PermaCycleGNS v1.0.0 — code, data, and trained models for "Nonlinearly Coupled Permafrost Dynamics From Graph Neural Networks

Code, datasets, and trained model checkpoints accompanying Liu, You, Ren & Tang, "Nonlinearly Coupled Permafrost Dynamics From Graph Neural Networks" (submitted). PermaCycleGNS is a Graph Network-based Simulator (GNS) surrogate for the 1-D thermo-hydro-chemical evolution of Arctic subsea permafrost. From an initial state, it autoregressively predicts 400 kyr of jointly evolving pore pressure, temperature, salinity, and ice saturation over a 51-node depth column, 8,800–200,000× faster than the physics-based simulator PermaCycle.

Contents: (i) PermaCycleGNS_v1.0.0_src — Python/PyTorch source matching GitHub tag v1.0.0; (ii) data — training/validation/test sets for the random split (Dr, 661/138/142 cases) and corner-extrapolation split (De, 662/144/135), as normalized .npz with metadata; (iii) models — the two headline checkpoints (multistep K=1→6 curriculum, step 10⁶), one per split; (iv) results — per-variable evaluation summaries reproducing the paper's tables.

The physics-based simulator PermaCycle (You, 2026) that generated the training data is not included; it is archived separately at doi:10.5281/zenodo.20043352. The GNS architecture adapts EQGNS (Liu & Becker, 2025, doi:10.1029/2025JB031981).

License: MIT. Reusers of the data should cite both this archive and PermaCycle.