InferStudio

InferStudio is NSF NCAR's web application for running AI weather model inference and visualizing the results. It's run with Panel and runs on NCAR's Casper HPC system via Open OnDemand, giving researchers a single interface to launch inference for multiple AI weather models, browse the resulting forecasts, and verify them against GFS analysis — without needing to manage each model's environment or write plotting code by hand.

Supported models

Model	Backend	Status
AIFS	earth2studio	Fully supported (visualization + verification)
Aurora	earth2studio	Fully supported (visualization + verification)
Pangu	earth2studio	Fully supported (visualization + verification)
WXFormer	MILES-CREDIT	Inference supported; visualization not yet wired up (different output format)

The two tabs

Visualization

• Datasets — browse available simulation runs (each one a folder containing one subfolder per model, e.g. AIFS/, Aurora/). Only one dataset can be viewed at a time.
• Variable & Level — a single shared set of controls (Time, Variable, Level, Colormap) drives every plot on screen at once, so comparing models means looking at the exact same field, level, and forecast step across all of them.
• Field plots — one panel per model, each labeled with the actual valid time and forecast lead time (not just a raw time-step index).
• Forecast Verification Statistics — on request, computes RMSE and CRPS (which, for a single deterministic forecast, is equivalent to MAE) for the current variable/level at every available lead time, verified against GFS analysis fetched via earth2studio. Handles grid mismatches between models (e.g. some models use a 720-point latitude grid, GFS uses 721) via interpolation.
• Metadata — quick reference for a selected dataset: time range, grid size, available variables, number of forecast steps, etc.

An example dataset ships with the app and is selected automatically on startup, so the Visualization tab has something to show before you've run any inference yourself.

Inference

• AI Model — choose one or more models to run.
• Output Parameters — simulation name (auto-generated from the selected models and a timestamp, but editable), output directory, and which surface/upper-air variables to save.
• Time Settings — pick a start date and a lead time (in forecast steps); combined with the time-step increment, this determines the run's end date automatically.
• Launcher — runs each selected model sequentially, showing live status (pending/running/done/error/cancelled) and streaming console output per model. Runs can be cancelled mid-flight. When a run finishes, it's automatically selected back in the Visualization tab.

Project layout

File	Purpose
app_layout.py	Top-level page layout, dataset scanning, and wiring between tabs
datasetPlot.py	Shared plot controls and per-dataset field plot rendering
earth2StudioPlot.py	Reads earth2studio-format NetCDF output and renders individual field plots
forecastStatsPanel.py	RMSE/CRPS verification against GFS analysis
inferenceTab.py	Inference tab logic: model runs, status tracking, console output
outputParams.py	Simulation name, output directory browser, variable selection
timePicker.py	Start date / lead time / time-step-increment controls
datasetSelector2.py	Dataset browser (checkbox list with single-selection)
metadata.py	Metadata display panel
earth2StudioRunner.py, milesCreditRunner.py	Per-backend model execution
static/styles.css	App-wide styling

Running InferStudio

From a Jupyter notebook cell:

from pathlib import Path
import panel as pn
from app_layout import build_app

DATA_DIR = Path("/path/to/your/output/directory")
pn.extension('modal', notifications=True)

template = build_app(DATA_DIR)
template.servable()

Each AI model backend (AIFS, Aurora, Pangu, WXFormer) runs from its own per-model Python environment on Casper, since the models have incompatible dependency stacks (different CUDA/PyTorch/library requirements).

Known limitations

• WXFormer output uses MILES-CREDIT's ERA5-style format (a true level dimension rather than earth2studio's flattened u850-style variables) and isn't yet plotted or verified — it currently shows a placeholder in the Visualization tab.
• Only one dataset can be viewed at a time in the Visualization tab.
• Forecast verification requires network access to fetch GFS analysis data, and can take a noticeable amount of time for datasets with many forecast steps.