Artificial Intelligence-Driven Magnetic Property Prediction and Materials Discovery for Next-Generation Spintronics

The review establishes a unified framework linking electronic structure to descriptors and then to machine-learning models, arguing that the electronic-structure origins of magnetism (d/f orbital occupancy, exchange splitting, d–d and p–d hybridization, spin–orbit coupling, Dzyaloshinskii–Moriya interaction) should serve as the physical basis for predictive modeling, and proposes an extended PSPP framework that inserts electronic structure as a mediating layer into the conventional processing–structure–performance chain. Compared with prior data-driven materials-informatics reviews, this work emphasizes incorporating electronic structure as a mediating layer and systematically organizes compositional, structural, electronic, and topological descriptors. A conceptual framework and literature synthesis based on existing theory and methods, without new experimental or computational data.

The review systematically compares regression, ensemble learning (random forest, gradient boosting), deep learning, graph neural networks, and physics-informed learning for magnetic property prediction, concluding that no single model universally dominates and that model choice should match data scale, descriptor quality, and the physics of the target property. By aggregating quantitative metrics from multiple studies (e.g., Curie temperature prediction R²≈0.92, MAE≈40 K; spin Hall conductivity MAE≈115 (ℏ/e) S cm⁻¹, AUC≈0.86; magnetocrystalline anisotropy energy RMSE≈0.22–0.25 meV), it provides a cross-model, cross-property comparison. Based on aggregation and comparison of quantitative metrics reported in multiple independent studies; the metrics come from the original works and the review itself does not perform a unified benchmark.

The review summarizes AI discovery progress in specific spintronic material classes: in Heusler alloys, random forest and gradient boosting can predict magnetic moment, Curie temperature, and half-metallicity and identify Slater–Pauling trends; in topological materials, XGBoost, graph neural networks, and generative inverse design can screen topological insulators and semimetals; in two-dimensional magnets, descriptor engineering, active learning, and graph neural networks can discover high spin Hall conductivity and magnetic topological candidates. It integrates AI discovery cases scattered across material classes into a unified descriptor-based paradigm and distills design principles such as heavy-element substitution enhancing spin–orbit coupling, symmetry breaking generating Berry curvature and DMI, and magnetic ordering enabling quantum anomalous transport. A review-based synthesis of multiple independent case studies, including specific material examples (e.g., Ta₃P, Ta₃As, CdAu₅, Li₂YBi₂, Cr₂NF₂, Os₂Cl₄) and reported performance metrics.

The review proposes a future closed-loop AI roadmap, advocating the integration of physics-informed machine learning, explainable AI, generative inverse design, multiscale simulation, uncertainty quantification, and autonomous experimental platforms, advanced over short-term (1–3 years), mid-term (3–7 years), and long-term (7–10+ years) horizons toward fully autonomous discovery ecosystems. Compared with prior reviews focused on single prediction tasks, this work provides a time-phased development roadmap comprising five stages: data generation, descriptor engineering, model development, inverse design, and experimental validation. A forward-looking roadmap and perspective based on extrapolation from current trends, not empirical results.

This review targets researchers in spintronics, magnetic materials, and materials informatics, and applies to scenarios of magnetic property prediction and materials screening at the level of compositional, structural, and electronic descriptors; its proposed closed-loop roadmap addresses research environments combining high-throughput computation with autonomous experimental platforms.

The review notes that most models are trained on DFT data assuming zero temperature, perfect crystallinity, and collinear magnetic configurations, inadequately capturing finite-temperature magnetic fluctuations, disorder, strong correlations, and non-collinear magnetism, so predicted Curie temperatures and magnetic anisotropies often deviate from experiment; there is also a trade-off between predictive accuracy and physical interpretability, and limited extrapolation across chemical space. In addition, as a review text it provides no unified benchmark, and performance metrics come from different original studies, so readers should note differences in data sources and evaluation conditions when comparing.