How we built predictive models that anticipate weight, density, granulometry, and moisture — and delivered an operational simulator capable of forecasting reprocessing outcomes before they happen.
The challenge was to turn critical variables from Amorim Cork Solutions' production process into reliable predictions. To do that, we mapped two data universes — the industrial history from SCADA and the lab samples from ACC Labs — and precisely defined which models would be needed for each variable and each measurement point along the line.
The solution covers 14 distinct product-variable combinations, using Linear Regression models for weight, density, and moisture, and Logistic Regression for granulometry — the only categorical variable in the process. Each model was designed around a clear business objective: predict REP 1 output before the end of the shift.
| Predicted Variable | Model Type | Data Source |
|---|---|---|
| Weight (MLI 1/2, 0.25/0.5, 0.5/1) | Linear Regression | SCADA |
| Weight Biomass · Excess | Linear Regression | SCADA |
| Density (MLI 1/2, 0.5/1) | Linear Regression | ACC Labs |
| Moisture (MLI 1/2, 0.5/1) | Linear Regression | ACC Labs |
| Granulometry (MLI 0.5/1) | Logistic Regression | ACC Labs |
* Full methodology walkthrough available in Chapter 1 of the video.
Every data point consumed by the models has a documented origin and an auditable transformation. We built a full lineage graph — from raw factory data all the way to the model in production — running on Palantir Foundry infrastructure.
| Source System | Data Consumed | Status |
|---|---|---|
| Industrial SCADA | Process sensors · 25M+ rows | Automated |
| ACC Labs | Lab samples per shift | Automated |
| Shift Emails | OCR via Gemini · Screen data and notes | Automated |
| MES / EPC | Batch types and input moisture | Automated |
* Architecture discreetly structured on Palantir Foundry infrastructure.
With clean, unified data in place, the platform delivers three layers of intelligence: detailed evaluation of each predictive model, investigation of process hypotheses with statistical validation, and a shift-by-shift operational view that connects what happened on the factory floor to what the models predict.
Each model is evaluated with error distributions, Predictions vs. Actuals, and baseline comparisons — across separate train and test sets to ensure real-world generalization.
The platform investigates and documents hypotheses such as the impact of raw material moisture and the influence of origin (China vs. other batches) on the stability and quality of the final product.
Scale evolution over time, setpoints vs. actual mill speeds, and Rotex/screens used each shift — all accessible through a single date selector.
The Simulator is where analytical intelligence becomes an operational tool. The plant manager configures shift parameters — raw material type, mill settings, and maintenance history — and receives, within seconds, a production forecast and expected final product specifications, complete with confidence intervals and failure probability per variable.
Raw material, scale, and equipment parameters all adjustable from a single screen.
Active ML models deliver real-time estimates for every input combination configured.
Every prediction includes a lower and upper prediction interval, quantifying model uncertainty.
Failure probability per product flags spec deviations before they occur on the production line.
Amorim Cork Solutions no longer operates in the dark. Every shift starts with a quality forecast for the final product — and operators can simulate decisions before making them. If your industrial operation deals with untracked process variability or unpredictable output quality — this problem already has a solution. We can do the same for you.
