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

Koki Otsuka; Anh Khoa Augustin Lu; Koji Shimizu; Satoshi Watanabe

doi:10.1016/j.commatsci.2026.114987

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

;

Download file (1.16 MB)
Download as zip (848 KB)

Collection

Citation

Koki Otsuka, Anh Khoa Augustin Lu, Koji Shimizu, Satoshi Watanabe. Distance-insensitive Graph Neural Networks and multi-task learning for accurate prediction of adsorption energies on alloy nanoclusters. Computational Materials Science. 2026, 274 (), 114987. https://doi.org/10.1016/j.commatsci.2026.114987

(BibTeX)

Description:

(abstract)

Developing a methodology that enables accurate yet computationally efficient
prediction of adsorption energies is pivotal for accelerating the discovery of high-
performance alloy catalysts. However, current data-driven approaches face significant
challenges, particularly the scarcity of high-accuracy distance-insensitive machine-
learning models suitable for screening unknown structures and the limited structural
diversity of available adsorption energy datasets. To address these challenges, we
developed an enhanced distance-insensitive graph neural network model, named
Bond-type Embedded Orbital Graph Convolutional Neural Network (BE-OGCNN), that
integrates orbital interaction features to maximize expressivity without geometric
dependency. In addition, we employed a multi-task learning framework using d-band
center and total energy as auxiliary tasks. This strategy overcomes data scarcity by
effectively exploiting abundant bulk crystal data that was previously underutilized for
surface property prediction. Our model achieved a mean absolute error of 0.042 eV on
44-atom alloy clusters, demonstrating accuracy comparable to that of a state-of-the-art
distance-sensitive model. Moreover, the multi-task approach successfully improved
prediction accuracy on larger 85-atom clusters, suggesting high potential of our
framework for rapid and reliable screening of realistic catalyst nanoparticles.

Rights: