Published Date
Author
Category
Machine learning is genuinely useful in climate work, and the useful applications look nothing like the ones that get announced. They are narrow, unglamorous and mostly about scheduling.
Where it earns its place
- Forecasting wind and solar output well enough to reduce the reserve a grid has to hold
- Predictive maintenance on turbines and transformers, where a failure avoided is generation kept
- Optimising building heating and cooling against occupancy and price rather than a fixed schedule
- Detecting methane leaks in satellite imagery at a rate no inspection programme could match
Where the claims outrun the evidence
Materials discovery, fusion control and climate modelling all attract confident predictions. The research is real; the timelines attached to it in press coverage generally are not, and treating them as planning assumptions has a cost.
The footprint on the other side of the ledger
Training and running these models consumes electricity, and data-centre demand is now material in several grids. The honest accounting is net: an application that saves more than it spends is worth having, and one that cannot show it does is worth questioning.
That is a solvable arithmetic problem rather than an objection in principle — but it is arithmetic that ought to be shown.
