Distance-insensitive Graph Neural Networks and multi-task learning for accurate prediction of adsorption energies on alloy nanoclusters

Developing methodologies for accurate yet computationally efficient prediction of adsorption energies is important for accelerating alloy catalyst screening. However, current data-driven approaches still face challenges when applied to finite alloy nanoclusters, where domain-matched adsorption-energy datasets remain limited and robust screening methods that do not strongly rely on fully relaxed geometries are desirable. To address these issues, we developed a distance-insensitive, adsorbate-oriented orbital graph neural network model (BE-OGCNN) that incorporates orbital interaction features without using explicit interatomic distance or angle descriptors. On a 44-atom alloy-cluster benchmark, the proposed model achieved a mean absolute error of 0.042 eV, outperforming the other distance-insensitive architectures examined here and achieving accuracy comparable to distance-sensitive reference models without using explicit geometric descriptors. We further examined whether bulk-derived auxiliary labels can improve adsorption-energy prediction through multi-task learning with d-band center and total energy auxiliary tasks. The multi-task approach yielded consistent improvement for larger 85-atom clusters, suggesting that physically motivated auxiliary tasks may improve data efficiency in adsorption-energy prediction for alloy nanoparticles.