Researchers build an AI model to predict water consumption at Chilean desert copper plants

A study using real data from two copper concentrator plants in the Antofagasta region tested four machine learning models to anticipate water demand in one of the driest regions on Earth.

Researchers published, in the scientific journal Water (MDPI), a study that uses real data from two copper concentrator plants in Chile's Antofagasta region — one of the most arid regions in the world, with less than 50 mm of rain a year — to build water consumption forecasting models.

The study compared four different machine learning algorithms — Support Vector Regressor (SVR), Extreme Gradient Boost (XGBoost), an artificial neural network (ANN) and Random Forest — all trained on a full year of real operating data from the two plants. Instead of testing a single model and assuming it works, the researchers used cross-validation to compare each approach's performance and identify which operating variables most influence water consumption — the kind of methodological rigor that separates a reliable model from one that only looks good on training data.

In a region where rainfall barely reaches 50 mm a year, getting the water consumption forecast wrong isn't a technical detail — it's an operational continuity risk. Having a model that anticipates water demand from data the plant already produces, instead of relying only on historical averages, gives the process team a decision margin that didn't exist before.