Physics-informed reduced modeling reveals the dominant cross-plane thermal bottleneck in magnetic tunnel junctions

Localized self-heating in nanoscale magnetic tunnel junctions (MTJs) critically affects the thermal reliability of spintronic devices, yet high-fidelity three-dimensional finite-element method (3D-FEM) simulations remain computationally expensive for rapid design exploration. Here, we develop a physics-informed dimensional-reduction framework that reformulates the three-dimensional heat-diffusion problem into a layered two-and-a-half-dimensional thermal network. An effective convection coefficient, m, is extracted through model-constrained parameter inversion, rather than prescribed empirically, to characterize interlayer heat transfer in MTJs. The reduced model retains nonuniform in-plane heat diffusion while representing interlayer thermal transport in a compact lumped form. Compared with 3D-FEM simulations, the model achieves comparable accuracy, with a hotspot error below 0.76% and an absolute temperature difference of 0.45 K, while achieving an approximately three-orders-of-magnitude speedup in post-calibration forward inference. The parameters inverted under steady-state conditions are further transferable to transient temperature-rise prediction under write-pulse excitation. These results provide an efficient and physically interpretable route for cross-plane thermal characterization and thermally aware design of MTJ devices.