PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Alessandro Lucatelli, Ferdinand Unterhuber, Katharina Krenzer, Mennatallah El-Assady, Benedikt Soja, Christina Humer, Joel Oskarsson

arXiv:2610.35693v1cs.LGcs.HCphysics.ao-ph

TL;DR

Graph-based Earth-system models are challenging to develop because graph construction is difficult to validate and forecast errors are hard to explain. GNN-VIEW provides interactive graph inspection and attribution-based forecast explanations, demonstrated on an ocean graph and Hurricane Laura precipitation. The authors present its visualizations as evidence for debugging, not verdicts.

  • Problem

    Large algorithmically constructed graphs are difficult to validate, and aggregate forecast metrics indicate failures without explaining why a particular forecast was inaccurate.

    Source: §1 · pp. 1–2

  • Method

    GNN-VIEW is an interactive tool for exploring and inspecting graph structure before training and attributing selected trained-model forecasts to their inputs afterward.

    Source: §1 · p. 2, §2 · p. 2

  • Results

    GNN-VIEW demonstrates graph-flaw detection on an ocean forecasting graph and attribution-based analysis of poorly predicted Hurricane Laura precipitation.

    Source: §1 · p. 2, §3 · pp. 3–4

  • Takeaways & Limitations

    The authors present GNN-VIEW as providing interactive evidence that can give Earth-model developers actionable insights for debugging and improving models.

    Source: §4 · p. 4

    The authors identify design-choice limitations and call for structure-aware explainers, rollout and probabilistic-ensemble explanations, and user studies of workflow impact.

    Source: §4 · p. 4

Abstract

from arXiv · show

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.

1 Introduction

Graph-based Earth-system models use connectivity to constrain learning, but constructing and validating large graphs is difficult, and forecasts can be hard to explain. GNN-VIEW addresses both challenges with graph inspection before training and attribution-based forecast explanations afterward.

1 Introduction

  • GNN connectivity encodes spatial structure and constrains model representations before training, while improper graph initialization can degrade forecast capabilities.The models connect input-grid states to a coarser mesh through grid-to-mesh, mesh-to-mesh, and mesh-to-grid edges.Source: §1 · p. 1
  • Algorithmically constructed graphs can hide structural anomalies in raw tensors, while custom static plots are not scalable and are domain-dependent.These limitations make graph construction difficult to validate.Source: §1 · p. 1
  • Held-out aggregate metrics such as RMSE show how much and where a model fails, but not why it made an inaccurate forecast.Understanding why matters for debugging and uncovering potential shortcut learning.Source: §1 · p. 2
  • Existing visual analytics lack the geographic semantics needed for Earth-system models; nearby tools target node classification or data assimilation, and plotting utilities cannot effectively inspect dense, multilayer graphs.GNNLens targets node classification, while CloudNine addresses data assimilation rather than forecasting.Source: §1 · p. 2
  • GNN-VIEW supports the Earth-system GNN development pipeline across domains, with demonstrations on an ocean graph and a mispredicted Hurricane Laura field.The demonstrations address graph-construction flaws and forecast explanation.Source: §1 · p. 2

2 GNN-VIEW

GNN-VIEW organizes graph exploration, precise structural inspection, and post-training prediction explanation into modes aligned with a developer’s workflow. It combines geographic views and thresholded graph statistics with input attributions presented as maps and ranked features.

2 GNN-VIEW

  • Explore and Inspect show the information pathways available in the graph, while Explain measures how the trained model uses them.The three modes approach model representational capabilities from different angles.Source: §2 · p. 2
  • Explore presents the complete encoder-processor-decoder graph on a projected global map; Inspect uses node degrees and edge lengths to threshold selected subgraphs.Geographic patterns paired with statistical selection support graph diagnostics.Source: §2 · p. 2
  • Explain first lets users select a target prediction by time, variable, and grid point, then attributes it to input timesteps, variables, and grid points.Time-, space-, and variable-resolved error metrics guide users toward poorly performing cases.Source: §2 · pp. 2–3
  • Input × Gradient provides a fast qualitative overview, while Integrated Gradients is slower and more accurate; both return signed feature scores relative to a baseline.Summed absolute scores, normalized by the total, give each input field’s attribution share.Source: §2 · p. 3
  • Attribution heatmaps show spatial patterns, while ranked input fields ordered by attribution share average over space to summarize behavior across time and variables.The views address the high dimensionality of per-target attribution tensors.Source: §2 · p. 3
  • Explore and Inspect require graph topology; Explain additionally requires a trained TorchScript model and dataset, and the tool accepts Neural-LAM or Anemoi graph formats.The tool supports flat and hierarchical graphs and makes no assumptions about mesh generation.Source: §2 · p. 3

3 Use cases: GNN-VIEW in action

The use cases show GNN-VIEW surfacing ocean-graph connectivity flaws and contrasting attribution patterns for well- and poorly-predicted Hurricane Laura fields. The visualizations inform diagnosis, while the precipitation explanation remains a testable hypothesis.

