[Paper Review] AIFS -- ECMWF's data-driven forecasting system

Simon Lang, Mihai Alexe+14 more|arXiv (Cornell University)|Jun 3, 2024

Forecasting Techniques and Applications63 citations

TL;DR

AIFS is ECMWF’s data-driven weather forecasting system using a graph neural network encoder-decoder and a sliding window transformer processor trained on ERA5 and ECMWF analyses; it runs four times daily alongside the physics-based NWP model and shows competitive forecast skill versus IFS and ERA5 for upper-air, surface, and tropical cyclone forecasts.

ABSTRACT

Machine learning-based weather forecasting models have quickly emerged as a promising methodology for accurate medium-range global weather forecasting. Here, we introduce the Artificial Intelligence Forecasting System (AIFS), a data driven forecast model developed by the European Centre for Medium-Range Weather Forecasts (ECMWF). AIFS is based on a graph neural network (GNN) encoder and decoder, and a sliding window transformer processor, and is trained on ECMWF's ERA5 re-analysis and ECMWF's operational numerical weather prediction (NWP) analyses. It has a flexible and modular design and supports several levels of parallelism to enable training on high-resolution input data. AIFS forecast skill is assessed by comparing its forecasts to NWP analyses and direct observational data. We show that AIFS produces highly skilled forecasts for upper-air variables, surface weather parameters and tropical cyclone tracks. AIFS is run four times daily alongside ECMWF's physics-based NWP model and forecasts are available to the public under ECMWF's open data policy.

Motivation & Objective

• Motivate the development of data-driven, operational weather forecasting systems as alternatives or complements to physics-based NWP.
• Describe the architecture and training regime of AIFS, including data sources, graph neural networks, and transformer-based processing.
• Assess forecast skill of AIFS relative to ECMWF IFS analyses and ERA5 observations across upper-air and surface variables and tropical cyclones.
• Demonstrate scalability and parallelism enabling high-resolution data handling and potential probabilistic ensemble capabilities.

Proposed method

• Encoder-processor-decoder design that maps ERA5 input onto a processor grid via a cut-off radius, then projects latent states back to the ERA5 grid.
• GNN blocks with multi-head graph transformer convolution for encoder/decoder and a pre-norm transformer with shifted window attention for the processor.
• Processor grid based on an octahedral reduced Gaussian grid (N320) with around 1° resolution; eight learnable features added to each node and edges.
• Activation checkpointing and sequence/tensor parallelism to enable training on high-resolution data and mixed-precision training.
• End-to-end training with auto-regressive rollout, pre-training on ERA5, rollout training, and fine-tuning on operational IFS analyses; loss is area-weighted MSE with per-variable scaling.

Experimental results

Research questions

• RQ1 How does AIFS's data-driven forecast quality compare to ECMWF’s IFS analyses and ERA5 observations across lead times up to 10 days?
• RQ2 What architectural choices (GNN encoder/decoder, transformer processor, reduced Gaussian grids) enable scalable, high-resolution data-driven forecasting?
• RQ3 Can data-parallel and tensor-parallel training regimes support large ensembles and probabilistic forecasting within operational-like workflows?
• RQ4 What are the strengths and limitations of AIFS in upper-air versus surface fields and tropical cyclone tracking?
• RQ5 What refinements (loss scaling, fine-tuning strategy, probabilistic objectives) could further improve forecast skill?

Key findings

• AIFS delivers highly skilled forecasts for upper-air and surface variables and tropical cyclone tracks, with competitive ACC improvements over IFS in several metrics.
• Forecast skill advantage over IFS is observed for longer lead-times, with substantial improvements in tropospheric forecasts up to about 100 hPa when verified against analyses and radiosondes.
• Surface forecasts (e.g., 2 m temperature, 10 m wind) show consistent improvements over IFS when verified against analyses and observations.
• AIFS tends to smooth forecast fields with lead time, a behavior linked to the area-weighted MSE loss used during training; probabilistic objectives are being explored to mitigate this.
• Compared to GraphCast, AIFS performs similarly, with differences varying over time and depending on training/fine-tuning approaches.
• Tropical cyclone track forecasts are notably better with AIFS than IFS, primarily due to reduced slow bias in TC propagation speed, though TC intensity can be underestimated relative to IFS.