NiceShot!

An interactive dashboard for exploring tokamak plasma shot data. Point it at a shot-statistics file and get an instant browser UI for slicing, visualising, and comparing shots.

Features

• Projection — UMAP or PCA scatter of every shot, coloured by any column. Backed by a content-hash cache so reloads are instant.
• Pairwise scatter — any two numeric columns plotted against each other, with linear/log axis toggles.
• Correlation — interactive Pearson correlation heatmap for any selection of numeric columns.
• Data table — sortable, virtualized table with shot-ID search, cross-highlight with scatter plots, and CSV export.
• Time traces — per-shot signal plots loaded on click. Supports local parquet/CSV files, live UDA, live SAL, and FAIR MAST (Zarr/netCDF, local or remote) backends.
• Filters — up to 6 simultaneous column filters combinable with AND / OR logic. All plots update live.
• Clustering — run K-Means, DBSCAN, or Agglomerative clustering on any set of numeric columns. Results colour the scatter plots immediately; clusters can be given human-readable class names.
• Cluster centroid traces — mean time-series per cluster, computed automatically after clustering and relabelled live as class names change.
• Outlier detection — flag anomalous shots with Isolation Forest or Local Outlier Factor. Outliers are highlighted in red on the scatter plots and sample traces are loaded automatically.
• Classification — train a Gradient Boosting, Random Forest, or Gaussian Process model on a target column and any set of numeric features. Predicted labels colour the scatter plots, and an optional decision-surface background shows the predicted probability of a chosen class. See docs/classification.md.
• CSV export — download the full data table with cluster_id, cluster_name columns appended when clustering has been run, and label plus per-class probability columns when classification has been run.
• SHAP decision plots — per-shot feature attribution rendered inline, for a supplied SHAP file or for a trained classifier (optional, requires --shap-data or a trained model).
• Reference graph — overlay the full reference-shot lineage on any scatter plot (optional, requires reference_shot_col in config).
• Lineage changes — see which variables changed between a shot and the shots before it in the same reference lineage. Shows a ranked summary of every variable, a colour-coded history table, a lineage tree, the operator's notes, and sparklines (optional, requires reference_shot_col in config).
• Semantic search — find shots similar to a selected one via nearest-neighbour search in feature space. Results are highlighted on the scatter plots with gold ring markers.
• Long-format files — point NiceShot at a table with one row per (shot, variable) and pick the variable from a header dropdown. Only the selected variable's rows are read, and the projection is computed for that variable alone.
• Extensible backends — add support for new data sources (MDSplus, HDF5, custom APIs, …) by subclassing TraceBackend or ShotDataBackend and registering via plugins: in config.

Requirements

Python ≥ 3.12

Install

pip install nice-shot
pip install "nice-shot[shap]"       # + SHAP plots, xarray, matplotlib

Run

nice-shot path/to/shot_stats.parquet

SHOT_DATA is the only required argument — everything else has a default. Open http://localhost:8050 in a browser.

By default nice-shot runs under gunicorn with 4 worker processes, which supports multiple concurrent users. On first run, UMAP/PCA is computed in the master process and cached; subsequent starts are instant.

For local development with hot-reload use --debug:

nice-shot path/to/shot_stats.parquet --debug

Common flags

Config-backed flags

Every option in nice_shot/config.yaml also has a CLI flag of the same name. An explicit CLI flag overrides the config file, which overrides the built-in default:

Run nice-shot --help for full descriptions.

Configuration

Edit nice_shot/config.yaml (or pass --config to point elsewhere):

backend: parquet        # parquet | uda | sal

signals:                # columns shown in the time-trace panel
  - ip
  - ne
  - dalpha

time_window:
  min_time: 0.0
  max_time: 1.0

projection_method: umap # umap | pca

projection_options:     # hyper-parameters of the projection
  n_components: 2       # the plots draw the first two
  n_neighbors: 15       # umap only
  min_dist: 0.1         # umap only

umap_features:          # omit to use all numeric columns
  - ip_max
  - ne_max
  - bt_max

reference_shot_col: reference__number   # omit to hide the feature

Any of these can be overridden for a single run with the matching CLI flag — see Config-backed flags above.

Most of them can also be changed while the app runs, from the Configuration tab. Those changes apply to your browser only, and a restart returns to the config file — see Configuration tab.

Data

Shot statistics file (SHOT_DATA) — a flat .parquet or .csv with one row per shot. The shot ID column is detected automatically (shot_id, shot, pulse, number, …).

Per-shot traces (--data-dir) — one .parquet or .csv per shot, laid out as:

<data-dir>/<any-subdir>/<shot_id>.parquet

Each file needs a time column and one column per configured signal.

Pre-computed projection (--projection) — a .npy (shape (n,2) or (n,3)), .csv, or .parquet with shot ID and two coordinate columns.

SHAP values (--shap-data) — an xarray NetCDF file with shot_id and feature dimensions.

See docs/data-formats.md for full schema details.

Docs

uv run --dev zensical serve

Opens the full documentation at http://localhost:8000.