ChelatorDiff: Energy-Aware 3D Chelator Complex Generation for F-block Metals

The accurate generation of three-dimensional metal complexes is a key enabling capability across many applications ranging from catalysis, energy conversion and storage, cancer therapeutics, to rare earth processing. This task is particularly crucial in modern discovery workflows where vast chemical space can be rapidly screened for the target properties using high-throughput computational methods. Yet, reliably generating such structures remains a formidable challenge, as these occupy a complex high-dimensional configurational landscape in which metal coordination environment, ligand conformations, molecular connectivity, and inter-ligand packing must all be consistently determined. This challenge is particularly acute for f-block complexes, which exhibit high coordination numbers, high-denticity ligands, and often explicit inner-sphere solvation. We introduce ChelatorDiff, a generative framework that addresses this challenge by decomposing complex construction into chemically distinct stages. ChelatorDiff combines learned metal–donor coordination refinement, conformer-aware and geometry-preserving ligand placement, conditional inner-sphere water generation, and sanity-aware repair and candidate selection. Across a diverse corpus of lanthanide and actinide complexes, this structured treatment proves especially effective for high-denticity ligands, where competing approaches struggle to place multiple donor atoms coherently. ChelatorDiff also achieves substantially higher build reliability than direct exhaustive construction approaches while reducing median recorded wall time by a factor of approximately four. Subsequent geometric refinement brings generated structures close to putative local energetic minima. These results show that explicitly preserving known chemical structure while learning the remaining geometric degrees of freedom provides a robust route to efficient high-throughput generation of complex f-block coordination structures.