Data-driven prediction of single-molecule magnets through spatial-electronic cooperation

The fast and precise prediction of single-molecule magnets (SMMs) is important for discovering applicable magnets for quantum computing, ultra-high-density storage, etc. However, the trial-and-error synthesis and limited mechanistic understanding have made the development of high-performance SMMs even more challenging, reflecting that only 1.2% of the complexes have an effective energy barrier (Ueff) over 600 cm-1 in the "SIMDAVIS" dataset. Herein, a universal magneto-structural relationship was established to predict Ueff barriers, the scope of which covers hundreds of Dy-SMMs with diverse ligands and varied coordination atoms. These efficient predictions rely on new spatial descriptors θ and the designed theoretical dataset through machine learning methods and multi-scale calculations, including the full magnetic relaxation process of 132 molecular structures at the CASSCF level of theory. This work provides high-throughput theoretical screening for high-performance SMMs, validated by not only 391 previous Dy-SMMs but also the 44 recently synthesized complexes. Our descriptors and theoretical dataset expand a useful interface for the intensive involvement of artificial intelligence in the field of SMMs.