A device-agnostic framework for modeling resistive random-access memory devices using neural ordinary differential equations

With the rise of foundation models in artificial intelligence (AI), developing energy-efficient hardware architectures to address the resulting immense computational demands becomes increasingly critical. Resistive random-access memory (ReRAM)-based crossbar arrays offer promising in-memory computing capabilities, but practical deployment necessitates accurate compact models for circuit-level simulation. This work presents the neural ordinary differential equation-based random-access memory modeling framework (NeurODE-RAM), a data-driven approach to compact modeling of ReRAM devices using deep neural networks (DNNs), for advanced devices with limited physical understanding. Motivated by equation-based compact models, NeurODE-RAM employs neural ordinary differential equations to approximate the continuous internal memory-state dynamics within an ordinary differential equation-recurrent neural network (ODE-RNN) framework, using experimentally calibrated simulation data. This is important for circuit-level integration, as Simulation Program with Integrated Circuit Emphasis (SPICE) transient analysis uses adaptive time-steps, yielding irregularly sampled time series that require robust handling. NeurODE-RAM accurately predicts long-term potentiation (LTP) and long-term depression (LTD) characteristics under various voltage pulse conditions, with high test-time accuracy and strong extrapolation—beyond three times the observed period, at under 5% relative linear error. Importantly, the framework's device-agnostic nature is evaluated across three ReRAM devices with different material stacks (TiN/TaOx/Pt, Ta/HfO 2 /Pt, and Pd/Ta 2 O 5 /TaOx/Pd) and geometric dimensions, demonstrating consistent accuracy and extrapolation without architectural modifications. Finally, NeurODE-RAM is evaluated at the circuit level through SPICE simulations of a 7 × 7 crossbar array, accurately reproducing multi-level conductance patterns under non-uniform local voltages, demonstrating its applicability as a practical component in electronic design automation (EDA) workflows for large-scale neuromorphic and in-memory computing.