GNN-VIEW: Visual Inspection and Explanation of Graph-based Earth-System Models

AI summary

Overview

Research area: Machine learning for Earth-system modeling, specifically graph neural network (GNN) based weather and ocean forecasting, with a focus on visual analytics and explainable AI (XAI).

Technical level: Intermediate. The paper assumes familiarity with the encoder–processor–decoder GNN paradigm and with attribution methods, but the tool itself is described in accessible terms.

Scope: A workshop paper presenting GNN-VIEW, a browser-based interactive tool for inspecting the graph structure of Earth-system GNNs before training and for explaining individual forecasts after training, demonstrated on an ocean graph and a hurricane precipitation forecast.

What This Paper Is About

Graph-based Earth-system models encode the spatial connectivity of the planet in a graph, and that graph constrains what the model can learn before training even begins. Developers currently build these graphs algorithmically and check them with ad-hoc static plots, which does not scale to large multi-layer graphs, and after training they can only see how much and where a model fails via aggregate metrics like RMSE — not why. GNN-VIEW addresses both gaps with one interactive tool: it exposes the graph's information pathways for pre-training debugging, and it uses attribution-based explainability to reveal which input features drove a specific forecast.

Key Contributions

1. An interactive visual tool spanning the full development pipeline. GNN-VIEW is organized around three modes — Explore, Inspect, and Explain — that follow a developer's workflow, covering both pre-training graph validation and post-training forecast explanation within a single browser-based application.

2. Structure-aware graph diagnostics. The tool characterizes local subgraph structure with three quantities — node in-degree, node out-degree, and edge length — and lets developers set upper and lower thresholds on these statistics to highlight suspicious graph elements directly on a projected global map.

3. Attribution-based forecast explanation. The Explain mode attributes a chosen target prediction (time, variable, grid-point) back to its inputs (timesteps, variables, grid-points), offering two methods: Input × Gradient (IxG) for fast qualitative overview and Integrated Gradients (IG) for slower, more accurate attributions. Results are shown as spatial heatmaps plus a ranked list of input variables ordered by attribution share.

4. Two worked use cases plus validation appendices. The paper demonstrates the tool on a flawed Njord ocean graph and on a mispredicted precipitation field during Hurricane Laura, and reports latency measurements, randomization tests, deletion tests, and IG completeness experiments.

Main Findings

• Isolated ocean grid points were made visible. Setting the out-degree threshold to 0 in the grid-to-mesh (G2M) subgraph of the Njord ocean graph surfaced grid points whose states are never encoded into the model, undermining the representation of the Kara Sea. This informs whether to run a disconnection repair algorithm or increase the number of mesh nodes.

• Two further graph failure modes were detected. Thresholding mesh edge length and mesh node degree revealed a mesh edge crossing over Wales (connecting water masses separated by land) and isolated mesh nodes on the western shores of Denmark, leaving the Baltic Sea disconnected from the rest of the ocean.

• Threshold sensitivity is directly visible. Comparing two edge-length cutoffs, 250 km isolates the land-crossing edge, while 200 km also captures legitimate long-range links — making the calibration trade-off apparent rather than hidden.

• Accurate and inaccurate forecasts show qualitatively different attribution patterns. For temperature at 925 hPa, the attribution ranking is dominated by a single input — the same field at t−6 h — and the attribution map is compact and centered near the target. For 6-hour precipitation, the ranking is relatively flat, led by geopotential and temperature at higher altitudes, which characterize the surroundings rather than the storm. Known physical drivers of precipitation such as moisture and winds enter marginally or not at all, and precipitation at t−6 h does not enter the top ten.

• A testable hypothesis emerged from the explanation. The authors propose that at this coarse resolution the model falls back to correlating precipitation with larger-scale circulation patterns instead of drawing on relevant physical processes — the kind of hypothesis the tool is designed to generate.

• Latency scales linearly with graph size and is dominated by edges. Measured server latency was 5.73 μs per edge element versus 3.71 and 4.19 μs for grid points and mesh nodes. An oceanography graph with approximately 22× more edges required almost 20× more time.

• Sanity checks supported both attribution methods. In randomization tests with 5 seeds, correlation between the trained model's explanation and randomized models' explanations was always lower than similarity among randomized models, for both attribution maps and rankings. Deletion tests showed IG reliably identifies the input fields that change model prediction the most rather than those with the highest absolute magnitude; IxG also performed well. The authors note that prior literature marks IG as performing poorly under randomization tests, but that it performs fairly well in this specific context.

• IG with few integration steps is close to converged. With m = 4 steps, IG recovers 96% of the prediction shift at the median for the highest-error target pool. Attributions computed with m = 8 are well correlated with those at m = 4, which is why m = 4 is the tool's default. For the Hurricane Laura case, m = 4 recovers 99.42% of the temperature prediction shift and 98.8% of the precipitation shift relative to m = 64 (which accounts for 99.99%). Completeness at m = 4 still varies and at times exceeds the 5% limit proposed in prior work, so the authors suggest m = 8 when deploying locally with GPU acceleration.

