source. But a validated research model on one alloy family is not the same as a production system.

The gap between research and live operations involved three hard problems:

1. Scale: The foundry produces dozens of alloy families. A single model trained on one material cannot generalize without first understanding which materials share the same defect drivers.
2. Feature quality: Several variables that correlated strongly in the research dataset turned out to be proxies for process conditions rather than independent causal factors — including nitrogen content, which tracked furnace behavior rather than chemistry.
3. Operational edge cases: Cold-start scenarios (new materials with short history) and material number transitions (production history split across renamed articles) both break naive time-series approaches.

The Approach

VulcanOps introduced a three-layer architecture to address these problems systematically:

Layer 1 — Product Group Clustering

Unsupervised clustering groups alloy families by shared defect signatures. Materials within the same cluster share a model — this solves the cold-start problem for new materials that are chemically similar to existing ones, and provides the training data volume needed for reliable predictions.

Layer 2 — 3-Tier Model Hierarchy

Each prediction is generated by the most specific model available:

• Tier 1 (Article-level): Sufficient historical data for a dedicated model
• Tier 2 (Cluster-level): Pooled training from the product group
• Tier 3 (Global baseline): Conservative fallback for genuinely novel materials

The system selects the highest applicable tier automatically based on available data volume.

Layer 3 — Feature Engineering for Edge Cases

• Correlation pruning: Features with high correlation to confounded process proxies (e.g., nitrogen) were removed from the feature set after systematic validation
• Lag features: Chemistry and process parameters from the preceding charges are included as sequence context
• Succession stitching: Production history is matched across material number changes, preserving continuity that would otherwise cause the model to treat renamed articles as cold-starts

Technical Stack

• Python — Scikit-learn, Pandas, NumPy for model training and feature pipelines
• Custom ETL — Production SQL warehouse to feature matrix via validated extraction scripts
• Clustering — Unsupervised grouping of alloy families by spectral and process signatures
• Model selection logic — Automated tier assignment based on data availability thresholds

Outcome

VulcanOps extended the NOSTRADAMUS pre-production gate from a single-alloy proof-of-concept to a multi-product-group system covering the majority of the production portfolio. The 3-tier hierarchy means that even materials with limited history benefit from a principled prediction rather than a naive default.

The feature engineering work — particularly the correlation pruning and succession stitching — established a reusable methodology for any future ML work on this production dataset.