Predictability of Storms in an Idealized Climate Revealed by Machine Learning

TL;DR

This work addresses the limits of midlatitude storm predictability by training a CNN on over 200,000 cyclone tracks from a 200-year aquaplanet GCM to forecast 42-hour growth and displacements with probabilistic uncertainty (μ and σ^2). The approach links background flow features—notably baroclinicity and jet meandering—to forecast skill, showing that growth is less predictable than track and that stronger baroclinicity reduces growth predictability while aiding Δy predictability. Explainable AI (gradient-based sensitivity) localizes uncertainty amplification to downstream jet structure, effectively doubling the predicted uncertainty sensitivity in more meandering jets. Overall, the study demonstrates that ML can diagnose and quantify fundamental predictability mechanisms in an idealized climate context, offering a path toward physics-informed improvements in storm forecasts and emphasizing the need to validate these insights in more realistic settings.

Abstract

The midlatitude climate and weather are shaped by storms, yet the factors governing their predictability remain insufficiently understood. Here, we use a Convolutional Neural Network (CNN) to predict and quantify uncertainty in the intensity growth and trajectory of over 200,000 storms simulated in a 200-year aquaplanet GCM. This idealized framework provides a controlled climate background for isolating factors that govern predictability. Results show that storm intensity is less predictable than trajectory. Strong baroclinicity accelerates storm intensification and reduces its predictability, consistent with theory. Crucially, enhanced jet meanders further degrade forecast skill, revealing a synoptic source of uncertainty. Using sensitivity maps from explainable AI, we find that the error growth rate is nearly doubled by the more meandering structure. These findings highlight the potential of machine learning for advancing understanding of predictability and its governing mechanisms.

Figures (12)

• Figure 1: A schematic diagram illustrating the input and output structure of the neural network. The model takes atmospheric conditions around the storm at genesis as input and predicts storm growth, $\Delta x$ and $\Delta y$ (defined in Section \ref{['sec:machine']}) up to 42 hours ahead in 6-hour intervals.
• Figure 2: (a) Mean squared error (MSE) normalized by $42^{\mathrm{nd}}$-hour variance of the CNN-predicted $\Delta x$, $\Delta y$ and growth . (b) Joint probability density function (PDF; shading) of the predicted $\boldsymbol{\mu}$ and $\boldsymbol{\sigma}^2$ (defined in Equation 1) from the machine learning models for the $42^{\mathrm{nd}}$-hour growth. The thin black contour encloses 70% of the maximum density, and the thick contour indicates the peak density within each bin of predicted $\boldsymbol{\mu}$. The two boxes denote regions used to separate good and poor predictions. Contours are plotted at intervals of 0.03 from 0.0 to 0.3. (c) Vertical structure of the sensitivity of the prediction to perturbations in each input variable $\mathcal{S}(\mathbf{x})$ calculated for growth at 42$^{nd}$ hour, averaged horizontally for each variable and pressure level.
• Figure 3:
• Figure 4: Composites of 275 hPa zonal wind fields at storm initialization for (a) well and (d) poorly predicted storms located south of 40$^\circ$ N; (b) and (e) are as (a) and (d), but for storms north of 40$^\circ$ N. The number of samples used for the composites are: (a) 6824; (b) 8677; (d) 7406; and (e) 6353. (c) Horizontal distrubution of $\mathcal{E}(\mathbf{\chi})$ (equation 4), the derivative of uncertainty in the 42-hour growth to the initial zonal-wind field (unit: 10$^{-3}$), shown for the good prediction set, the sensitivity vector is quantified as the vertical-column mean of the gradient components.; (f) Same as (c), but for the poor prediction set.
• Figure S2: Comparison of MSE among the three neural-network models predicting storms in this study. The figure shows how the prediction error evolves over 400 training epochs, illustrating the skill achieved by each model. MSE is computed and averaged across all variables, each normalized by its mean and standard deviation (Table S1).