Machine-learning-accelerated discovery of synthesizable high-temperature altermagnets with giant spin splitting

Abstract

Altermagnets offer a route to spin-polarized electronic states without macroscopic magnetization, because compensated magnetic order can generate momentum-dependent spin splitting through crystal-symmetry-controlled exchange fields. However, experimentally viable altermagnets combining large spin splitting, thermodynamic stability and high magnetic ordering temperatures remain scarce. Here, we develop a machine-learning-accelerated high-throughput framework to explore the tetragonal AB$_2$C$_2$D compounds. Screening 8640 variants identifies 1347 compensated antiferromagnetic candidates satisfying altermagnetic symmetry. An interpretable XGBoost model trained on first-principles spin-splitting data then isolates 34 low-hull-energy candidates,including four previously reported, with giant non-relativistic spin splittings exceeding 1.5 eV near the Fermi level. Detailed first-principles calculations of the representative RbMn$_2$Te$_2$O confirm a maximum spin splitting of $\sim$1.88 eV with dynamical stability and an estimated Néel temperature of $\sim$390 K. The giant splitting originates from symmetry-locked Mn-sublattice exchange fields amplified by directional Mn-d/Te-p hybridization. Furthermore, we uncover a profound soft-mode-driven structural transition associated with an interlayer dimensionality crossover in SrMn$_2$Te$_2$O, yet the unfolded electronic structure demonstrates that the altermagnetic spin splitting remains robust after lattice reconstruction. Hydrostatic pressure provides an additional tuning route, producing non-monotonic modulation of the spin-split Fermi surface governed by local coordination and orbital hybridization. These results establish tetragonal AB$_2$C$_2$D compounds as a tunable materials platform for stray-field-free spintronic devices and provide a general data-driven strategy for discovering robust giant-splitting altermagnets.