Heatwaves at kilometre-scale: choosing the right months for training reduces simulation costs by a factor of four

Using CASPER, researchers have shown that the error in a downscaling model depends on the climatic discrepancy with its training data far more than on the volume of that data.

Adapting cities to heat waves requires temperature, humidity and wind data to the nearest kilometre. Downscaling networks can produce them, but training them requires high-resolution simulations that are sometimes more expensive than what they save. A team including David Rolnick sought to quantify the real need with CASPER, a U-Net that goes from 32 km reanalyses to a 1 km grid, tested in 24 configurations ranging from one to eight months of data.

The result is a straight line: the error over the periods not seen increases linearly with the climatological distance from the training data (RMSE = 0.83 + 2.95 d). This distance explains 90% of the variance of the error, compared to 7% for the simple volume of data, and makes it possible to predict in advance the performance over an unprecedented month. Over extreme summer weeks, CASPER retains the fine structure and physical coherence between variables, and sticks to weather stations to within 1.8 K during documented heat waves.

Transpose to another city

The model degrades when it is moved geographically, but eleven days of local simulation are enough to reduce the error in Vancouver from 3.8 to 1.3 K. The authors' conclusion: by covering the target climate rather than accumulating months, we obtain the same precision for four times less simulation, which puts these maps within the reach of teams without supercomputers.