Improving failure rates in pulsed SOT-MRAM switching by reinforcement learning

Abstract

Finding and optimizing robust schemes for field-free switching remains a challenging problem in spin-orbit torque magnetoresistive random access memories. In this work reinforcement learning is employed for the optimization of switching schemes for such memory cells. A cell is switched purely electrically by applying pulses to two orthogonal metal wires. It is shown that a neural network model trained on a fixed material parameter set is suitable to determine optimal pulse sequences for reliable switching in the presence of thermal fluctuations, material parameter variations and reduction of the current to a sub-critical value. Multiple realizations of switching by means of simulation prove the reliability of magnetization reversal based on the pulse sequences found via reinforcement learning and show that the failure rate due to material parameter variations in these memory devices can be significantly reduced.