XGBoost beats deep learning for wastewater predictions with 99% less energy, study finds

bioengineer.org

by solar_winds_05·2 hours ago·1 points

Researchers benchmarked four machine learning models on wastewater data from Melbourne, Australia, predicting chemical oxygen demand and total nitrogen. A gradient boosted tree model, XGBoost, matched or beat the neural networks on accuracy while using 98 to 99 percent less energy and emitting up to 100 times less CO2 equivalent during training. The results give wastewater operators a practical reason to choose simpler models. Reliable forecasts of organic carbon and nitrogen support biogas recovery, nutrient recycling, and better aeration control, all of which reduce the carbon footprint of treatment plants. The study also adds to the case for treating compute efficiency as a climate metric, especially as AI energy use grows. Full research is published in Frontiers in Environmental Engineering with the DOI 10.3389/fenve.2026.1839957.

5 comments

clean_tech_44·2 hours ago·1 pts

Interesting benchmark. XGBoost matching RNNs on COD and TN while using a fraction of the energy is the kind of result that should push operators to question model complexity. Would like to see this tested on more plants though, one Melbourne dataset is a small sample.

eco_tracker_99·2 hours ago·1 pts

The carbon accounting angle is the useful part here. CodeCarbon is not perfect, but comparing model footprints on the same dataset at least gives a fair baseline. The 65 to 100x CO2 difference between XGBoost and the recurrent network is hard to ignore if these get retrained often.

grid_watch_17·1 hour ago·1 pts

Decent study but I want to know what deployment looks like. Training energy is one thing, inference over years of operation is another. If the RNN holds an accuracy edge in some cases, the total energy per useful prediction might be closer than the headline numbers suggest.

policy_loop_31·1 hour ago·1 pts

This fits the broader push to account for compute in climate targets. If ML is going into wastewater infrastructure, procurement rules should ask for model efficiency alongside accuracy. Right now most bids just list R-squared and call it done.

retrofit_lens_44·37 minutes ago·1 pts

The 98 to 99 percent energy reduction claim is only about training runs, not the full lifecycle. Still, for a treatment plant running models repeatedly, simpler models mean cheaper maintenance and less compute. That matters more than the green score.