Research & Papers

GeoXplain lets you see why weather AI models predict what they do

Open-source Python toolkit visualizes attribution maps for climate foundation models.

Deep Dive

GeoXplain is a new open-source Python toolkit designed to make weather and climate foundation models more interpretable. Developed by researchers and submitted to IEEE VIS 2026, it accepts attribution bundles (grids plus metadata) and renders them as interactive maps or globes in notebooks or browsers. Users can explore explanations across atmospheric pressure levels, forecast time, and target annotations, with optional ERA5 overlays.

The toolkit is model-agnostic but ships with an Aurora Adapter—its first computation backend—that computes explanations for the Aurora foundation model using gradient saliency, Integrated Gradients, RISE, ViT-CX, and multi-frame variants. It runs locally on GPU, via a GPU listener, or through SLURM. Analysts can install it via `pip install geoxplain` and access the code on GitHub.

Key Points
  • Supports 6 explanation methods: gradient saliency, Integrated Gradients, RISE, ViT-CX, multi-frame saliency, and Integrated Gradients rollouts
  • Includes an Aurora Adapter for on-the-fly explanations via local GPU, GPU listener, or SLURM
  • Renders attribution maps in map or globe mode with linked timelines and ERA5 overlays

Why It Matters

Makes black-box weather models interpretable, helping scientists trust and debug climate forecasts.

📬 Get the top 10 AI stories daily