Machine learning assisted prediction of dynamics in current-driven nested skyrmion bags

Abstract

Nested skyrmion bags are topological magnetic structures with tunable topological charge, offering potential for spintronic applications. Predicting their current-driven dynamics, particularly the skyrmion Hall angle, across diverse structural and material parameters remains challenging due to the complexity of the underlying physics. Here we show, through micromagnetic simulations validated by Thiele equation analysis, that zero-topological-charge bags move linearly without transverse deflection, while non-zero-charge configurations exhibit a widely tunable Hall angle. Transverse elongation at high nesting levels can be suppressed by additional domain wall layers. To enable rapid prediction, we implement twelve machine learning models, among which gradient boosting methods and neural networks achieve high accuracy, whereas linear regression fails, confirming the inherent nonlinearity of the system. Leveraging this predictive capability, we demonstrate a demultiplexer device that routes information based on the Hall angle. This work provides a framework for designing topology-based spintronic devices such as racetrack memory and signal routers.