Role of the atmospheric circulation in the observed warming over Europe using a neural network.

Daily temperature variations over Europe are strongly linked to fluctuations in the large‐scale atmospheric circulation over the North Atlantic basin. Recently, Europe has been warming rapidly, and it is important to accurately estimate the contribution of atmospheric circulation to this trend.

Here, we present an innovative dynamical adjustment framework based on a convolutional neural network (UNET) trained on CMIP6 simulations and fine-tuned on reanalysis, to estimate the observed circulation-induced temperature at the daily timescale and the subsequent trends over 1979-2024. This approach offers robust estimators at the daily scale, and performs generally better than the commonly used methods for dynamical adjustment (e.g. analogues).

When applying this method on temperature averaged over western Europe, and using the winds at 850 hPa as the circulation predictor, we find that the temperature trends induced by the dynamics between 1979 and 2024 are of 0.05 [-0.03,0.14]°C/decade annually and greater in summer (0.08 [-0.00,0.17]°C/decade) and in winter, but with higher uncertainty (0.09 [-0.11,0.29]°C/decade).

Further, we conduct sensitivity tests to the circulation predictor. Considering the wind at 700 hPa rather than 850 hPa makes no substantial difference, but considering the SLP can increase the estimated dynamical trends up to a factor of 2. This discrepancy might be due to surface processes affecting the temperature-SLP relationship, and our findings suggest that dynamical adjustment methods can be sensitive to the predictor used.