ML-Driven Compact Models for RRAMs: Addressing Variability and Simulation Efficiency

Abstract

Machine Learning (ML)-based compact modeling provides a promising alternative to traditional physics-based methods, enabling faster development of compact models for novel devices while offering improved predictive performance. For Resistive Random Access Memory (RRAM) devices, several ML-based compact models have been developed. However, these models often face two key challenges: they fail to capture stochastic cycle-to-cycle variations effectively, and they are difficult to accurately convert into Verilog-A models for SPICE simulations. To address these challenges, we propose a novel variation-aware ML-based compact model for RRAM, using modified deep ensemble techniques to account for cycle-to-cycle variations and model uncertainty, along with a newly designed state determination function to accurately capture resistive switching characteristics. Furthermore, by introducing knowledge distillation combined with a pruning-retraining process, the proposed model achieves a 67% reduction in simulation turnaround time while maintaining predictive accuracy, ensuring strong compatibility with SPICE simulations.