Angular relational knowledge distillation of machine learning interatomic potentials for scalable catalyst exploration

State-of-the-art equivariant Graph Neural Networks (GNNs) achieve DFT-level accuracy for molecular simulations but remain computationally prohibitive for high-throughput screening and long-timescale dynamics. Knowledge distillation (KD) offers a promising solution, yet its application to GNNs remains nascent compared to the mature methodologies developed for language and vision models. The few existing approaches rely on atom-wise feature matching and have struggled to achieve the accuracy necessary for capturing the relational physics underlying the potential energy surface (PES). We introduce Angular Relational Knowledge (ARK) distillation, a framework that distills relational knowledge from pretrained GNNs by modeling each interatomic interaction as a relational vector. Through a contrastive objective, ARK guides compact student models to preserve the geometric structure of the teacher’s learned PES rather than isolated node features. On the OC20, OMat24, and SPICE benchmarks, our ARK-trained student consistently outperforms baselines in energy and force prediction, achieving faithful physical knowledge transfer at a fraction of the computational cost. In a practical high-throughput catalyst screening application, the distilled model achieves an 11.9 × acceleration while preserving chemical coherency. In a large-scale oxygen reduction reaction catalyst screening of over 580,329 structures, ARK reduces the computational cost from 59.4 CPU-years to 11.6 GPU-hours while successfully identifying 30 DFT-confirmed catalyst candidates. Notably, the screening reveals a high prevalence of tellurium-containing materials among the top candidates, corroborating recent experimental reports and demonstrating ARK’s capacity for identifying computationally promising candidates in underexplored spaces.