PU-learning-guided discovery of synthesizable multiferroic nitride perovskites with altermagnetic order

ABN3-type nitride perovskites offer a rich platform for multifunctional materials but remain synthetically elusive. Here, we develop a machine learning (ML)-guided framework to expand the library of nitride perovskites and identify multiferroic candidates. By integrating positive-unlabeled (PU) learning with crystal graph convolutional neural networks (CGCNN), we screen 1465 ABN3 compositions and predict 96 synthesizable compounds. Further symmetry and magnetic filtering yield 4 altermagnetic ferroelectric (AM-FE) perovskites. Among them, CeCrN3 emerges as a promising candidate, exhibiting a bandgap of 0.30 eV, a spontaneous polarization of 0.59 μC/cm2, a high Curie temperature of 650 K, and a low polarization switching barrier of 53 meV, as confirmed by density functional theory (DFT) calculations. In addition, CeCrN3 demonstrates a pronounced bulk photovoltaic effect (BPVE), with the shift current reaching 44 μA/V2 and an injection current reaching 2.4 × 109 A/V2, both of which reverse upon FE switching. These findings not only advance the understanding of nitride perovskites but also provides a validated ML-DFT framework to guide experimental efforts in realizing novel functional materials.