{
  "train_dir": "/scratch/jtb3sud/processed_elasto_plastic/global_max/normalized/small/train",
  "val_dir": "/scratch/jtb3sud/processed_elasto_plastic/global_max/normalized/small/val",
  "file_pattern": "*.pt",
  "seq_len": 16,
  "stride": 16,
  "num_static_feats": 2,
  "num_dynamic_feats": 2,
  "hidden_dim": 128,
  "processor_size": 4,
  "num_harmonics": 5,
  "num_layers_node_processor": 2,
  "num_layers_edge_processor": 2,
  "num_layers_edge_encoder": 2,
  "num_layers_node_decoder": 2,
  "aggregation": "sum",
  "mlp_activation": "relu",
  "mask_eroding": true,
  "use_loss_decay": false,
  "loss_decay_gamma": 0.9,
  "ss_schedule": "linear",
  "ss_initial_ratio": 0.0,
  "ss_final_ratio": 0.0,
  "epochs": 1500,
  "lr": 5e-05,
  "grad_clip_norm": 1.0,
  "device": "cuda",
  "num_workers": 4,
  "output_dir": "/scratch/jtb3sud/delta/elasto",
  "resume": null,
  "reset_best": false,
  "fresh_scheduler": false,
  "architecture": "MeshGraphKAN (NVIDIA PhysicsNeMo, PyG reimplementation)",
  "kan_implementation": "Learnable Fourier coefficients [2, out, in, harmonics] with einsum",
  "normalization": "global_max",
  "input_dim_nodes": 4,
  "input_dim_edges": 3,
  "references": [
    "Pfaff et al., Learning Mesh-Based Simulation with Graph Networks, 2021",
    "Liu et al., KAN: Kolmogorov-Arnold Networks, 2024",
    "Peng et al., Interpretable physics-informed GNNs for flood forecasting, 2024",
    "NVIDIA PhysicsNeMo: github.com/NVIDIA/physicsnemo"
  ]
}