Single-molecule magnets (SMMs), as a class of advanced molecular materials with immense application potential in high-density magnetic storage, spintronics and quantum information science, have attracted extensive attention in recent years. Over the past three decades, experimental synthesis of novel SMMs has made remarkable progress: the effective magnetic reversal barrier (U_eff) and blocking temperature (T_B)—the two key figures of merit characterizing SMM performance—have been pushed from the initial values of 42 cm^-1 and 4 K to the current records of 1843 cm^-1 and 100 K, respectively. Despite these advances, the practical device-level implementation of SMMs remains highly challenging, most notably because the blocking temperatures achieved so far still fall far short of the requirements for robust real-world operation.
From the theoretical perspective, ab initio quantum chemical calculations play a central role in the study of SMMs by enabling a microscopic understanding of magnetic relaxation mechanisms and the establishment of quantitative structure- magnetism relationships. However, an accurate description of molecular magnetism requires high-level electronic structure methods that can treat strong electron correlation and relativistic effects with high fidelity, which leads to prohibitive computational costs, especially for large or multi-nuclear systems and for simulations that must probe long-time spin dynamics. In this context, machine learning (ML) offers a powerful complementary route to bridge the gap between accuracy and effciency.
This review provides a comprehensive account of ML methods in the theoretical study of SMMs along the entire chain from data to models to applications. We first introduce the basic concepts of SMMs, the physical picture of magnetic relaxation, effective spin Hamiltonians and their parametrization, and the current status of ab initio approaches to SMMs, including recent developments in first-principles spin- phonon coupling and spin relaxation theories. Building on this foundation, we then systematically survey ML applications in SMM research from five interrelated perspectives: (i) construction and curation of SMM-related datasets, such as databases of lanthanide complexes derived from experimental measurements and high-throughput quantum chemistry calculations; (ii) development of ML force fields tailored for molecular magnetic systems, exemplified by SNAP-type potentials and related symmetry-adapted representations that retain rotational, translational and permutational invariances while reaching near- ab initio accuracy; (iii) ML models for predicting key magnetic quantities—including magnetic anisotropy tensors, zero-field splitting parameters and magnetic exchange couplings—using physically informed descriptors and, in particular, covariant ML architectures that respect the tensorial transformation properties under molecular rotations; (iv) ML-assisted spin dynamics simulations, in which ML force fields and ML-predicted spin Hamiltonian parameters are combined to effciently evaluate spin- phonon coupling and relaxation times over extended time scales; and (v) ML-driven SMM design strategies that integrate high-throughput screening, automated quantum-chemical workflows and data-driven optimization in large chemical spaces, including frameworks that build open databases and couple them with ML models to identify promising high-performance SMM candidates.
Across these topics, we emphasize the core methodological ideas and emerging best practices: the design of symmetry-adapted molecular representations for magnetic systems, the strategy to reduce the cost of generating high-quality training data, and the construction of closed loop that link electronic structure calculations, ML models and screening or design tasks. We conclude by summarizing the current opportunities and outstanding challenges for ML in SMM research, including data scarcity, transferability and interpretability of models, and the extension from single-ion to multinuclear systems. We also highlight the promising outlook for ML-augmented spin dynamics as enabling technologies for the rational discovery and design of next-generation, high-performance SMMs.