A Generative Framework for Predicting Antiferromagnets

Predicting antiferromagnets (AFMs) is crucial for advancing ultrafast spintronics. However, traditional methods are constrained by the complexity of correlated electrons and magnetic ordering, limiting systematic exploration of new chemical spaces. Herein, an AFM design framework integrating a crystal diffusion variational autoencoder is presented with data augmentation (CDVAE-DA), crystal graph convolutional neural networks (CGCNNs), a genetic algorithm (GA), and density functional theory (DFT) validation. CDVAE-DA generates chemically diverse and structurally valid candidates, achieving a composition validity rate of 90.68%. Leveraging transfer learning on an AFM dataset, the CDVAE-DA is biased toward generating AFM structures. The CGCNNs efficiently screen for potential AFMs using three properties: formation energy (FE), total magnetic moment (TMM), and band gap (BG). The GA regulates the direction of structure generation using these properties, while DFT calculations rigorously validate the lattice stability and AFM order of generated structures. Employing this framework with GA yields three AFM semiconductors (MnS, FeO₄P, and MnO) from 2000 generated structures. In contrast, omitting GA identifies two metallic AFMs (LiVO₂, LiFeN) from 5000 structures. This underscores the GA's role in optimizing latent vectors to facilitate efficient AFM generation. This work establishes a design paradigm for AFMs, accelerating next-generation spintronic material discovery.