AI-Guided Codesign Framework for Novel Material and Device Design applied to MTJ-based True Random Number Generators

TL;DR

This work investigates the design and optimization of spin orbit torque and spin transfer torque magnetic tunnel junction models as the probabilistic devices for true random number generation.

Abstract

Novel devices and novel computing paradigms are key for energy efficient, performant future computing systems. However, designing devices for new applications is often time consuming and tedious. Here, we investigate the design and optimization of spin orbit torque and spin transfer torque magnetic tunnel junction models as the probabilistic devices for true random number generation. We leverage reinforcement learning and evolutionary optimization to vary key device and material properties of the various device models for stochastic operation. Our AI guided codesign methods generated different candidate devices capable of generating stochastic samples for a desired probability distribution, while also minimizing energy usage for the devices.

Figures (10)

• Figure 1: Our AI-Guided Framework for Device Discovery and Optimization for a given application. Overview of the device model, AI-guided discovery and optimization strategy, and RNG algorithm workflow. Given a target distribution, the optimization approach (b) uses a device model (a) to simulate a true random bit according to the RNG algorithm (c). The optimization algorithm designs novel device configurations (d) that must pass device checks to be viable. The viable devices are used to produce the target distribution for a given application (e).
• Figure 2: Overview of tree algorithm. Truncated target distribution (a) is used in the tree algorithm (b) to optimize the MTJ devices to match the desired distribution.
• Figure 3: Schematic illustrations of the modeled SOT-MTJ (a--c) and STT-MTJ (d--f) with perpendicular magnetic anisotropy. (a) The SOT charge current ($J_{\text{SOT}}$) is applied through terminal T3 to T1 to rotate the free layer (FL) in-plane, and the change of magnetoresistance is read through an MTJ via terminals T2 and T1 after $J_{\text{SOT}}$ is removed. An additional STT charge current $J_{\text{STT}}$ between T1 and T2 provides biasing of the coin. (d) The current ($J_{\text{STT}}$) induces a stochastic switching of the free layer, which is reset with $J_{\text{reset}}$ after each read operation. (b, e) Example device S-curve showing how STT current amplitude ($J_{\text{STT}}$) between T1 and T2 biases the bit probability for both device types. (c, f) Two example random bitstreams generated using the simulation model maicke2023magnetic, with accompanying pulsed SOT or STT currents in units of mA.
• Figure 4: RL device optimization results. (a, b) PDF comparison of top 5 device configurations, best configuration, PRNG, and target distribution for both SOT and STT devices. (c, d) Parameter configurations of top 5 devices with energy and KL divergence metrics compared against the default configurations and a PRNG for both SOT and STT devices. (e, f) Pareto fronts comparing energy and KL divergence metrics of various SOT and STT device configurations. (g, h) Probability distributions of the parameter ranges that were explored for both SOT and STT devices.
• Figure 5: EA device optimization results. (a, b) PDF comparison of top 5 device configurations, best configuration, PRNG, and target distribution for both SOT and STT devices. (c, d) Parameter configurations of top 5 devices with energy and KL divergence metrics compared against the default configurations and a PRNG for both SOT and STT devices. (e, f) Pareto fronts comparing energy and KL divergence metrics of various SOT and STT device configurations. (g, h) Probability distributions of the parameter ranges that were explored for both SOT and STT devices.