AIFS

• Machine learning-based weather prediction.

Our forecasts are initialized from the ERA5T reanalysis using the 10 members of the ERA5 Ensemble of Data Assimilations (EDA).

For real-time forecasting, AIFSgaia uses two consecutive states (t, t–24 h) from ERA5T and its associated EDA as input. Input variables are: - temperature, humidity, geopotential height and horizonal winds at 1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, 50 hPa, - surface variables: 10m winds, 2m-temperature, mean sea level pressure, total column water vapour, - forcing variables: orography, land sea mask, insolation.

The model was trained on the ERA5 reanalysis from 1979 – 2024 with 6-hourly temporal resolution on an o96 grid (~1°). In addition to the input variables, the model predicts precipitation, cloud cover, and top of atmosphere radiation as diagnostic outputs.

AIFSgaia follows the architecture introduced in Lang et al. (2024b). It combines a transformer-based graph neural network (GNN) encoder and decoder with a sliding-window transformer processor. Attention is computed along spiral longitudinal bands (Lang et al. 2024a). The processor consists of 16 layers with an embedding dimension of 1024 and 8 attention heads, resulting in approximately 230 million trainable parameters. Forecast ensembles are generated by conditioning on random noise. For each member, independent noise samples from a standard normal distribution (dimension = latent grid × 4 noise channels) are transformed by a two-layer MLP and injected via conditional layer normalization in each processor layer. The model is trained to predict the 24-hour difference between t and t+24 h using two consecutive inputs (t, t–24 h). The loss is the fairCRPS computed over 4 ensemble members against the ERA5 target. Training is run for 200k iterations with a batch size of 16, initial learning rate 1e-3, and cosine annealing. During inference, forecasts are generated autoregressively for 28 steps, producing 200 ensemble members to represent the predictive distribution.

AIFSgaia is based on Lang et al., AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score (2024b), http://arxiv.org/abs/2412.15832

During development, the model was trained on ERA5 data up to 2018 and verified in the period 2020–2024. After fine-tuning on recent years (2018–2024), additional sanity checks were performed for March–May 2025 forecasts.

The main challenges were technical—automating the end-to-end forecast generation and submission pipeline to ensure reproducibility and timely uploads.

Current forecasts tend to be overconfident, with ensemble spread smaller than observed variability. Future work will focus on improved uncertainty calibration and probabilistic spread correction.

The full model code and utilities are openly available through the anemoi Python packages, developed as part of the broader ECMWF community effort enabling reproducible AI-based weather forecasting.

The development of AIFSgaia would have not been possible without the great support by the AIFS Team at ECMWF, covering data preparation, code development, computing architecture and deployment.

• Machine learning-based weather prediction.

AIFShera is initialized from the ERA5T reanalysis using the 10 members of the ERA5 Ensemble of Data Assimilations (EDA).

For real-time forecasting, AIFShera ingests two consecutive 6-hourly states (t, t–6 h) from ERA5T EDA. The input variables are: - temperature, humidity, geopotential height, and horizontal winds at 1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, 50 hPa, - surface variables: 10m winds, 2m-temperature, mean sea level pressure, total column water vapour, - forcing variables: orography, land sea mask, isolation.

The model was trained on the ERA5 reanalysis from 1979 – 2024 with 6-hourly temporal resolution on an o96 grid (~1°). In addition to the input variables, the model predicts precipitation, cloud cover, and top-of-atmosphere radiation as diagnostic outputs.

AIFShera follows the AIFS-CRPS architecture described by Lang et al. (2024b). It combines a transformer-based graph neural network (GNN) encoder and decoder with a sliding-window transformer processor. Attention is computed along spiral longitudinal bands (Lang et al. 2024a). The processor consists of 16 layers with an embedding dimension of 1024 and 8 attention heads, resulting in approximately 230 million trainable parameters. Forecast ensembles are generated by conditioning on random noise. For each member, independent noise samples from a standard normal distribution (dimension = latent grid × 4 noise channels) are transformed by a two-layer MLP and injected via conditional layer normalization in each processor layer. AIFShera predicts 6-hourly increments using two consecutive inputs (t, t–6 h). The model is trained with the fairCRPS loss over 4 ensemble members against ERA5 targets. Training is divided into two stages: 1. Pre-training: one-step prediction for 200k iterations (batch size = 16, initial lr = 1e-3, cosine annealing) over 1979–2017. 2. Fine-tuning: progressive autoregressive training (2–12 steps), each trained for 2k iterations on 2017–2024. At inference, the model performs 112 autoregressive steps (28 days) with 200 ensemble members to sample the predictive distribution.