3 Use cases: GNN-VIEW in action

  • The Njord graph’s mesh is generated by spatially clustering sea-surface grid points, and its cluster count is a free parameter.The use case examines iterative choices including an edge-length cap and a repair pass for isolated elements.Source: §3 · p. 3
  • Thresholding G2M out-degree at zero reveals isolated grid points whose states are not encoded, undermining representation of the Kara Sea and informing whether to repair disconnections or add mesh nodes.The Njord graph’s irregular ocean domain requires balancing land crossings, basin coverage, and connections to every sea grid point.Source: §3 · p. 3
  • Mesh-edge length and mesh-node degree thresholds expose a Wales-crossing edge and isolated western Denmark nodes that disconnect the Baltic Sea.A 250 km cutoff isolates the land-crossing edge, whereas 200 km also captures legitimate long-range links, making threshold sensitivity visible.Source: §3 · pp. 3–4
  • The Hurricane Laura case uses a GraphCast-style GNN trained on 1.5° ERA5 data and predicts the storm’s collapse six hours after Category 4 landfall from two earlier atmospheric states.The inputs are states at 00:00 and 06:00 UTC on 2020-08-27; the target time is 12:00 UTC.Source: §3 · p. 4
  • For poorly predicted precipitation, attribution rankings are led by higher-altitude geopotential and temperature, while known drivers such as moisture and winds rank marginally or not at all.Precipitation at t−6 h is outside the top ten, and its attribution map is weak and mixed in sign; the authors hypothesize that coarse resolution may prompt reliance on large-scale circulation patterns.Source: §3 · p. 4, Figure 3 · p. 4
  • For accurately predicted 925 hPa temperature, attribution is dominated by the same field at t−6 h and is compact around the target, unlike the precipitation pattern.Figure 3 pairs target-error heatmaps with signed input-attribution maps and ranked input fields for the two target fields.Source: §3 · p. 4, Figure 3 · p. 4

4 Summary and future work

GNN-VIEW brings graph validation before training and prediction debugging after training into developers’ view. The authors frame its outputs as interactive evidence rather than verdicts and identify extensions and further workflow evaluation as future work.

4 Summary and future work

  • GNN-VIEW provides interactive evidence, not verdicts, in scenarios involving a poorly connected ocean graph and a mispredicted precipitation field.The authors say the tool offers actionable insights to help developers debug and improve models, supporting a more transparent development process.Source: §4 · p. 4
  • Future work includes structure-aware explainers, explanations for forecast rollouts and probabilistic ensembles, and user studies of workflow impact.The authors describe these as ways to explore limitations of GNN-VIEW’s design choices and broaden its impact.Source: §4 · p. 4

A.1 Single-step inference and evaluation metric

The tool evaluates single-step forecasts and provides error summaries, while local deployment lets users use their own datasets without changing downstream pipelines.

A.1 Single-step inference and evaluation metric

  • Given normalized states at t−12 h and t−6 h plus forcing, the model predicts the state at t; per-grid-point error is computed on demand.Mean and maximum RMSE, plus error concentration, are precomputed across the test period for a ready overview.Source: §A.1 · p. 8

A.2 Local deployment

  • Local deployment supports users’ own datasets without changing inference, error, or attribution pipelines.Source: §A.2 · p. 8

B GNN-VIEW latency

Latency grows linearly with graph size, with edge rendering the largest per-element cost; full-graph rendering is slower than the tool’s usual zoomed-in use.

  • Synthetic tests vary grid points, mesh nodes, or edges one at a time while holding the other component counts fixed; edge rendering costs 5.73 µs/element versus 3.71 for grid points and 4.19 for mesh nodes.Edges also tend to be more numerous in real graphs.Source: B GNN-VIEW latency · p. 8, Figure 4a
  • An oceanography graph with approximately 22× more edges requires almost 20× more rendering time, and latency scales linearly for real graphs.Source: B GNN-VIEW latency · p. 8
  • Full-graph latency is measured with the entire graph on screen, whereas typical finer zoom levels render only a portion and therefore have considerably lower latency.Source: B GNN-VIEW latency · p. 8

C Sanity Checks

The sanity checks assess whether explanations reflect learned model parameters and input-output behavior, using randomized parameters and deletion tests across selected targets.

