Predicting extreme events from data using deep machine learning: when and where

The paper clearly defines the when/where labeling, datasets, CGLE setup, and reports consistent headline metrics (e.g., CGLE AUC≈0.91/0.94 at |u|=6, p=10; AUC≈0.72/0.82 at |u|=5, p=20; 5×5 top-4≈98.37% with ResNet-50; wind 2×2 where≈92.8%) and provides enough numerical detail to be broadly reproducible . However, methodological choices like epoch selection using test-set AUC and mixed temporal splits for wind leave minor gaps in experimental rigor, and Appendix B’s ROC exposition blurs whether the score is the DCNN output or a simple intensity threshold . The model’s submission is expressly incomplete (no experiments), and while its protocol largely matches the paper, it diverges in a few key places (e.g., wind region and temporal split; handling overlapping events in where-labeling), so it cannot validate the reported targets.