Investigating Inductive Biases for Machine Learning Emulation of Sudden Stratospheric Warmings in Idealised Isca Simulations
Researchers have investigated how different machine learning architectures impact the emulation of sudden stratospheric warming (SSW) events. Using idealised Isca simulations, they found that while convolutional, transformer, and graph-based models performed similarly during quiet stratospheric periods, their performance diverged significantly when SSW-like variability was present. The study highlights explicit three-dimensional vertical coupling as a crucial inductive bias for accurately emulating stratospheric dynamics, though it also notes that low forecast error does not always equate to physically accurate wave-mean-flow interactions. AI
IMPACT This research could lead to more accurate weather prediction models by improving the emulation of complex atmospheric phenomena.