AI reconstruction of European weather from the Euro-Atlantic regimes

TL;DR

This work introduces lightweight AI models that reconstruct European surface temperature and precipitation anomalies from Euro-Atlantic Weather Regimes (WR) indices, exploiting non-linear WR-to-ground-variable relationships. Using ERA5 data for WR computation and SEAS5 as a forecast benchmark, the study demonstrates strong winter skills for temperature and notable, though less robust, precipitation performance, with four WR indices offering a practical balance between accuracy and forecastability. The models outperform simple WR composites and remain robust to input errors, and in a hybrid NWP-AI setup with SEAS5 WR forecasts, achieve comparable or superior winter temperature and summer precipitation skill in a forecasting context. The results suggest WR-based anomaly reconstruction as a viable pathway for sub-seasonal to seasonal European forecasting, with clear avenues for probabilistic extension and broader regional adaptation.

Abstract

We present a non-linear AI-model designed to reconstruct monthly mean anomalies of the European temperature and precipitation based on the Euro-Atlantic Weather regimes (WR) indices. WR represent recurrent, quasi-stationary, and persistent states of the atmospheric circulation that exert considerable influence over the European weather, therefore offering an opportunity for sub-seasonal to seasonal forecasting. While much research has focused on studying the correlation and impacts of the WR on European weather, the estimation of ground-level climate variables, such as temperature and precipitation, from Euro-Atlantic WR remains largely unexplored and is currently limited to linear methods. The presented AI model can capture and introduce complex non-linearities in the relation between the WR indices, describing the state of the Euro-Atlantic atmospheric circulation and the corresponding surface temperature and precipitation anomalies in Europe. We discuss the AI-model performance in reconstructing the monthly mean two-meter temperature and total precipitation anomalies in the European winter and summer, also varying the number of WR used to describe the monthly atmospheric circulation. We assess the impact of errors on the WR indices in the reconstruction and show that a mean absolute relative error below 80% yields improved seasonal reconstruction compared to the ECMWF operational seasonal forecast system, SEAS5. As a demonstration of practical applicability, we evaluate the model using WR indices predicted by SEAS5, finding slightly better or comparable skill relative to the SEAS5 forecast itself. Our findings demonstrate that WR-based anomaly reconstruction, powered by AI tools, offers a promising pathway for sub-seasonal and seasonal forecasting.

Figures (12)

• Figure 1: Cluster mean of the $Z500$ field (black isocontours) and corresponding average $Z500$ anomalies in winter (December, January, February), associated with the seven year-round Euro-Atlantic WR. The average is performed in the 1981-2010 climatology. The "No Regime" refers to the mean atmospheric circulation when there is no WR attribution according to the index life-cycle criteria grams_balancing_2017.
• Figure 2: Schematic representation of the model's architecture. The input vector $x$, which includes the $n$WR indices, the year, and the encoded month, is passed as input to the model. After each layer, the output $y$ is passed to a ReLU activation function $\max(0,y)$. The $N = 28$ dimensional vector, output of the second linear ResNet block, is stacked pixel-by-pixel onto the static data (latitude and longitude grids, land-sea mask, and digital terrain model). The resulting $(N + 4) \times n_\lambda \times n_\phi$ tensor, where $n_\phi$ and $n_\lambda$ are the number of latitude and longitude grid points, respectively, is the input of the first convolutional ResNet layer. The number of features is then reduced by two by the second ResNet block and to one (temperature or precipitation field) by the final ResNet layer
• Figure 3: Qualitative comparison of reconstructed monthly mean two-meter temperature (left column) and total precipitation (right column) against ERA5. The top row displays the cases with the lowest mean squared error, while the bottom row shows those with the highest. The best (worst) reconstruction for temperature corresponds to July 2016 (February 2011), while for precipitation it is August 2013 (November 2019).
• Figure 4: AI models spatial MAE (top), ACC (middle) and CE (bottom) in reconstructing the winters (first and third columns) and summers (second and fourth columns) mean two-meter temperature and total precipitation anomalies using seven WR indices in the test years (from January 2011 to =1sp 2024 . Hatched areas mark regions with correlations that are not significant at $p > 0.05$. Black contours indicate regions with ACC (CE) higher than 0.5 (0.3).
• Figure 5: Relative improvement of the AI models with respect to the regressions on individual WR as benchmark. First (second) [third] row shows the relative improvement in MAE (ACC) [CE] for both two-meter temperature and total precipitation separately for the winter (December, January, February) and summer (June, July, August) from 2011 to =1sp 2024 . Solid (dashed) contours indicate regions with relative improvement higher (lower) than 30%.