Methodology in Plain English

The authors built a web application that reads a graph produced by existing Earth-system modelling libraries (Neural-LAM or Anemoi format, covering both flat and hierarchical graphs) and renders it on a projected global map, making no assumptions about how the mesh was generated. Before training, developers use the Explore mode to see the whole graph and the Inspect mode to select a subgraph and compute simple local statistics — how many edges point in, how many point out, how long each edge is — then drag thresholds on those statistics to flag anomalies geographically.

After training, the model is uploaded as a TorchScript .pt file alongside the dataset. The tool computes error metrics resolved over time, space, and variables to guide the developer toward weak spots. Once the developer picks a target prediction, the tool runs an attribution method that scores each input feature by how much it pushed the prediction above or below a baseline. Because the resulting tensor is extremely high-dimensional, the tool collapses it two ways: heatmaps for spatial distribution (two side by side for comparison across timesteps or variables) and a ranked list by attribution share, obtained by summing absolute scores over space and normalizing by the total.

To check that the explanations are meaningful rather than artifacts, the authors ran two families of tests adapted to spatio-temporal regression: randomization tests (comparing explanations from the trained model against models with weights randomized under 5 seeds, with control pairings) and deletion tests (progressively zeroing the highest-ranked inputs in normalized space, i.e. to the training mean, and measuring the absolute change in prediction in units of the target variable's standard deviation). These were run on the Hurricane Laura case plus two pools of 50 targets each: S0, randomly chosen from the ERA5 test set (1 Jan 2020 to 10 Jan 2021), and S1, the 50 highest-error targets, with a cap of 10 per variable family to keep the pool representative.

Why This Matters

Impact on research. The paper argues that validating graph construction and understanding forecast failures are both currently served only by custom static plotting, which is not scalable and is domain-dependent. Existing visual analytics tools for GNNs are model-agnostic and do not incorporate the geographic semantics Earth-system models need; the two closest works, GNNLens and CloudNine, target node classification only and the data assimilation stage respectively. GNN-VIEW fills that gap and, importantly, frames its output as interactive evidence rather than verdicts, supporting an iterative debugging loop.

Real-world applications:

• Operational weather forecasting — diagnosing why a GNN forecast fails for a specific extreme event, as shown for Hurricane Laura's precipitation six hours after landfall.
• Ocean forecasting — catching disconnected basins, land-crossing edges, and unencoded sea grid points before committing to expensive training runs.
• Other Earth-system domains — the authors note the same design challenges extend to wildfire, hydrology, and ice sheet modeling.
• Trustworthy model development — surfacing potential shortcut learning that aggregate RMSE hides.

Industry relevance. GNN-based weather models have matched and often surpassed operational numerical prediction at a fraction of the inference cost and are increasingly adopted operationally, so tooling that shortens the debug cycle and increases transparency has direct value for organizations deploying these models. The tool ships with two default datasets — ERA5 at 1.5° and 6-hourly resolution for weather, and an adaptation of GLORYS12 re-gridded to 0.25° with sea-ice added for oceanography — and supports local deployment on users' own datasets.

Future Directions

1. Structure-aware explainers. The authors suggest connecting model behavior more closely to the graph structure itself, rather than treating attribution as a property of input fields alone.

2. Explainability of forecast rollouts. Extending attribution from single-step forecasts to multi-step rollouts would broaden the tool's reach, since operational use involves iterated forecasting.

3. Probabilistic model ensembles. Supporting explanation for ensemble and probabilistic models is named as another avenue for wider impact in the community.

4. User studies. The authors call for further evaluation of the tool's actual impact on developers' workflows through user studies — currently the evidence is two demonstration scenarios plus technical sanity checks.

Target Audience

Developers and researchers building graph-based Earth-system models — particularly those working with Neural-LAM, Anemoi, or GraphCast-style architectures — who need to validate mesh construction or debug specific forecast failures. It is also relevant to visualization and XAI researchers interested in domain-specific adaptation of attribution methods, and to practitioners in operational weather and ocean forecasting evaluating whether such models are trustworthy enough to deploy. Readers seeking benchmark accuracy comparisons against numerical weather prediction will not find them here; the paper is a tool demonstration, not a model-performance study.

Authors’ abstract

Machine learning Earth-system models are becoming increasingly important, and so is their development. Graph-based models, one of the most prominent approaches, allow developers to constrain the learning process before training via graph connectivity. Yet, developing these models remains challenging: the graph is built algorithmically with little intuition about the final result beyond static plotting, which requires extra work and is often not detailed enough. Even after successful training, how the model arrives at forecasts remains obscure. We present GNN-VIEW, an interactive browser-based tool that supports the development of such systems. GNN-VIEW exposes the graph's information pathways, allowing model developers to debug the graph structure before training. After training, GNN-VIEW uses attribution-based explainability to reveal which input features are used by the trained model for a given forecast. This helps developers understand why a model fails. We showcase the usefulness of GNN-VIEW in the construction stage of a graph for an ocean forecasting model, and in explaining poorly predicted precipitation of a trained weather forecasting model during Hurricane Laura.