Relaxation Experiments on the Sub-seasonal Time Scale Using ML-based Weather Prediction Models

Weather predictions on the subseasonal time-scale remain a challenge. Though sources for extratropical subseasonal predicability lie in slowly evolving components of the Earth System, these sources are not fully exploited yet due to systematic errors in numerical weather prediction models. A prominent example are tropical modes of variability (e.g., the MJO and tropical waves), where short-term forecast errors may deteriorate the subseasonal forecast skill in the extratropics (Dias et al. 2021). However, it is not yet clear in which tropical regions short-term forecast errors have the strongest impact on the extratropical forecast skill. One way to identify these regions are relaxation experiments using numerical weather prediction models. However, these experiments quickly become expensive especially when being run in an ensemble configuration. In this study, we employ machine learning-based weather prediction models and conduct relaxation experiments for different regions in the tropics on the subseasonal time-scale. Probabilistic predictions are obtained by initializing models from the Ensemble of Data Assimilations (EDA) system of ECMWF. Though the ensemble predictions remain underdispersive, our study reveals that relaxation experiments in machine learning-based weather prediction models are in principle technically possible. On a case study basis, we successfully identify regions where short-term forecast errors in the tropics have the greatest impact on the extratropical subseasonal forecast skill. Further, we can show that due to the comparably low computational costs, machine-learning based models are a suitable research tool for universities with limited computational ressources where large-sets of relaxation experiments with numerical weather prediction models would be otherwise unfeasible.

Reference:

Dias J, Tulich SN, Gehne M, Kiladis GN. Tropical origins of weeks 2–4 forecast errors during the Northern Hemisphere cool season. Monthly Weather Review. 2021 Sep;149(9):2975-91.