Semi-recurrent Physics Informed Neural Networks for Modelling Resistive Memories

Accurate modelling of the non-linear and stochastic behaviour of resistive random access memory (RRAM) devices remains a significant challenge, as conventional models often fail to capture the intricate dynamics of resistive switching and stochastic ion migration. In this work, we propose a novel semi-recurrent physics-informed neural network (PINN) framework to model RRAM devices by integrating both data-driven learning and the underlying physical laws governing memristive switching. This hybrid approach combines recurrent neural networks with device-specific physics, enabling the proposed model to account for both static and dynamic behaviors of RRAM devices, capturing not only the static I-V characteristics but also synaptic-like responses. Extensive experimental validation demonstrates the model’s ability to generalize across various RRAM devices, providing a robust tool for in-memory and neuromorphic computing. This framework reduces the reliance on empirical fitting, automating and streamlining the modelling process, while maintaining high fidelity to physical principles. The integration of PINNs into memristor modeling represents a substantial advancement toward efficient, scalable, and physically consistent simulations of RRAM devices, advancing their use in next-generation AI hardware accelerators.