Feature: Add diagnostic visualization for isolated/disconnected mesh nodes

[ArnabTechiee]

Motivation

When configuring Local Area Models (LAM) and tuning parameters like mesh_node_distance, it is easy to accidentally create spatial configurations where certain grid nodes become completely disconnected (0-degree nodes). It is highly beneficial to visualize these "graph holes" early in the data pipeline.

I built a standalone diagnostic tool to map these boundaries (original Gist here: https://gist.github.com/ArnabTechiee/7fc757f79379709943300d879af3baec), and following a discussion on Slack with @joeloskarsson and @leifdenby, I would like to integrate this directly into weather-model-graphs.

Current Standalone Output

Proposed Implementation in WMG

Currently, my standalone script uses a mock generator and raw matplotlib. To integrate this cleanly into the WMG architecture, I propose adapting the logic to utilize WMG's native networkx data structures and existing visualization modules.

Here is the proposed integration plan:

1. New Utility Function: Add a new function (e.g., plot_isolated_nodes(graph, ax=None)) inside src/weather_model_graphs/visualise/plot_2d.py.
2. Identify Isolates: Instead of my custom distance thresholds, leverage the backend by using networkx.isolates(graph) to find all nodes with a degree of 0.
3. Attribute Tagging: Iterate through the graph and tag the nodes with a new string attribute (e.g., graph.nodes[node]["Connection Status"] = "Isolated" vs "Connected").
4. Native Rendering: Pass the tagged graph directly into WMG's existing nx_draw_with_pos_and_attr function. This avoids redundant matplotlib code and ensures the diagnostic plot perfectly matches WMG's native aesthetic, colorbars, and discrete legends.
5. Exposure: Expose this function in visualise/__init__.py so downstream users (or CLI arguments in neural-lam) can easily call wmg.visualise.plot_isolated_nodes(graph).

Does this architectural approach look good to you both? If so, I’d be happy to open a draft PR for this!

[yuvraajnarula]

@ArnabTechiee
This is a great addition. Drawing from the spherical mesh prototype, a couple of thoughts:

1. Global graphs also need this – Even with quasi-uniform meshes, grid points near poles could become disconnected if radius_factor is too small. This visualization would help users debug connectivity issues in spherical graphs too.

2. Integration with global graph work – Once this lands, I can add a test case that verifies no isolated nodes exist for reasonable radius_factor values (e.g., 0.6). This gives us a quantitative check alongside your visual one.

3. CLI hook – If we later add a --plot-isolated flag to the graph generation CLI (#384), users could auto-diagnose after creation.

Happy to help test this with icosahedral graphs once you have a draft PR.

[ArnabTechiee]

Hey @yuvraajnarula, thanks for the feedback and for offering to help test!

Those are excellent points. Making sure this tool is agnostic and works seamlessly for both flat LAM grids and spherical/icosahedral meshes is definitely the goal, and tying it into a --plot-isolated CLI hook later sounds like a perfect use case.

Since this touches the core visualization module, I’m going to hold off on opening the draft PR just until @leifdenby and @joeloskarsson get a chance to weigh in on the proposed architectural approach. Once they give the green light on the networkx integration plan, I will get the code up and ping you so we can test it against your global graphs!

[ArnabTechiee]

Hi @leifdenby and @joeloskarsson, just a gentle ping on this when you have a moment!

I am ready to start building and testing this out across the various LAM and global mesh configurations. Does the networkx attribute-tagging approach outlined above align with how you want the visualise module structured?

If the architecture looks good to you both, please let me know and I can get a draft PR opened up this week!