Accurate Machine Learning Predictions of Magnetization Barriers and Relaxation Times in Dnh (n = 4–6) Dy(III) Single-Molecule Magnets

Single-molecule magnets (SMMs) represent a frontier in molecular magnetism, offering transformative opportunities for quantum computing, high-density information storage, and spintronic technologies. Among them, lanthanide-based SMMs—particularly Dy(III) systems—exhibit the highest magnetic anisotropy barriers and blocking temperatures, owing to their strong unquenched orbital angular momentum. Despite these advantages, the factors governing magnetic relaxation and blocking temperatures remain poorly understood, as they originate from subtle and highly non-linear structure–property relationships. Furthermore, pseudo-Dnh Dy-based SIMs offer the additional advantage of air stability, making them promising candidates for practical applications. A critical challenge, however, is whether barrier heights and blocking temperatures can be predicted directly from geometry, without relying on computationally demanding multireference ab initio methods which is computationally expensive. In this work, we introduce a machine learning (ML)-driven framework capable of predicting key SMM performance metrics directly from structural descriptors, thereby reducing dependence on high-cost CASSCF calculations. We focused on Dy(III)-based pseudo-Dnh (n = 4, 5, 6) single-ion magnets, where strong axial fields and weak equatorial interactions promote enhanced anisotropy. A comprehensive dataset of 1429 unique structures (256 D4h, 872 D5h, and 299 D6h systems) was curated, spanning a broad chemical and magnetic space: barrier heights (U) from 0–1855 cm⁻¹, axial bond lengths from 1.72–3.31 Å, equatorial bond lengths from 1.37–3.58 Å, axial bond angles from 128.5–180°, and relaxation prefactors (log τ0) from –5.1 to –12.3. Using ML models trained on structural parameters (bond lengths, bond angles) in combination with CASSCF-derived electronic descriptors (LoProp charges, axial crystal field parameter B20, and gZZ), we achieved predictive accuracies exceeding R² = 0.90 log τ0. This work establishes machine learning as a powerful accelerator for SMM discovery, capable of uncovering hidden structure–relaxation relationships, guiding ligand design, and bridging quantum chemical predictions with experimental synthesis. Beyond enabling the rational design of air-stable, high-barrier SIMs, our approach paves the way toward predictive molecular engineering of quantum materials.