Case study — Industrial coatings
Reducing batch failure rate from 12% to 2.1%
Three years of batch records, a gradient boosting model, and one overlooked raw material interaction — how a coatings manufacturer cut rework costs by 83%.
12% → 2.1%
Batch failure rate
83%
Reduction in rework
£340k
Annual cost saving
6 weeks
From data to findings
The challenge
The client manufactured a range of industrial protective coatings for the construction and infrastructure sectors. Over an 18-month period, their batch failure rate had climbed from around 4% to 12% — a level that was generating approximately £340,000 per year in rework, scrap, and delayed deliveries.
The quality team had investigated the obvious suspects: raw material lot changes, equipment maintenance records, operator changes, and seasonal temperature variation. None of the one-variable-at-a-time analyses had produced a clear explanation. The failures appeared to be random.
They were not random. The cause was a two-way interaction between raw material moisture content and mixing temperature — a combination that conventional analysis is structurally unable to detect.
The data
The client had three years of batch records stored across two systems: a LIMS containing raw material lot data and in-process measurements, and a separate spreadsheet-based system for process parameters and final quality results.
After structuring and merging the two datasets, we had 847 complete batch records with 34 candidate variables: 12 raw material properties, 14 process parameters, and 8 in-process measurements. The quality outcome was binary: pass or fail against the viscosity, adhesion, and gloss specifications.
Dataset
Methodology
Data audit and structuring
We merged the LIMS and spreadsheet data, resolved inconsistencies in lot numbering, and imputed the 3.2% of missing values using median imputation within raw material supplier groups. Batches with more than 20% missing values were excluded, leaving 821 complete records.
Exploratory analysis
Before modelling, we ran univariate analysis on all 34 variables. No single variable showed a strong association with batch failure — the highest individual AUC was 0.61 for raw material moisture content. This confirmed that the cause was likely an interaction rather than a main effect.
Gradient boosting model
We trained a gradient boosting classifier (XGBoost) on 80% of the data, with 5-fold cross-validation for hyperparameter tuning. The final model achieved an AUC of 0.91 on the held-out test set — strong predictive performance that indicated a real, learnable signal in the data.
SHAP analysis
We used SHAP (SHapley Additive exPlanations) values to decompose the model's predictions and identify the variables and interactions driving batch failures. The SHAP interaction plot revealed a strong interaction between raw material moisture content and mixing temperature — the combination that conventional analysis had missed.
Key findings
The SHAP analysis identified a critical interaction: when the moisture content of the primary resin exceeded 0.8% and the mixing temperature was below 42°C, the probability of batch failure increased by a factor of 6.3. Neither variable alone was predictive — the effect only appeared in combination.
The interaction explained the apparent randomness of the failures. Moisture content varied with raw material lot and storage conditions; mixing temperature varied with ambient temperature and the time elapsed since the previous batch. The two variables were independently variable, and their dangerous combination occurred unpredictably.
Secondary findings included a weaker but significant effect of mixing time at high moisture content, and a protective effect of a specific pigment dispersant at elevated temperatures. These were incorporated into the revised process specification.
When resin moisture content exceeded 0.8% and mixing temperature was below 42°C, batch failure probability increased by a factor of 6.3.
Implementation
Tightened incoming material specification
The moisture content specification for the primary resin was tightened from ≤1.5% to ≤0.7%. The supplier was notified and agreed to the revised specification without a price increase, as the tighter limit was achievable within their existing process.
Minimum mixing temperature control
A minimum mixing temperature of 45°C was added to the batch manufacturing record. For batches produced in winter months, a pre-heating step was added to ensure the mixing vessel reached temperature before charging.
In-process moisture monitoring
A rapid moisture analyser was added to the incoming goods inspection process. Lots with moisture content between 0.5% and 0.7% are flagged for elevated mixing temperature (minimum 50°C) as an additional control.
Outcomes
In the six months following implementation, the batch failure rate fell from 12% to 2.1% — a reduction of 83%. The remaining failures were attributable to a separate cause (a specific pigment lot with anomalous particle size distribution) that was identified through ongoing monitoring.
The annualised cost saving from reduced rework and scrap was £340,000. The total project cost, including data structuring, analysis, and implementation support, was recovered within the first two months of operation.
The client subsequently extended the monitoring programme to two additional product lines, where similar analyses identified process improvements with a combined estimated saving of £180,000 per year.
Similar challenges?
If you have batch records and a failure rate you cannot explain, we can help. Most analyses take four to six weeks from data receipt to findings.
Get in touch