Machine Learning Prediction for Fe(II) Spin-Crossover Complex in the Same Spin State Using Geometrical and Topological Descriptors

Spin-crossover (SCO) complexes are molecular materials capable of reversibly switching between high-spin (HS) and low-spin (LS) states in response to external stimuli. Predicting SCO from crystallographic data allows for the efficient design of new SCO complexes. However, no extensive and diverse data sets of complexes with annotated SCO characteristics were available. In the present study, we manually assembled a data set of 500 Fe(II)-N₆ coordination complexes with explicit spin states and SCO potential, termed FeN₆-SSD. Using this data set, we built machine learning models to distinguish between SCO-undergoing and non-SCO complexes crystallized in the same spin state (either HS or LS). The classification results showed that the key factors for predicting SCO activity differed between the two spin states: in the HS regime, local geometric distortions, such as Fe-N bond elongation and octahedral deformation, were important, whereas in the LS regime, ligand-derived chemical and steric factors were important. Overall, the many-body tensor representation as a descriptor set achieved high prediction accuracy in both spin states. The influence of environmental factors, such as solvents and counterions, on SCO classification was inconsistent across representations.