Machine learning-enhanced numerical modeling of modern magnetoresistive memories

Abstract

The continuous down-scaling of semiconductor devices over the past decades has led to higher integration density but also higher standby power consumption due to increased leakage currents. Novolatile memory is a promising solution to this problem. Magnetoresistive random access memory (MRAM) poses to be a suitable non-volatile alternative due to its straightforward architecture and compatibility with CMOS technology. It offers the benefits of high speed and excellent endurance, making it an attractive choice for a variety of applications including IoT and automotive applications, as well as embedded DRAM and last-level cache memory. Effective simulation tools provide crucial insights for designing MRAM devices. The process of understanding how the magnetization changes over time in these devices, involves solving the Landau-Lifshitz-Gilbert (LLG) equation. This equation can be enhanced with additional terms that account for the torque acting on the magnetization, which is essential for MRAM functionality. This work is dedicated to the computational study and machine-learning assisted optimization of MRAM devices. The first part of this work focuses on the study of the finite element method (FEM) based computation of the demagnetizing field, a crucial contribution to the effective field originating from the long-range interaction of the magnetic moments. Computational solutions to remedy the challenges of the open-boundary problem are implemented, and their performance is evaluated. The second part of this thesis introduces novel computational approaches that combine reinforcement learning with micromagnetic simulations for MRAM device optimization. Traditionally, the application of current pulses to switch magnetoresistive memory cells relies on heuristics. However, this work demonstrates the effectiveness of reinforcement learning in MRAM device control. By autonomously interacting with the simulation, a reinforcement learning agent discovers optimal switching pulse sequences and optimizes various objectives, eliminating the need for manual experimentation.This approach offers a promising solution for enhancing the efficiency and effectiveness of MRAM switching. The approach demonstrates that an agent trained on a fixed set of parameters can effectively transfer its knowledge of magnetization dynamics in the free layer to scenarios with varying environmental conditions. It is shown that over a wide range of material parameters, the agent is capable of achieving reversal of the free layer magnetization. Additionally, the approach is extended to SOT-assisted STT-MRAM, and it is shown that by modifying the rewarding strategy the focus of the learned pulse scheme can successfully be shifted towards different objectives. Specifically, optimization for both fast magnetization reversal and energy-efficient switching is performed. By condensing the dynamically applied pulses of the reinforcement learning agent, static pulse sequences are obtained that perform well across a wide parameter range.