Topological magnetic structure generation using VAE-GAN hybrid model and discriminator-driven latent sampling

Area of Science:

• Materials Science
• Computational Physics
• Machine Learning

Background:

• Deep generative models are increasingly used for scientific data generation.
• Generating diverse magnetic topological structures is challenging due to energy barriers.
• Existing methods struggle to produce varied and valid spin structures.

Purpose of the Study:

• To develop a hybrid variational autoencoder (VAE) and generative adversarial network (GAN) model for generating 2D magnetic topological structures.
• To leverage the strengths of both VAEs (diversity) and GANs (fidelity) for scientific data generation.
• To explore the application of discriminator-driven latent sampling (DDLS) for enhancing generated sample quality.

Main Methods:

• Implementation of a hybrid VAE-GAN model.
• Utilizing the model to generate two-dimensional magnetic topological structure data.
• Applying discriminator-driven latent sampling (DDLS) to refine generated samples.

Main Results:

• The VAE-GAN hybrid model successfully generated a variety of plausible 2D magnetic topological structures.
• The model demonstrated an ability to overcome energy and topological barriers.
• Discriminator-driven latent sampling (DDLS) improved the quality and coverage of generated data, adhering to topological rules.

Conclusions:

• The VAE-GAN hybrid model provides an effective approach for generating diverse and valid magnetic topological structures.
• The method facilitates applications such as searching for desired samples within a large set of valid structures.
• DDLS is a valuable technique for enhancing the fidelity and coverage of generated scientific data.