Navigating Magnetic Chiral States with Autoencoder

In this study, we employed Autoencoder, a type of neural network, to interpolate and extrapolate between two distinct magnetic structures: the labyrinth and the skyrmion. The Autoencoder, acting as a computational observer, learned from instances of these magnetic structures, converting each instance into a compact representation known as a latent code. Analyzing this latent code space enabled the generation of novel magnetic structures responsive to external magnetic fields not included in the training dataset. We introduced two algorithms for modifying the latent codes. The first algorithm involved inversion and translation operations within the latent space, preserving the inherent chiral characteristics of the original data. The second algorithm, employing recursive flow with a modification bias, induced topological changes, leading to a diverse array of statistically plausible.



Our research aims to delve into the intersection of artificial intelligence and condensed matter physics, highlighting the potential applications of artificial intelligence technologies in advancing physics research.