<p>Training hyperparameters for HP-GNN with selection methods and rationale. Values were determined through systematic optimization including grid search, Bayesian optimization, and ablation studies on validation data. The physics learning rate is intentionally scaled to 0.1 × the main rate to ensure stable convergence of Kuramoto parameters.</p>

Training hyperparameters for HP-GNN with selection methods and rationale. Values were determined through systematic optimization including grid search, Bayesian optimization, and ablation studies on validation data. The physics learning rate is intentionally scaled to 0.1 × the main rate to ensure stable convergence of Kuramoto parameters.