Plastics & Masterbatch

Mine Colours Raised FTR from 68% to 93% in Masterbatch Production with ML

Mine Colours produced masterbatch with an annual capacity of 4,000+ tons, but FTR stood at just 68% with 2.3 correction rounds per batch. Corius's XGBoost model predicts color deviation (ΔE) with MAE=0.18 precision, raising FTR to 93%, FPY to 91%, and delivering full return on investment in 5 months.

Problem

Due to the complexity of formulation parameters, an average of 2.3 correction rounds were required per batch; low FTR (First Time Right) rates led to scrap, energy waste, and delivery delays.

Solution

Corius developed an XGBoost prediction model combining historical production data and spectrophotometer measurements; the model optimizes pigment ratios, temperature, and mixing parameters to predict correct first-time production.

Result

FTR rose from 68% to 93%, FPY from 74% to 91%, annual scrap cost dropped by 38%, full ROI in 5 months.

93% FTR (First Time Right) Starting point: 68% (+25 points)
91% FPY (First Pass Yield) Starting point: 74% (+17 points)
-38% Annual scrap cost Raw material + energy savings
MAE 0.18 ΔE prediction error 5× below the tolerance threshold
11× faster Formulation time From 4 hours to 22 minutes
ROI in 5 months Return on investment From scrap and rework costs

What Operational Challenges Was Mine Colours Facing?

Mine Colours produces masterbatch concentrates that add color and functionality to plastic products. Each batch depends on the precise combination of pigment type and ratio, carrier resin selection, mixing temperature, and duration. Even a minor deviation can push the color delta (ΔE) out of tolerance, causing the entire batch to be rejected. The company was managing this complexity through the intuitive decisions of experienced formulators; however, dependence on specialist personnel was constraining production speed and prolonging new product development times. Averaging 2.3 correction rounds per batch led to both raw material and energy waste, and — most critically — frequent missed customer delivery deadlines.

  • FTR (First Time Right) rate at just 68%; one in every three batches required at least one correction round.
  • FPY (First Pass Yield) at 74%, well below the industry average of 85%+.
  • Color tolerance measurement (ΔE < 1.0) was performed manually with a spectrophotometer at the end of production; errors could only be detected after the entire batch was complete.
  • When switching to a new pigment supplier, formulators relied on intuition rather than historical data; adaptation took 3–6 weeks.
  • Scrap rate had reached 11% of volume; rework cost accounted for a significant line item in the annual budget.

How Did Corius Solve These Challenges?

Corius developed a machine learning model that predicts ΔE by combining Mine Colours's 3 years of production records and spectrophotometer measurements. The XGBoost-based model takes pigment type and concentration, carrier resin, extruder temperature profile, and mixing duration as inputs to forecast color deviation before a batch begins, and provides the formulator with optimal parameter recommendations. The model was embedded into the production line via REST API integration with the existing MES (Manufacturing Execution System) infrastructure.

01
Data Consolidation and Exploration
Month 1

Production logs, spectrophotometer measurements, and raw material certificates from different sources were consolidated into a single database. Key variables affecting quality deviation were identified.

  • Extraction and consolidation of 3 years (2022–2024) of batch records from ERP and MES systems
  • Digitization of spectrophotometer measurements (L*, a*, b*, ΔE) and mapping to batch records
  • Integration of pigment supplier, lot number, and purity certificate data
  • Missing value analysis and interpolation; 420 invalid records cleaned
  • Identification of the 14 most critical features via correlation matrix
02
Model Development and Validation
Months 2–3

ΔE prediction was performed using an XGBoost regression model. After hyperparameter optimization and cross-validation, a test set result of MAE = 0.18 ΔE units was achieved — well below the tolerance threshold of ΔE < 1.0.

  • Baseline comparison: Linear Regression, Random Forest, XGBoost, LightGBM
  • XGBoost hyperparameter optimization: 300 trials with Optuna
  • Train / validation / test split: 70/15/15, lot-based leakage-proof separation
  • Feature importance analysis with SHAP values — top 5 variables identified
  • Field validation round with 30 batches under real production conditions
03
MES Integration and Go-Live
Month 4

The model was integrated into the production planning interface. Formulators can now enter batch parameters and instantly see the predicted ΔE and optimal formulation recommendation; the model is continuously updated with feedback.

  • Model service layer with FastAPI; integration with existing MES API
  • Formulator interface: parameter input → ΔE prediction + recommendation panel
  • Automatic addition of new batch data to the model update pipeline (MLflow)
  • Team training and 2-week shadow mode testing
Technologies Used
Python (scikit-learn) XGBoost SHAP Optuna MLflow FastAPI Pandas / NumPy PostgreSQL

How Was the Model Used on the Production Line?

New Pigment Supplier Adaptation

When switching to a new pigment supplier with different lot purity values, the model uses the purity-concentration relationship learned from historical data to automatically suggest an initial formulation for the new supplier.

Supplier adaptation time Reduced from 6 weeks to 3 days
FTR on first pilot batch 87% — with zero prior data

In-Line ΔE Early Warning

Extruder sensor data is monitored in real time; when the model detects a temperature or torque anomaly, it predicts that the batch will exceed color tolerance before completion and alerts the formulator.

At-risk batches caught early 8 batches saved in the first month
Scrap volume prevented Estimated 3.2 tons of raw material

Formulation Optimization Recommendations

When target color values (L*, a*, b*) are entered, the model recommends pigment blend ratios and process parameters that balance the lowest cost with the highest FTR probability.

Average formulation time Reduced from 4 hours to 22 minutes
Raw material cost optimization Avg. 6% savings per batch

Faster Customer Color Approval Process

During the color sample approval process presented to customers, the model pre-calculates the producibility score and potential deviation range of the requested color, minimizing revisions.

Average approval rounds Dropped from 2.3 to 1.1
Customer satisfaction score NPS 48 → 71

What Results Did the ML Model Produce?

93% FTR (First Time Right) Starting point: 68% (+25 points)
91% FPY (First Pass Yield) Starting point: 74% (+17 points)
-38% Annual scrap cost Raw material + energy savings
MAE 0.18 ΔE prediction error 5× below the tolerance threshold
11× faster Formulation time From 4 hours to 22 minutes
ROI in 5 months Return on investment From scrap and rework costs

We now catch quality deviations on the production line before they occur — that's revolutionary for us. Our annual scrap costs have dropped noticeably and our delivery reliability has improved.

Shall we improve your production quality metrics with data?

Request a free analysis for machine learning solutions that optimize your FTR, FPY, and scrap rates.