Closed-Loop Quality Intelligence: From Messy Data to a Production Fix
The Challenge
A precision-casting foundry had a recurring, sporadic microstructure defect that had resisted explanation for years. Investigations cycled through the usual suspects — furnace lining, gas absorption, equipment wear — and never produced a stable root cause. Meanwhile the defect drove scrap, and the only quality gate was a subjective visual test that couldn’t be trusted (see Project CERES).
The deeper problem: the evidence lived in five disconnected systems — spectrometer chemistry, melt and pour records, MES production data, warehouse movements, and shake-out quality outcomes. No single view existed to even ask “what actually predicts this defect?”
The Approach
This was a decision-science problem, not a modelling contest. The loop was integrate → diagnose → measure → intervene → standardize.
Integrate the evidence
Roughly 2,000 production charges were linked across all five source systems — incoming-material chemistry tied to the specific charges it ended up in, joined to process sensors, temperatures, and final QC outcomes.
Find the cause, not the correlation
Instead of chasing a predictive model, the analysis used dose-response stratification, stratified contingency tests, and chemistry-adjusted logistic regression — and, crucially, designed experiments. A controlled doping trial and an overnight “bath-memory” natural experiment (re-pouring the same melt the next day) isolated causation that observational data alone could not.
The finding: the defect was driven by a conjunction — elevated incoming-supplier iron and an adverse furnace-atmosphere state, acting together as a strong interaction effect. A nine-element screen confirmed iron as the single chemical axis. The defect was actually over-refinement via iron-phosphide nucleation — the opposite of what its name implied.
Close the loop
The findings became operator-facing decisions: a risk calculator (decision-tree, AUC 0.787), a zero-CAPEX pour-temperature and batch-sequencing policy, an incoming-material specification change, and quantitative acceptance criteria that replaced the subjective visual test.
Technical Stack
- Data integration: Python (Pandas), SQL warehouse extraction, multi-key entity resolution across 5 systems
- Causal analysis: dose-response stratification, stratified χ²/Fisher tests, adjusted logistic regression, Design of Experiments
- Measurement: SEM/EDX particle analysis; objective grain measurement via CERES; Gage R&R
- Delivery: interactive risk calculator + structured technical and management reporting
Outcome
- Operations adopted the pour-temperature and sequencing policy with immediate effect and zero capital cost.
- A process fix went live in production, with early quality stabilization, and supplier-iron limits were written into specification.
- A subjective, not-capable inspection was replaced with quantitative, auditable acceptance criteria.
- Behaviour changed across production, quality, supply chain, and suppliers — the hardest part of any quality improvement.
The value here wasn’t a clever algorithm. It was closing the full loop — from fragmented plant data to a measured, adopted change on the shop floor.