Graph BSP

CPU inference package for reproducing the backward-facing step figures used in the paper. The notebook inference.ipynb loads the saved Base, Chebyshev BSP, and GLEAM checkpoints, reuses the cropped BFS data in data/BFS, and displays the paper figures inline.

Contents

• inference.ipynb: paper-focused CPU inference notebook with embedded outputs.
• saved_trained_models/: saved Base, Chebyshev BSP, and GLEAM checkpoints.
• data/BFS/: cropped BFS graph, fields, mesh cells, and normalization stats.
• train.py: Base BFS forecaster training entry point.
• train_chebyshev.py: Chebyshev BSP training entry point.
• train_gleam_multilevel.py: GLEAM multilevel training entry point.
• gleam_multilevel_loss.py: GLEAM low-rank graph-spectral loss.

Notes

Several required .pt files are large. If GitHub rejects them, track them with Git LFS or host them as release assets. The raw CG2 BFS tensor is intentionally not included because the cropped CG1 inference workflow does not use it.