NcDashboard bridges the gap between legacy viewers (like NcView) and complex, code-heavy workflows (Xarray/Iris). It enables browser-based exploration of remote HPC datasets, providing a high-performance, reactive interface for atmospheric and oceanographic data.
- Web-Native Remote Access: High-performance exploration of large-scale data directly on HPC clusters via a simple web interface.
- Linked Multi-Modal Views: 1D to 4D visualizations (maps, profiles, time-series) through a recursive node-based tree model. Viewport zoom and position are synchronized across related plots.
- LLM-Driven Analysis: Natural-language custom fields (e.g., "calculate vorticity") generated as containerized Python code and auto-integrated into the UI.
- Automatic Robust Scaling: Uses percentile-based scaling (2% - 98%) to ensure visualizations look good immediately, even with extreme outliers.
| Feature | NcView | Iris/Xarray | NcDashboard |
|---|---|---|---|
| Remote/HPC Workflow | Limited | Manual | Native |
| Interactive Pan/Zoom | Basic | Coding Req. | Dynamic |
| AI-Custom Analysis | No | No | Built-in |
| Session Sharing | No | Scripts | JSON State |
This project uses uv for lightning-fast Python package and environment management.
If you don't have uv installed, follow the instructions at astral.sh/uv.
git clone https://github.com/olmozavala/ncdashboard.git
cd ncdashboard
uv syncThis will automatically create a .venv directory and install all required dependencies as defined in pyproject.toml.
To run the dashboard, provide the path to the NetCDF file or directory using uv run:
# Single file
uv run ncdashboard.py test_data/gom_t007.nc
# Multiple files with regex
uv run ncdashboard.py test_data --regex "*.nc"
# Specify a custom port
uv run ncdashboard.py path --port 8055Bulk load NetCDF files; variables are auto-mapped to 1D–4D visual nodes for one-click plotting. Navigate through time and depth dimensions seamlessly with synchronized sliders.
Compute derived fields (e.g., vorticity, velocity, mean temperature) via LLM prompts. The AI generates isolated Python code, executes it, and inserts results directly as new interactive figures.
Save/load your full workspace—including active plots, zoom levels, selected indices, colormaps, and even AI-generated code—to a single JSON file.
Built on a modern Python stack (Panel, Xarray, Datashader), the NcDashboard architecture prioritizes performance and scalability by utilizing a recursive tree-based model.
- Recursive FigureNodes: A tree-based model where nodes (1D–4D) inherit data context and handle dynamic rendering updates.
- Lazy Evaluation: Data is streamed via Dask/Xarray and rendered on-demand, allowing large-scale exploration without memory overflow.
- AI Execution Sandbox: LLM-generated code runs in an isolated environment with restricted imports and restricted memory.
Our next milestone is the integration of vector GIS layers (Points, Lines, Polygons; e.g., Shapefile/GeoJSON). This will enable:
- Advanced Spatial Analytics: Perform shape-based operations such as calculating scientific quantities along transects (lines).
- Regional Focus: Crop variables within complex polygons (states, EEZs, etc.).
- In-Situ Comparison: Automated value extraction at station locations (points).