AI-Based Anomaly Detection for Extreme Event Attribution: An Analysis of European Heatwaves

Formal extreme event attribution traditionally relies on large, computationally intensive climate-model ensembles, which often hampers the provision of information in real-time attribution. To overcome this, we propose a lightweight, interpretable framework that couples unsupervised anomaly detection with Bayesian deep learning enabling near-real-time attribution without statistical assumptions or costly climate model simulations. A convolutional variational autoencoder (VAE) is trained on daily 2 m temperature fields from the CMIP6 HadGEM3-GC31-LL Hist-NAT experiment (1850–2020) and early ERA5 reanalysis (1940–1980). The VAE achieves a spatial-mean correlation of ~ 0.9 over land and exhibits robust skill in discriminating warm days from non-extreme conditions across the full domain. Spatial gradients of the VAE’s reconstruction mean-squared error and Kullback–Leibler divergence yield three threshold-free metrics quantifying local departures from pre-industrial conditions. We examine their relationship with more statistical metrics like the counterfactual exceedance probability pₑₓ , which measures the chance that an event drawn from the factual climate exceeds any event expected under the counterfactual scenario. Across the training period, these metrics exhibit a strong monotonic anti-correlation with pₑₓ , demonstrating that VAE loss gradients encode meaningful dynamical–thermodynamic information. A Bayesian multi-layer perceptron calibrated solely on these metrics reproduces grid-point pₑₓ with a mean spatial correlation of 0.92, delivering exceedance probabilities alongside interquartile ranges. Case studies of four exceptional European heatwave summers (2010, 2018, 2019, and 2022) show that our framework accurately captures spatio-temporal patterns, assigns near-zero probabilities ( p̂ₑₓ<0.05 ) to record-breaking temperatures, and yields return-period estimates consistent with large-ensemble attribution studies–all without assuming a parametric distribution or using global-mean temperature predictors. This novel approach provides an efficient tool for operational attribution of extreme events under a changing climate by quantifying changes in the probability of occurrence on daily maps, which would offer valuable insights for the development of proactive regional planning and effective adaptive strategies. Motivated by the growing demand for actionable climate-risk information, the proposed protocol-ready framework supports decision-making by delivering quantitative, uncertainty-aware estimates of changing current risk levels. The graphical abstract presents a novel methodology for attributing extreme weather events to anthropogenic climate change using Variational Autoencoder-based anomaly detection (AI-Attribution Metrics). Three attribution metrics ( ξ _∇ LKLD , ξ _∇ LMSE , and ξ Total ), derived from the spatial gradients of the VAE loss functions ( LKLD and LMSE ), are introduced. These metrics exhibit a strong anti-correlation with the counterfactual exceedance probability ( pₑₓ ), indicating that higher metric values correspond to a lower likelihood that an event from the factual climate exceeds any event under the counterfactual scenario, and vice versa. A Bayesian multi-layer perceptron (MLP) is then employed to estimate the parameters of a beta distribution and predict p̂ₑₓ across spatial grids. The framework is applied to four major European heatwave events (2010, 2018, 2019, and 2022), successfully capturing spatio-temporal patterns, assigning near-zero probabilities ( p̂ₑₓ < 0.05 ) to record-breaking temperatures, and producing return-period estimates aligned with large-ensemble studies, all without assuming parametric distributions or relying on global-mean temperature predictors. The paper introduces a lightweight and interpretable framework combining unsupervised anomaly detection with Bayesian deep learning to perform near-real-time attribution of extreme climate events, reducing the reliance on computationally heavy climate-model ensembles. A convolutional VAE is trained on historical temperature data to detect anomalies via reconstruction errors and Kullback–Leibler divergence, yielding three threshold free metrics that quantify local deviations from pre-industrial climate conditions. The VAE-derived metrics show a strong monotonic anti-correlation with the counterfactual exceedance probability pex, with |r| > 0.85 and the Maximal Information Coefficient (MIC) ≈ 1, indicating that they encode robust dynamical and thermodynamical information. Applied to four major European heatwaves (2010, 2018, 2019, 2022), the framework accurately reproduces spatio-temporal heatwave patterns, estimates near-zero probabilities for record-breaking events, and aligns with results from large ensembles, without relying on global-mean temperature predictors or parametric assumptions.