Dataset and publication: Modelling magnetic material properties with uncertainty-aware neural networks (GNN coercivity study)

Figures and article of the graph-neural-network study (Section 5) of "Modelling magnetic material properties with uncertainty-aware neural networks" (Computational Materials Science 276, 115073, 2027, https://doi.org/10.1016/j.commatsci.2026.115073). Hc_test_total_unc_hexbin.png is Fig. 18A: parity and residual plot of the test set for the representative seed 123, coloured by the total predictive uncertainty. confidence_curve_Hc_all_samples.png is Fig. 18B: confidence curves of the mean absolute error over ten seeds. Wager_2027_CMS.pdf is the published article (open access, CC BY 4.0, © 2026 The Authors). test_predictions_all_seeds.csv contains the per-sample data behind both figures (10 seeds x 316 test samples = 3160 rows; Fig. 18A uses the rows of seed 123, Fig. 18B uses all rows). Columns: seed; dataset_index = position in the graph list built from the PKL files by scaler_helper.get_dataset (batches in the order edge_length_var, 1_V3, 2_V3, 3_V3, bigger, bigger2, fisch_2; one graph with fewer than two grains is removed); pkl_batch and pkl_index = batch and position of the graph in the PKL files; y_true and y_pred = coercivity μ₀H_c in T; std_total, std_aleatoric and std_epistemic = predictive uncertainties in T (total = square root of the sum of the squared aleatoric and epistemic parts). The values were obtained by re-evaluating the trained networks of the paper (seeds 123 to 132, weights in https://github.com/wagerc97/magnetic-uncertainty) on the test split of each seed with 50 Monte Carlo dropout passes, rounded to six decimals. They agree with the published figures within the dropout noise (about 0.0002 in R² and MAE) but are not bit-identical to the PNG files. plot_figures_from_csv.zip contains a script that recreates Fig. 18A and 18B from this CSV file (plotting modules by C. Wager, MIT license; usage and requirements are described in the README inside the archive). The figures were produced with the code archived at https://doi.org/10.5281/zenodo.23545084 (script reproduce_fig18.py). The training data (1054 synthetic microstructures as graphs, folder PKL/) are archived with Micromagnetic_GNN at https://doi.org/10.5281/zenodo.23360688 and are not repeated here.