Defining skyrmion phase boundaries and predicting the phase diagram among various spin textures remain critical challenges in the field of two-dimensional magnets. In this work, we combine atomistic spin simulations with machine learning (ML) to identify skyrmion phase boundaries across multiple parameter spaces, including frustration interaction $\eta$ $(|J_2/J_1|)$, Dzyaloshinskii-Moriya interaction $\left(D\right)$, single-ion anisotropy $\left(K\right)$, magnetic field $\left(B\right)$, and temperature $\left(T\right)$. Using convolutional auto-encoder, t-distributed stochastic neighbor embedding, $K$-means, and Inception-V3 models, we classify three distinct phases (ferromagnet, skyrmion, and mixed phase) with $98.2\%$ accuracy. A Bayesian active learning framework incorporating Gaussian process regression iteratively refines phase boundary predictions, achieving stable errors around $\sim 0.05$. In addition, phase boundary fitting reveals a functional relationship, $\lambda \sim \eta K/D^2$, linking magnetic parameters to skyrmion stability. These findings demonstrate the potential of ML to enhance the efficiency and precision of skyrmion phase boundary analysis, advancing the understanding of complex spintronic systems.

Close

PrevNext

View figure in article

• Magnetic anisotropy
• Magnetic phase transitions
• Phase diagrams
• Skyrmions
• Spin texture
• Magnetic systems
• Bayesian networks
• Convolutional neural networks
• Heisenberg model
• Landau-Lifschitz-Gilbert equation
• Machine learning
• Micromagnetic modeling
• Unsupervised learning