Merge global forecasting capabilities from `prob_model_global`

[joeloskarsson]

Global forecasting can be viewed as a special case of LAM, with no boundary conditions to consider. Therefore it is not a hard task to allow for global forecasting within the Neural-LAM framework. This has the benefit of being able to use any models in here also for global forecasting.

Global forecasting, as done in Oskarsson et al. (2024), is currently implemented on the branch prob_model_global. This issue is to keep track of merging of the prob_model_global branch into main.

The main changes in this branch include:

• Add script for creating of global forcing

  neural-lam/create_global_forcing.py

  Line 26 in f68a513

  def main():

  This will likely be taken care of through Creation of Static and Forcing Fields #34, rather than merging create_global_forcing.py directly.
• Add script for creating global static grid features

  neural-lam/create_global_grid_features.py

  Line 23 in f68a513

  def main():

  Also likely better handled in Creation of Static and Forcing Fields #34.
• Add script for creating global mesh graphs (multiscale and hierarchical)

  neural-lam/create_global_mesh.py

  Line 128 in f68a513

  def main():

  Will see if this should sit in this repo or in weather-model-graphs, and to what extent we keep the graphcast dependency for graph generation.
• Allow for 3d graph plotting of global graphs

  neural-lam/plot_global_graph.py

  Line 99 in f68a513

  def main():

  It is perhaps better to first move the 3d plotting functionality to weather-model-graphs and do this there. See point above.
• Enable generating variable weights based on pressure levels

  neural-lam/create_parameter_weights.py

  Line 53 in f68a513

  if global_ds:

  This is better handled by defining more general variable weighting setup where one can either set these weights manually per-variable in config files or use some pre-defined strategies.
• Add ERA5 dataset class

  neural-lam/neural_lam/era5_dataset.py

  Line 13 in f68a513

  class ERA5Dataset(torch.utils.data.Dataset):

  This should not be added as a separate new class, as all functionality in there should be possible through the generic dataset created in Multiple Zarr to Rule them All #54. But it needs to be checked that the result of Multiple Zarr to Rule them All #54 will work with global ERA5 data.
• Add functionality to save forecasts to zarr+xarray for later plotting or evaluation

  neural-lam/neural_lam/forecast_to_xarr.py

  Line 199 in f68a513

  def forecast_to_xarr(

• Add grid weighting (spatial points weighted differently) to loss and metrics

  neural-lam/neural_lam/metrics.py

  Line 41 in f68a513

  if grid_weights is not None:

• Make boundary forcing optional when rolling out forecasts

  neural-lam/neural_lam/models/ar_model.py

  Line 150 in f68a513

  def optional_boundary_forcing(self, pred_state, border_state):

• Plot example predictions for specific variables and lead times also during validation steps

  neural-lam/neural_lam/models/base_graph_model.py

  Line 183 in f68a513

  def validation_step(self, batch, *args):

• Allow for plotting of global data

  neural-lam/neural_lam/vis.py

  Line 79 in f68a513

  def plot_on_axis(

  Plotting functionality should be based on arbitrary projections rather than hard-coded to MEPS case. This should be done already by Multiple Zarr to Rule them All #54, but it needs to be checked that this works also for global forecasts.

[InFiNiTy0639]

Hi @joeloskarsson ,I noticed that ML-LAM is listed among the GSoC 2026 organizations, and I am keen to contribute. I would like to confirm whether this issue is open for contributors and if I may begin working on it as a potential GSoC proposal.

[joeloskarsson]

@InFiNiTy0639 yes, this issue is highly related to project idea number 4. Please see https://github.com/mllam/neural-lam/wiki/GSoC-ideas#4-global-weather-forecasting for some more details.

[yuvraajnarula]

Hi @joeloskarsson,

I’ve been reviewing the code related to this issue, particularly ARModel and metrics.py , to better understand how global forecasting is expected to integrate with the current LAM-specific rollout logic.

From what I can see:

• Boundary forcing is tightly embedded in unroll_prediction(), which may require some decoupling for a clean global implementation.
• The metrics layer already appears fairly general and likely needs minimal changes.

Before diving deeper, I wanted to clarify the intended scope and timing of this issue:

• Is Merge global forecasting capabilities from prob_model_global #63 currently open for incremental contributions?
• Or is it expected to be addressed mainly during the GSoC period?
• Are there specific sub-parts (e.g., boundary abstraction, mesh handling, dataset integration) that you would consider safe and useful to work on now?

I’d appreciate any guidance on where contributions would be most helpful at this stage.

[nehamanoj1105]

Hi @joeloskarsson ,

I’ve opened a small PR (#211) that makes boundary forcing optional in ARModel, as a first step toward clean global forecasting support.

This focuses only on decoupling boundary overwrite logic during rollout.

Happy to refine further if needed.

[selcuksntrk]

Hi @joeloskarsson @observingClouds @leifdenby,

I'm Selcuk Senturk, MSc Computer Science student and an AI/ML Engineer in IBM. I'm interested in contributing to the global weather forecasting extension as part of GSoC 2026 (175 hours).

My background connects to this project on multiple levels. My BSc in Physics means graph construction on spherical domains, coordinate systems, and the mathematics underlying GNN message passing are familiar ground, not just as ML tools but as physical representations. I work daily with PyTorch in production systems, and I've worked with graph-based deep learning architectures in my research and professional work. I'm also comfortable with data pipelines from working with structured scientific datasets.

The core challenge here, constructing a well-behaved global graph that handles the periodicity of longitude and the polar singularities is the kind of problem I find technically interesting.

I would like to discuss further if I'm a good candidate for this project under GSoC 2026 and how I plan to contribute.

Resume: Drive
LinkedIn: selcuksntrk

[joeloskarsson]

@yuvraajnarula we mainly view this as something that could be worked on throughout gsoc rather than now. But I will not stop any contributions already now.

@selcuksntrk Hi! Feel free to also join the slack and discuss further in the project channel there.