Machine learning assisted prediction of perpendicular Magnetic Tunnel Junction (pMTJ) final state

Application of machine learning in studying micromagnetic structure of magnetic materials has become of great interest in device optimization in spintronic systems [1]. In particular, magnetic tunnel junctions have been widely used and studied for magnetic memory and sensing applications [2]. In this work, we investigate which magnetic features affect junction switching using linear and non-linear machine learning algorithms. Various magnetic structures are achieved via spin transfer torque mechanism in a 32nm-dimeter pMTJ disk. A current pulse of amplitude slightly above the critical current for switching the junction was applied in the normal to the disk to induce an initial magnetic state. Resulting datasets were analyzed by one linear (linear regression) and two non-linear (XGBoost and Random Forest) machine learning algorithms. These algorithms were trained using 5-fold cross-validation and repeated k-fold cross-validation with k=10 to ensure robust model evaluation and reduce overfitting by testing performance across multiple data partitions. We will present on the prediction accuracy of the three models and identify the most important magnetic parameters that determine the final state of the junction. Our results show the importance of applying nonlinear modeling in interpreting micromagnetic data and provide insight into the characteristics that determine junction switching in magnetic materials.

[1] I. Labrie-Boulay et al., Phys. Rev. Applied 21 (2024)

[2] A. D. Kent et al., Nature 10 (2015)