Skilful Global Seasonal Predictions From A Machine Learning Weather Model Trained On Reanalysis Data Chris Kent Adam A Scaife Nick J Dunstone Doug Smith Steven C Hardiman Tom Dunstan Oliver Wattmeyer by Chris Kent & Adam A. Scaife & Nick J. Dunstone & Doug Smith & Steven C. Hardiman & Tom Dunstan & Oliver Watt-meyer instant download after payment.

Machine learning weather models trained on observed atmospheric conditions can outperform1234567890():,;1234567890():,;conventional physics-based models at short- to medium-range (1–14 day) forecast timescales. Herewe take the machine learning model ACE2, trained to predict 6-hourly steps in atmospheric evolutionand which can remain stable over long forecast periods, and assess it from a seasonal forecastingperspective (1–3 month lead time). Applying persisted sea surface temperature (SST) and sea-iceanomalies centred on 1st November each year, we initialise a lagged ensemble of seasonal predictionscovering 1993/1994 to 2015/2016. Over this 23-year period there is remarkable similarity in thepatterns of predictability with a leading physics-based model. The ACE2 model exhibits skilfulpredictions of the North Atlantic Oscillation (NAO) with a correlation score of 0.47 (p = 0.02), as well as arealistic global distribution of skill and ensemble spread. Surprisingly, ACE2 is found to exhibit a signalto-noise error as seen in physics-based models, in which it is better at predicting the real world thanitself. Examining predictions of winter 2009/2010 indicates potential limitations of ACE2 in capturingextreme seasonal conditions that extend outside the training data. This study reveals that machinelearning weather models can produce skilful global seasonal predictions and provide newopportunities for increased understanding, development and generation of near-term climatepredictions.