MPAS-Ocean Visualization & Benchmarks

MPAS-Ocean (Voronoi mesh, 235,160 cells x 60 layers) visualization toolkit and DiT model benchmark results for the FANR paper.

Structure

ocean/
├── visualization/          Exploratory vis of raw MPAS-Ocean NetCDF data
├── rendering/              Paper-quality Voronoi cell rasterization pipeline
├── benchmarks/             DiT architecture ablation charts (PSNR, time, size)

Data files (not in repo)

Large data files are gitignored. Obtain them from the compute server before deleting local copies.

File	Size	Location on server	Used by
visualization/ocean.nc	1.8 GB	/project/ml4scivis/fan/data/ocean/output.nc	plot_ocean_layers.py, prepare_coords.py, render.py
visualization/coord.npy	169 MB	Generated locally by prepare_coords.py	—
visualization/salinity_t0.raw	56 MB	Generated locally	—
visualization/mesh.nc (3.3 MB) and rendering/cell_id_map.npz (1.5 MB) are tracked in git.

visualization/

Matplotlib-based exploration of the raw ocean.nc dataset.

Script	Purpose
plot_ocean_layers.py	Per-layer global temperature & salinity scatter plots (Kindlmann colormap)
prepare_coords.py	Normalize lon/lat/layer to [-1, 1] → coord.npy for neural network input

Data files:

• ocean.nc — MPAS-Ocean output (1.8 GB, 235160 cells x 60 layers, temperature + salinity)
• coord.npy — Normalized coordinates (235160, 60, 3), order: [lon, lat, layer]
• colormap/ — Kindlmann colormap tables (various resolutions)

cd visualization
python3 plot_ocean_layers.py    # outputs to figs/
python3 prepare_coords.py       # outputs coord.npy

rendering/

Fast Voronoi cell rasterizer producing 3600x1800 equirectangular PNGs. Uses a pre-computed cell_id_map.npz (pixel → cell index lookup) so each frame is a pure array operation (~0.36s/frame).

File	Purpose
rasterizer.py	RasterRenderer class + Kindlmann LUT loader
render.py	CLI: render single .dat/.npy or batch a directory
cell_id_map.npz	Pre-computed (1800, 3600) int32 raster; -1 = land

cd rendering

# Render a single pre-normalized .dat (values in [-1, 1])
python3 render.py single path/to/0042.dat -o output.png

# Render a raw .npy (applies mask + normalization internally)
python3 render.py single path/to/0042.npy -o output.png --raw

# Batch render all .dat files in a directory
python3 render.py batch path/to/ens_dir/ -o output_dir/

Pipeline (on remote server)

The full paper-figure pipeline runs on the compute server:

ocean.nc (mesh topology)  →  cell_id_map.npz  (one-time precompute)
raw .npy (per-ensemble)   →  ocean_gt.py       →  .dat (masked + normalized)
.dat (per-method infer)   →  ocean_render.py   →  .png (3600x1800)
per-method .png           →  gen_compare.py    →  comparison grid
comparison grid           →  pick_compare.py   →  zoom boxes + crops

benchmarks/

Ablation results comparing DiT configurations (plane_size x DiT_size x patch_size, each S/M/L).

Script	Purpose
plot_benchmarks.py	Generate grouped bar charts from results.csv

Output directories: psnr/, train_time/, model_size/ — each contains overview + grouped-by-{patch, plane, dit} plots.

cd benchmarks
python3 plot_benchmarks.py