Answers are grounded in this paper and cite their sources.
Aydin et al. reported R2 = 0.921 for band gap prediction using hybrid graph neural networks. Reproduction shows this performance is inflated by data leakage from post-DFT proxies (is_metal, is_gap_direct) and embedding contamination. Controlled ablations reveal a genuine GNN gain of 0.08–0.11 in R2, while leakage contributes ∼0.06, lowering honest performance to 0.82–0.86 R2 (MAE 0.28–0.30 eV).
No takes yet. Share an insight, caveat, or question.
Shambhu Bhandari Sharma (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: