A digital twin that finds the lowest-energy setting

The R13 refining process is extremely energy-intensive and accounts for a significant share of the mill's operating cost. The relationships between plate gap, pressure and raw material mix are non-linear, which makes manual optimisation hard, particularly during recipe or rate changes.

Without a systematic way to find the optimal operating point, energy consumption stayed high and pulp quality varied whenever raw material or production rate changed. Out-of-spec pulp accumulated during startups and recipe changes.

The process has a 12 to 25 minute delay before pulp reaches the measurement points, which makes manual adjustment reactive by definition. Optimisation depended on operator experience and intuition rather than real-time data.

AEL built a digital twin of the R13 refining process, a simulator that predicts how the process will behave before the operator makes a change.

A Sweet Spot optimiser analyses more than 476,000 combinations of control parameters and recommends the setting that minimises energy consumption within the given quality limits. All recipes now run in one shared model: it is the same machine, and a larger coherent dataset gives more reliable predictions than many small models. Accuracy and data coverage per recipe are shown directly in the simulator.

Under the hood are RandomForest models trained on the mill's own process history, using sixteen features: eight raw control parameters plus eight engineered from domain knowledge. The models handle dynamic time lags so control parameters are correctly synchronised with process outcomes 12 to 25 minutes downstream. The model is retrained on fresh production data as the mill evolves; the August 2026 retrain added the whole first half of the year, about 50 percent more data.

The interface offers two starting points, the plant's actual settings or the recipe's target values, and warns when a sensor has gone quiet, so gaps in the data are never mistaken for process deviations.

Most importantly, the twin now validates itself. A timeline-aligned view compares each prediction against the real outcome for the same material, which turns every incoming measurement into an automatic test and replaces the manual test protocol entirely. Operators file deviation reports straight from the view; the system attaches every setting and timestamp automatically.

The platform was made configurable so the same twin can be rolled out to the R14, R15 and R16 refiners with more data and configuration rather than new development. A second twin for the bleach plant, where the target is minimal chemical consumption rather than energy, is being set up on the same platform.