Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Final model — May-22 PCA–U-Net predictions

final_model_unet.zip contains predictions for all 2,000 full-domain May-22 test samples, with no added measurement noise.

Unzip and open the extracted folder in MATLAB. Each sample file contains its coordinates and predictions; no separate grid file is needed.

P = load('test_sample142.mat');
plot_prediction_points('test_sample142.mat');
plot_prediction_volume('test_sample142.mat');

The plots hide air by default. Pass false as the second argument to show the whole grid. Only base MATLAB is needed.

Contents

  • P.points: 131072-by-3 coordinates [x,y,z], in metres.
  • P.values: 131072-by-1 predicted conductivity, in S/m, matched row for row with P.points.
  • P.prediction: 64-by-32-by-64 conductivity array, indexed (z,y,x).
  • P.solid_mask: the same shape as P.prediction; true at or below the peat surface.
  • P.xq, P.yq, P.zq: grid axes. Coordinates increase along each axis; z increases upward.

The point order matches xyz_points from extract_xyz_grid: y varies fastest, then x, then z. The 3D array matches uniform_sigma_grid. No reordering is needed when that code uses the same grid settings.

Dataset and domain

These predictions use the May-22, 2026 COMSOL test set. The prediction network was trained from random initialization with L2 + Dice + Chamfer losses from the beginning, keeping the existing 64-mode PCA decoder fixed. There is no separate fine-tuning stage. Training used Muon with AdamW for the remaining parameters; validation selected epoch 117 of 120. The checkpoint is saved_runs/pca_boundary_from_start_20261103/muon120/best.pt; P.model_run identifies the run.

Samples 1–1000 have one inclusion; samples 1001–2000 have two. Filenames match the original MATLAB sample numbers. The two-inclusion samples were not used for training or validation.

The full regular prediction grid has 64 x-points in [-15,15], 32 y-points in [-7.5,7.5], and 64 z-points in approximately [-10,2.2510135] metres. Use the saved axes for exact coordinates. These are regular-grid values, not values on the original COMSOL mesh nodes.

The training offset of 1e-6 S/m is included in the predicted conductivity. No ground truth, voltages, or model weights are included. The files use MATLAB v5 format. Array correspondence was checked in SciPy; the plotting scripts have not been run in MATLAB here.

Downloads last month
25