Zaninelli et al. (2026) AI-Based Anomaly Detection for Extreme Event Attribution: An Analysis of European Heatwaves

Identification

• Journal: Earth Systems and Environment
• Year: 2026
• Date: 2026-03-31
• Authors: Pablo G. Zaninelli, David Barriopedro, Jorge Pérez-Aracil, Marie Drouard
• DOI: 10.1007/s41748-026-01122-6

Research Groups

• Departamento de Matemática e Informática Aplicadas a las Ingenierías Civil y Naval, Escuela Técnica Superior de Ingenieros de Caminos, Canales y Puertos, Universidad Politécnica de Madrid, Spain
• Instituto de Geociencias (IGEO), Consejo Superior de Investigaciones Científicas - Universidad Complutense de Madrid (CSIC - UCM), Spain
• Department of Signal Processing and Communications, Universidad de Alcalá, Spain

Short Summary

This study introduces a lightweight, interpretable AI-based framework combining unsupervised anomaly detection with Bayesian deep learning for near-real-time extreme event attribution. It successfully quantifies the probability of European heatwaves under pre-industrial conditions, demonstrating consistency with traditional methods without relying on computationally intensive climate model ensembles.

Objective

• To develop a lightweight, interpretable AI-based framework for near-real-time extreme event attribution, coupling unsupervised anomaly detection with Bayesian deep learning.
• To define threshold-free attribution metrics derived from Variational Autoencoder (VAE) loss functions and link them to counterfactual exceedance probability ((p_{ex})).
• To validate the framework's performance and apply it to attribute European heatwaves, demonstrating its ability to quantify changes in event likelihood without computationally intensive climate model ensembles or parametric assumptions.

Study Configuration

• Spatial Scale: Europe (\left[ 22^{\circ }W-45^{\circ }E;\; 27^{\circ }N-72^{\circ }N \right]), interpolated to a (2^{\circ } \times 2^{\circ }) grid.
• Temporal Scale:
  • CMIP6 HadGEM3-GC31-LL Hist-NAT experiment: 1850–2020 (VAE training data).
  • ERA5 reanalysis: 1940–1980 (VAE training data, "Old World" baseline), 1940-2023 (evaluation).
  • Factual fields (Hist extended with SSP2-4.5 scenario): 1850-2020.
  • Daily 2 m temperature fields for extended summer (May to September).
  • Case studies: European heatwave summers (2010, 2018, 2019, 2022).

Methodology and Data

• Models used:
  • Variational Autoencoder (VAE): Convolutional 2D filter layers, latent space of 100 dimensions, trained with Adam optimizer.
  • Bayesian Multi-Layer Perceptron (BMLP): Single hidden layer with 50 units.
• Data sources:
  • Daily 2 m temperature fields from the CMIP6 HadGEM3-GC31-LL Hist-NAT experiment (1850–2020) for counterfactual climate training.
  • Daily 2 m temperature fields from ERA5 reanalysis (1940–1980 for training, 1940-2023 for evaluation) for counterfactual/Old World climate training.
  • Daily 2 m temperature fields from CMIP6 HadGEM3-GC31-LL Hist simulation extended with SSP2-4.5 scenario (1850-2020) for factual conditions.

Main Results

• The VAE achieved a spatial-mean correlation of approximately 0.9 over land for temperature reconstruction and demonstrated robust skill in discriminating warm days from non-extreme conditions.
• Three threshold-free metrics ((\xi {\nabla L{\text {KLD}}}), (\xi {\nabla L{\text {MSE}}}), and (\xi _{\text {Total}})), derived from the spatial gradients of the VAE loss functions, were introduced to quantify local departures from pre-industrial conditions.
• These VAE-derived metrics exhibited a strong monotonic anti-correlation with the counterfactual exceedance probability ((p_{ex})) across the training period (Pearson correlation > 0.85, Maximal Information Coefficient (\approx 1)).
• A Bayesian Multi-Layer Perceptron (BMLP), calibrated solely on these metrics, reproduced grid-point (p_{ex}) 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) showed that the framework accurately captured spatio-temporal patterns, assigned near-zero probabilities ((\hat{p}_{ex} < 0.05)) to record-breaking temperatures, and yielded return-period estimates consistent with large-ensemble attribution studies.
• An uncertainty decomposition revealed that approximately 61% of the total predictive uncertainty is aleatoric, 20% is epistemic (linked to BMLP weights), and 18% is propagated from the VAE stage.

Contributions

• Introduces a novel, lightweight, and interpretable AI-based framework (VAE-AD-BMLP) for near-real-time extreme event attribution, significantly reducing reliance on computationally intensive climate model ensembles.
• Develops threshold-free attribution metrics directly from the spatial gradients of VAE loss functions, providing physically meaningful and interpretable measures of departure from counterfactual climate conditions.
• Establishes a transparent and robust link between ML-derived anomaly diagnostics and classical probabilistic attribution quantities ((p_{ex})) without requiring parametric assumptions for event distributions or global-mean temperature predictors.
• Demonstrates operational applicability and generalization across different extreme events and regions, offering quantitative, uncertainty-aware estimates of changing current risk levels for proactive regional planning and adaptive strategies.
• Provides an unconditional (probabilistic) attribution framework that quantifies changes in the probability of events of a given magnitude, independent of their specific physical drivers, which is particularly relevant for adaptation planning.

Funding

• EU-funded H2020 project CLINT (Grant Agreement No. 101003876).
• Open Access funding provided thanks to the CRUE-CSIC agreement with Springer Nature.

Citation

@article{Zaninelli2026AIBased,
  author = {Zaninelli, Pablo G. and Barriopedro, David and Pérez-Aracil, Jorge and Drouard, Marie},
  title = {AI-Based Anomaly Detection for Extreme Event Attribution: An Analysis of European Heatwaves},
  journal = {Earth Systems and Environment},
  year = {2026},
  doi = {10.1007/s41748-026-01122-6},
  url = {https://doi.org/10.1007/s41748-026-01122-6}
}