CLINT

• Post-processing of numerical weather prediction (NWP) data.
• Machine learning-based weather prediction.
• Hybrid model that integrates physical simulations with machine learning or statistical techniques.

The CLINTDD model consists of two components: (1) an ML data-driven forecasting core trained on paleoclimate data, and (2) a hybrid ensemble component based on the ECMWF Extended-Range Ensemble Forecast. We initialise the data-driven component, while the dynamical ensemble is acquired from ECMWF and processed for calibration providing physically consistent forecasts used later for hybrid post-processing. The data-driven model was first trained on the PMIP6 Past2k paleoclimate simulation (MPI-ESM1.2-LR, 0–1850 CE) to identify the most relevant predictor variables and time-lag relationships controlling surface temperature and precipitation variability. An (evolutionary) optimisation algorithm was used to identify predictor combinations which provide the optimal seasonal forecast skill in the model world. Once the optimal configuration was determined in the model world, the same predictors in the ERA5 reanalysis are used for operational forecasts. For each forecast start date ERA5 provides the initial predictor fields describing the atmospheric, oceanic, and land-surface state (such as 2 m temperature, mean sea-level pressure, 500 hPa geopotential height, soil moisture, sea-surface temperature, sea-ice concentration, and others). These variables are converted to anomalies relative to ERA5 climatology, and spatially clustered using an enhanced k-means algorithm (which includes weighting by geographical distance). Lagged weekly averages of these clustered predictors (typically 1–12 weeks prior to the start date) form the initial condition vector of the AI model. Lagged weekly averages of these clustered predictors (typically 1–12 weeks before the start date) form the initial condition vector of the AI model, following the lag structure optimised from the paleoclimate training.

The CLINTDD model relies on three main datasets for real-time operation: - ERA5 near-real-time reanalysis, providing the most recent atmospheric, land, and ocean predictors used to initialise each weekly forecast. - Past2k paleoclimate simulation (MPI-ESM1.2-LR), used at each forecast cycle to perform feature selection and identify the most relevant predictors and lag relationships. - ECMWF Extended-Range Ensemble Forecasts, which supply the operational ensemble members up to 45 days ahead and are post-processed through the hybrid AI framework.

The CLINTDD data-driven model is trained entirely on the Past2k paleoclimate simulation (MPI-ESM1.2-LR, years 0–1850 CE). This dataset provides a long, stationary record used to optimise the feature-selection algorithm and train the machine-learning models linking lagged predictors to temperature and precipitation variability. The ERA5 reanalysis is not used for training but to define the spatial clusters of predictors and to apply the trained model for real-world forecasts.

The CLINTDD model combines a data-driven forecasting framework with a hybrid ensemble post-processing step. Its architecture consists of four main stages: 1 - Dimensionality reduction – Predictor fields (e.g. temperature, soil moisture, mean sea-level pressure, 500 hPa geopotential height, etc) are spatially clustered using an enhanced k-means algorithm to reduce input dimensionality while retaining key regional variability. 2- Feature selection (FS) – An evolutionary algorithm (Probabilistic Coral Reef Optimization) identifies the optimal combination of clustered predictors, domains, and time-lags that maximise predictive skill. This optimisation is performed on the Past2k paleoclimate simulation and repeated for each forecast date. 3- Forecasting model – The selected predictors are used as inputs to machine-learning regressors (e.g. Random Forest, LightGBM, AdaBoost) that predict weekly mean temperature and precipitation anomalies. 4 - Hybrid ensemble layer – The ECMWF Extended-Range Ensemble Forecasts are first calibrated against ERA5 and then post-processed using the AI data driven forecast as a first-guess to select ensemble members, improving both deterministic and probabilistic forecast skill.

The data-driven forecast framework applied to seasonal forecast of european heatwaves was published: McAdam, R., Pérez-Aracil, J., Squintu, A., Peláez-Rodríguez, C., Hansen, F., Torralba, V., Loukos, H., Zorita, E., Giuliani, M., Cavicchia, L., Salcedo-Sanz, S., & Scoccimarro, E. (2025). Feature selection for data-driven seasonal forecasts of European heatwaves. Communications Earth & Environment, 6(1), 842. https://doi.org/10.1038/s43247-025-02863-4 The driver detection framework was first applied to heatwaves in: Pérez-Aracil, J., et al. Identifying key drivers of heatwaves: A novel spatio-temporal framework for extreme event detection. Weather and Climate Extremes. https://doi.org/10.1016/j.wace.2025.100792

Validation of the subseasonal version of the CLINTDD model is ongoing. The framework follows the same approach used for the seasonal prototype: cross-validation against ERA5 reanalysis as an independent reference after training on Past2k, using deterministic (correlation, N-RMSE) and probabilistic (CRPSS) metrics. Additional benchmarking of the hybrid ensemble configuration is being carried out within the AI-Weather Quest competition, comparing forecasts to ECMWF operational outputs under the same evaluation protocol.

The main challenge was the computational cost of the feature-selection (FS) algorithm, which must be rerun for each forecast date. To address this, code optimisation, memory management, and batch parallelisation were implemented, reducing execution time and making the method suitable for operational use.

Yes. The subseasonal implementation is still experimental, while most validation work has so far focused on the seasonal counterpart. The operational domain is global, and we are addressing several methodological and computational aspects during the AI-Weather Quest competition — including optimisation of the feature-selection algorithm, refinement of lag and cluster choices across regions, and improvement of probabilistic calibration for weekly forecasts. As a small team, we are carrying out this development in parallel with our other activities, progressively integrating these advances into the operational workflow.

The main innovation is the integration of a paleoclimate-trained feature-selection framework (Past2k) into an operational subseasonal forecasting workflow. This approach allows the model to identify physically consistent predictors and time-lags that remain robust under different climate regimes. As a result, this work provides two key outcomes: skillful and computationally-economic forecasts, and identification of drivers (crucial to scientific understanding of process studied here e.g. heatwaves) Additional innovations include the hybrid ensemble post-processing guided by AI forecasts, and the global, cloud-based implementation that makes large-scale feature selection and inference feasible within the weekly competition schedule.

The CLINTDD model was developed through a collaboration between The Climate Data Factory (TCDF), CMCC Foundation, and the Universidad de Alcalá (UAH) in the context of the H2020 CLINT project. Harilaos Loukos (TCDF) led the design of the hybrid architecture, coordinated the integration of the data-driven and ensemble components, and supervised the operational implementation for the AI-Weather Quest competition. Ronan McAdam (CMCC) developed the data-driven forecasting framework and led the scientific validation of the method. Jorge Pérez-Aracil (UAH) and colleagues developed the feature-selection algorithm (PCRO-SL) and the initial version of the driver-detection framework. Thomas Noël (TCDF) implemented the cloud infrastructure, automated workflows, and operational deployment. Guilhem Vignal (TCDF) handled code development and optimisation, data integration, and data-processing routines of the data-driven framework.