A Generative Framework for Predicting Antiferromagnets.

Area of Science:

• Materials Science
• Condensed Matter Physics
• Computational Chemistry

Background:

• Predicting antiferromagnets (AFMs) is vital for ultrafast spintronics but hindered by complex electron correlations.
• Traditional methods limit systematic exploration of new chemical spaces for AFM discovery.

Purpose of the Study:

• To develop an efficient computational framework for designing and discovering novel antiferromagnetic materials.
• To accelerate the identification of new AFMs for next-generation spintronic applications.

Main Methods:

• Integrated a crystal diffusion variational autoencoder with data augmentation (CDVAE-DA) for candidate generation.
• Employed crystal graph convolutional neural networks (CGCNNs) for high-throughput screening based on formation energy, magnetic moment, and band gap.
• Utilized a genetic algorithm (GA) to optimize structure generation, with density functional theory (DFT) for validation.

Main Results:

• The CDVAE-DA achieved a 90.68% composition validity rate and generated chemically diverse candidates.
• The integrated framework successfully identified three novel AFM semiconductors (MnS, FeO4P, MnO).
• The genetic algorithm significantly improved the efficiency of discovering targeted AFM structures.

Conclusions:

• The developed AI-driven framework establishes a new paradigm for designing antiferromagnets.
• This approach accelerates the discovery of materials crucial for advancing ultrafast spintronic technologies.
• The study highlights the synergy between generative models, machine learning screening, and physics-based validation.