TRIADS is a parameter-efficient recursive attention architecture for materials property prediction in small-data regimes. By combining weight-tied recursive reasoning, per-cycle deep supervision, and physics-informed features, TRIADS achieves strong performance across six Matbench benchmarks with fewer than 250K parameters. Results include 0.9655 ROC-AUC on matbench expt is metal (44K–100K parameters), 0.3068 eV MAE on matbench expt gap, 35.89 meV/atom on matbench jdft2d, and 41.91 cm⁻¹ on matbench phonons without external pretraining. Controlled ablations show that deep supervision reduces MAE by 23.3% under identical architecture, highlighting the importance of architecture-coupled training in small-data settings.
Rudra Tiwari (Tue,) studied this question.