Statistical Modeling of Metal-Oxide RRAM SET/RESET Behavior Using Deep Neural Networks

We propose the use of end-to-end deep neural network (DNN) models to predict the statistical behavior of resistive RAM (RRAM) during SET/RESET programming. We demonstrate that such models may accurately generate the probability distributions describing the final conductance, without the need for fitting parameters by hand (R
²
= 0.983 for SET, R
²
= 0.817 for RESET). We also attempt to connect some of our models’ qualitative behavior to physics-based understanding of RRAM SET/RESET processes. Finally, we describe a use case in which our models serve as key testing infrastructure in functional verification of a multiple-bits-per-cell RRAM controller.