Interpretable Deep Learning for Probabilistic MJO Prediction. Issue 16 (27th August 2022)

• Record Type: Journal Article
• Authors: Delaunay, Antoine; Christensen, Hannah M.
• Is Part Of: Geophysical research letters. Volume 49:Issue 16(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-08-27
• Subjects: Madden‐Julian oscillation -- predictability -- deep learning -- XAI -- MJO; Geophysics -- Periodicals; Planets -- Periodicals; Lunar geology -- Periodicals; 550
• DOI: 10.1029/2022GL098566
• Languages: English
• ISSNs: 0094-8276

Abstract

Abstract: The Madden‐Julian oscillation (MJO) is the dominant source of sub‐seasonal variability in the tropics. It consists of an Eastward moving region of enhanced convection coupled to changes in zonal winds. It is not possible to predict the precise evolution of the MJO, so sub‐seasonal forecasts are generally probabilistic. We present a deep convolutional neural network (CNN) that produces skilful state‐dependent probabilistic MJO forecasts. Importantly, the CNN's forecast uncertainty varies depending on the instantaneous predictability of the MJO. The CNN accounts for intrinsic chaotic uncertainty by predicting the standard deviation about the mean, and model uncertainty using Monte‐Carlo dropout. Interpretation of the CNN mean forecasts highlights known MJO mechanisms, providing confidence in the model. Interpretation of forecast uncertainty indicates mechanisms governing MJO predictability. In particular, we find an initially stronger MJO signal is associated with more uncertainty, and that MJO predictability is affected by the state of the Walker Circulation. Plain Language Summary: The Madden‐Julian oscillation (MJO) is an important tropical climate phenomenon. It consists of enhanced convective thunderstorms and anomalous winds that propagate eastward along the Equator for a few weeks. The MJO is difficult to predict and exhibits great variability. This means that forecasts are often probabilistic. However, current models have difficulty in correctly predicting the uncertainty in the forecast based on the current conditions. In this paper, we propose a model using neural networks capable of making reliable probabilistic forecasts. We interpret the behavior of the algorithm to verify its consistency with the known physical mechanisms of the MJO and to highlight new physical conditions that affect MJO prediction uncertainty. Key Points: A deep convolutional neural network (CNN) is used to produce probabilistic forecasts of the Madden‐Julian oscillation (MJO) The forecasts provide well‐calibrated state‐dependent estimates of forecast uncertainty The CNN forecasts are used to probe sources of predictability for the MJO