Demonstrating the BRIS AI Weather Model on JASMIN and Exploring UK Domain Adaptation

Artificial intelligence is rapidly reshaping the landscape of weather prediction. In this post, I demonstrate how the BRIS AI weather forecasting model, developed within ECMWF’s Anemoi framework, can be deployed on JASMIN and explore whether it can be adapted from its native Scandinavian domain to the UK.

---

Why AI Weather Models Matter

Traditional numerical weather prediction (NWP) models solve the governing equations of atmospheric dynamics explicitly, requiring substantial computational resources. In contrast, AI-based models learn mappings between atmospheric states from historical datasets such as ERA5.

Once trained, these models:

• Run significantly faster during inference
• Exploit GPU architectures efficiently
• Use larger effective timesteps

Recent systems such as GraphCast, Pangu-Weather, and FourCastNet have demonstrated competitive skill with operational NWP systems.

However, most of this progress has been at the global scale. Extending these approaches to regional forecasting, where mesoscale processes dominate, remains an open challenge.

---

The BRIS Model: Bridging Global and Regional Scales

BRIS represents a hybrid approach combining:

• Global atmospheric context
• Regional refinement via a stretched grid

It is implemented as a graph neural network (GNN):

• Nodes represent atmospheric state variables
• Edges define spatial connectivity and message passing
• Multi-scale structure enables interaction between global and regional domains

This architecture is closely aligned with ECMWF’s emerging AI Forecasting System (AIFS), making BRIS a useful prototype for regional AI forecasting within an operational ecosystem.

---

Running BRIS on JASMIN

Infrastructure Setup

The BRIS model was deployed on JASMIN using:

• NVIDIA A100 GPUs (Orchid cluster)
• Python environment with:
  • anemoi-inference
  • anemoi-datasets
  • earthkit
  • PyTorch / PyTorch Geometric

Despite non-trivial dependency management, the model ran successfully on a single GPU.

---

Out-of-the-Box Execution (Scandinavian Domain)

Running BRIS on its native domain worked seamlessly.

The workflow consisted of:

1. Preparing initial conditions
2. Loading pretrained checkpoints
3. Running inference
4. Generating forecast outputs

Outputs included:

• Near-surface temperature
• Wind fields
• Surface pressure
• Precipitation

These results confirm that the BRIS inference workflow can be executed efficiently within the JASMIN infrastructure.

---

Moving Beyond Scandinavia: UK Domain Adaptation

A central question of this work was:

Can a regional AI weather model trained in one region be applied to a different geographical domain?

---

The Core Challenge: Domain Dependence

Unlike traditional NWP models, AI weather models are strongly dependent on:

• Training climatology
• Grid configuration
• Graph connectivity

The BRIS model was trained using a Nordic-centred stretched grid, tightly coupled to:

• Specific node distributions
• Graph topology
• Regional datasets (MEPS)

Applying the model to a UK domain introduced several challenges:

• Graph–checkpoint incompatibilities
• Sensitivity to domain geometry
• Failure when grid alignment was inconsistent

This highlights a fundamental limitation: the learned representation is not easily transferable across domains without adaptation.

---

A Working Approach: Forecast-in-a-Box (FIAB)

To address these challenges, the demonstrator adopted the Forecast-in-a-Box (FIAB) workflow from Destination Earth.

This involved:

• Defining graph structures via YAML
• Constructing nodes using Anemoi graph builders
• Defining connectivity through:
  • K-nearest-neighbour edges
  • Multi-scale hidden-node interactions
• Preparing:
  • Global datasets (IFS-like)
  • Regional cutouts (~0.05°–0.1° resolution)

Two domains were explored:

1. UK-focused regional domain
2. Extended North Atlantic domain capturing upstream dynamics

---

What Worked—and What Didn’t

Technical Success

From a systems perspective:

• The inference workflow executed successfully
• Forecast outputs were generated
• Outputs could be analysed and post-processed

Scientific Limitations

However, forecast quality degraded compared to the native configuration:

• Reduced spatial coherence
• Less realistic large-scale structure
• Inconsistent evolution of meteorological fields

These results indicate that successful execution does not imply reliable forecast skill when the model is applied outside its training region.

---

Integration with CSET

BRIS outputs are generated on an unstructured grid, requiring post-processing before integration with standard analysis tools.

A lightweight Python workflow was developed to:

1. Read BRIS NetCDF outputs
2. Interpolate onto a regular latitude–longitude grid
3. Convert to Iris cubes
4. Feed into CSET plotting routines

This demonstrated that BRIS outputs can be integrated into existing analysis frameworks with minimal additional code.

---

Key Lessons

1. AI Models Are Infrastructure-Ready

BRIS can be run efficiently on JASMIN GPU infrastructure.

2. Workflow Portability Is High

The Anemoi framework enables flexible experimentation with domain configurations.

3. Generalisation Is the Bottleneck

Forecast quality degrades significantly outside the training domain, highlighting the need for domain-specific training or adaptation.

---

Next Steps

Domain-Specific Training

Fine-tuning BRIS using UK datasets (e.g. UKV) is likely required for improved performance.

Integration with Operational Pipelines

Future work could explore integration with:

• UKV analyses
• Operational forecasting systems
• FastNet datasets

Evaluation Against NWP

Systematic comparison with traditional NWP models is needed to quantify skill.

Expansion to Other Regions

The workflow developed here could support experiments over other regions, such as India.

---

Conclusion

This work demonstrates that AI weather models are operationally deployable within modern HPC environments, but also highlights a key limitation:

AI weather models remain strongly tied to their training domain.

Addressing this limitation will require:

• Domain-specific datasets
• Adaptation strategies
• Hybrid AI–physics approaches

The workflows developed here provide a foundation for future research into regional AI weather prediction.

---

Acknowledgements

I would like to thank:

• Harrison Cook and Mario Santa Cruz (ECMWF)
• Even Marius Nordhagen and Thomas Nipen (Norwegian Meteorological Institute)
• Huw Lewis and the FastNet team (Met Office)

for their technical support and collaboration throughout this work.