Climate models: halving the error would require 28 times more computation, and no one is measuring the energy consumed

Having analysed 159 CMIP models since 1995, an EU roadmap quantifies the diminishing returns on computing and calls for standardised energy reports ahead of CMIP8.

The models of the Earth system form the basis of carbon neutrality policies, but their own energy consumption almost never comes into the assessment of their quality. A roadmap signed by a large team of climatologists builds on the archived outputs of 159 models covering five generations of the CMIP intercomparison project. Finding: Since 1995, the error has declined at a constant rate of 13.5% at each doubling of the computing power. Splitting the CMIP6 error by two would therefore require 28 times more calculation.

Unequal progress

Gains vary according to the variables: since CMIP1, the error in sea-level pressure was divided by 3.5 and the error on surface temperature by 2.7, but the error on precipitation decreased by only 7%. Weather models based on machine learning equalize or exceed operational forecasts for a fraction of the calculation time, but none publish a measurement of its consumption, and integrate the cost of training changes their ranking.

The authors propose a "competence-energy" framework, combining an empirical Pareto front and a ranking score, and argue for a standardized declaration of energy use and performance as soon as CMIP8 is prepared.