Visual Analytics Workbench for Embedding-Based Weather Data Exploration
An open-source visual analytics workbench has been created by researchers to facilitate the exploration of embedding-based representations from extensive, high-dimensional datasets in Earth system science, sourced from both physics-based and AI-driven weather and climate models. This tool tackles the issue that nearest neighbors in latent space might lack scientific significance due to preprocessing, geographical factors, or model biases. It connects embedding experiments with source data, metadata, spatial context, and model configurations, enabling users to trace latent-space findings back to physics. Users can investigate latent spaces for various models and perform both global and local queries. The workbench aids in examining how embeddings structure meteorological data, comparing representation models, and validating findings against physical evidence. The research paper is accessible on arXiv with ID 2605.00972.
Key facts
- Open-source visual analytics workbench for embedding-based exploration of weather and climate data.
- Addresses the problem that nearest neighbors in latent space may not be scientifically meaningful.
- Links embedding experiments to source data, metadata, spatial context, and model configurations.
- Supports similarity search and analog retrieval through embedding-based representations.
- Enables inspection of how embeddings organize meteorological data.
- Allows comparison of representation models and development of retrieval strategies.
- Facilitates verification of results against physical evidence.
- Paper published on arXiv with ID 2605.00972.
Entities
Institutions
- arXiv