Apart from minor differences in training schedule, period, and input/output variables, the model is equivalent to Lang et al., 2024b: “AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score” (arXiv:2412.15832).

During development, the model was trained on ERA5 data up to 2018 and verified in the period 2020–2024. After fine-tuning on recent years (2018–2024), additional sanity checks were performed for March–May 2025 forecasts.

The technical challenges were primarily related to automating forecast generation and submission.

Current forecasts tend to show narrower ensemble spread than observed variability. Future work will focus on improved uncertainty calibration and probabilistic spread correction.

Besides its coarser O96 grid resolution, AIFShera closely follows the operational AIFS-ENS medium-range model at ECMWF. It is encouraging to see that a model optimized for medium-range forecasting generalizes well to sub-seasonal timescales.

AIFShera was developed by the AIFS Team at ECMWF, encompassing efforts in data preparation, model architecture, code development, high-performance computing, and operational deployment.

• Machine learning-based weather prediction.

Our forecasts are initialized from two distinct reanalysis products. - The atmospheric variables are initialized from the ERA5T reanalysis using the 10 ensemble members of the ERA5 Ensemble of Data Assimilations (EDA). - The ocean variables are initialized using the ORAS6 ocean and sea-ice reanalysis. Because ORAS6 becomes available with a latency of about 8–12 days (depending on the alignment between the forecast start time and the 5-day assimilation window of the reanalysis), a separate deterministic machine-learning model was trained to bridge this short gap. This autoregressive model, which uses surface ocean and sea-ice conditions forced by the operational atmospheric analysis, provides the most accurate possible ocean initial conditions during this delay period. Unlike the atmosphere, the ocean's initial conditions are identical for all ensemble members.

For real-time forecasting, AIFSthalassa uses two consecutive states (t, t–24 h), similar to AIFSgaia and AIFShera. As described above, the atmospheric input comes from ERA5T and its associated EDA, while the ocean and sea ice input comes from an ORAS6-like model. Input variables are: - Temperature, humidity, geopotential height, and horizontal winds at 1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, 50 hPa. - Surface variables: 10m winds, 2m-temperature, mean sea level pressure, total column water vapour. - Ocean variables: sea surface temperature, sea surface height, sea surface salinity, and sea surface velocity (zonal and meridional components). - Forcing variables: orography, land-sea mask, insolation.

The model was trained on the ERA5 and ORAS6 reanalyses from 1993 – 2024 with 6-hourly temporal resolution on an o96 grid (~1°). In addition to the input variables, the model predicts precipitation and cloud cover as diagnostic outputs.

AIFSTalassa follows the AIFS-CRPS architecture described by Lang et al. (2024b) and summarized in the AIFSgaia documentation. The main differences with AIFSgaia are: • The embedding dimension of the processor is 1536 for approximately 500M trainable model parameters. • Training is run for 300k iterations. In AIFSthalassa, there is no explicit coupling between the ocean and atmospheric components. The fields are trained jointly in a single model with a common loss, at the same spatiotemporal resolution.

AIFSthalassa is based on Lang et al., AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score (2024b), http://arxiv.org/abs/2412.15832. The specific treatment of the ocean variables is not yet described in a peer-reviewed publication.

Model verification follows the same procedure as for AIFSgaia, using ERA5 data for out-of-sample evaluation and one March-May of this year for sanity checks after fine-tuning.

The technical challenges were largely identical to those encountered in AIFSgaia, primarily related to automating forecast generation and submission. A forced ML ocean model has been developed to obtain the updated ocean initial conditions that enable us to run the forecast system in real time.

Like AIFSgaia, AIFSthalassa tends to produce an ensemble spread narrower than the observed variability. Future work will focus on improving ensemble calibration.

AIFSthalassa is the first model of the AIFS family that explores a prognostic representation of the surface ocean and sea ice as sources of S2S predictability.

AIFSthalassa was developed by the AIFS Team at ECMWF, encompassing efforts in data preparation, model architecture, code development, high-performance computing, and operational deployment.