A High-Throughput Workflow for Identifying Spin-Crossover Complexes via Artificial Intelligence

Experimental and computational design of spin-switching materials is cumbersome and expensive because of the vast combinatorial chemical space provided by the choice of ligands coordinating a metallic center [Chem. Rev. 121, 9927 (2021)]. Development of tools for the swift elucidation of these intricate relationships thus is key to discovering promising material candidates. To this end, we compiled a data set of 1,436 medium-sized metal-organic complexes, extracted from the Cambridge Structural Database [Acta Crystallogr. B: Struct. Sci. Cryst. Eng. Mater. 72, 171 (2016)], and equipped with the spin-crossover energy computed with the r²SCAN meta-generalized gradient density functional approximation [J. Phys. Chem. Lett. 11, 8208 (2020)]. Our study describes the development of a high-throughput workflow based on an equivariant graph neural network [arXiv:2102.09844v3] for efficient identification of potential spin-crossover candidates. We show that our model increases the chances for finding metal-organic complexes of interest by a factor of four with only 915 trainable parameters. Further, we provide evidence that the model achieves that gain by learning the importance of the coordination shells surrounding the metallic core of the complex. Finally, we test our model on a larger data set with over eleven thousand complexes and discuss the importance of selecting distinct thresholds for the spin-switching energy.