Introducing a Neural Network Approach to Memristor Dynamics: A Comparative Study with Traditional Compact Models
Abstract
Modeling the switching dynamics of memristive devices poses significant challenges for real-world applications, particularly in achieving long-term operational stability. While conventional compact models are effective for short-term simulations, they fail to capture the degradation effects and complexities associated with extended switching behavior. In this work, we propose a novel framework for forecasting memristor switching series using state-of-the-art deep learning architectures. Experimental data from Au/Ta/ZrO₂(Y)/TaOx/TiN/Ti-based memristors were used to compare a classical compact model—featuring a linear drift model with ARIMA corrections—against advanced neural networks, including TimesNet, FredFormer, ATFNet, and SparseTSF. Our results demonstrate that deep learning models, particularly TimesNet, significantly improve predictive accuracy and robustness over long-term switching series. This study provides a foundation for integrating deep learning into memristor modeling, paving the way for more reliable and scalable simulations.