  • Randomizing model parameters and performing deletion tests assess explanation sensitivity to learned parameters and faithfulness to model input-output behavior.The checks are adapted to graph-based regression with large inputs and direct skip-connections for the target field at t−6 h, and results are compared with reference controls.Source: C Sanity Checks · p. 8
  • The checks cover 50 random ERA5 test-set targets (S0) and 50 highest-error targets (S1), with a cap of 10 high-error targets per variable family.The cap avoids a pool dominated by high errors that often occur for Vertical Velocity at 50 hPa.Source: C Sanity Checks · pp. 8–9
  • The Hurricane Laura case provides a direct sanity check, while the other two target pools provide more robust results; checks use the GraphCast-like trained model.Source: C Sanity Checks · pp. 8–9

C.1 Randomization tests

The randomization tests compare trained-model explanations with randomized-model controls and inputs, finding trained explanations less similar to randomized ones than randomized explanations are to each other.

  • All weights are randomized together using five seeds; paired tests compare trained versus randomized models and trained explanations versus inputs, with randomized-model and input comparisons as controls.Attribution tensors use Pearson correlation and rankings use Kendall correlation; Pearson is chosen because localized receptive fields produce many near-zero entries that could inflate Spearman correlation.Source: §C.1 · p. 9
  • Trained-model explanations are less similar to randomized-model explanations than randomized models are to one another, for both attribution maps and rankings.This result holds for the Hurricane Laura temperature and precipitation targets.Source: §C.1 · p. 9, Figure 5 · p. 10
  • Similarity between explanations and inputs is almost null for spatial patterns but remains present for per-variable magnitudes.Magnitude share sums absolute input values over space and normalizes across variables.Source: §C.1 · p. 11
  • Integrated Gradients randomization results are stable across different target predictions, while Input × Gradient has similar sensitivity and is suited to fast attributions but lacks completeness by design.Source: Figure 6a

C.2 Deletion tests

Deletion tests check whether removing highly ranked inputs changes predictions sooner or more than control orderings. The reported tests support IG and IxG for identifying influential inputs, rather than relying only on input magnitude.

C.2 Deletion tests

  • Inputs are set to zero in normalized space, equivalent to their training mean; forcing and static features and the target’s t−6 h field are not removed.The target field is excluded because a skip-connection copies it directly to the output.Source: §C.2 · p. 11
  • Field deletion ranks whole variable-and-timestep fields by attribution share; single-entry deletion ranks individual inputs, with random and input-magnitude controls.Faithfulness correlations compare each field’s prediction change with its IG share and, as a control, its input-magnitude share.Source: §C.2 · pp. 11–12
  • The MoRF curve measures absolute target-prediction change as inputs are removed, and AOPC summarizes its mean over removal steps on a normalized log-k axis.The computation excludes k = 0 and full removal, where all orderings coincide.Source: §C.2 · pp. 11–12
  • IG identifies influential fields better than high-magnitude fields in the Hurricane Laura tests, and both IG and IxG perform well across the S0 and S1 target pools.The tests also support using IxG as a fast attribution method.Source: §C.2 · p. 12, Figure 7 · p. 12

D Integrated Gradients

Integrated Gradients (IG) attributes prediction shifts from a climatological-mean reference input to the actual input, with completeness and attribution convergence assessed as integration steps increase. Four steps recover 96% of the prediction shift at the median for the worst-points pool, while the Hurricane Laura tests report high recovery for temperature and precipitation.

D Integrated Gradients

  • IG integrates gradients from each field’s climatological mean to the actual input; positive and negative scores indicate increases and decreases relative to the baseline output.Completeness means the attribution scores sum to the prediction shift from the reference input to the actual input.Source: D Integrated Gradients · pp. 12–13
  • 96% of the prediction shift is recovered at the median by IG with m = 4 for the worst-points pool, and attributions appear stable after four steps.Maps and input rankings computed with m = 8 are well correlated with those computed with m = 4, as are larger-step results.Source: D Integrated Gradients · p. 13, Figure 9 · p. 14
  • More integration steps improve completeness at the cost of latency; m = 4 is the tool default, while m = 8 is suggested for local GPU use because m = 4 completeness varies significantly.The completeness gap compares explained and actual prediction shifts; convergence is assessed by correlation between attributions at m and 2m.Source: D Integrated Gradients · pp. 13–14
  • With m = 4, IG recovers 99.42% of the temperature prediction shift and 98.8% of the precipitation shift in the Hurricane Laura use case.The m = 64 attributions account for 99.99% of the shift and serve as the comparison ground truth in this use case.Source: D Integrated Gradients · p. 14, Figure 10 · p. 15

Related papers

Loading 2610.35693v1…