Model‐Based Time‐Modulated Write Algorithm for 1R Analog Memristive Crossbar Arrays

ABSTRACT Memristive crossbar arrays are a key technology for analog in‐memory computing in AI accelerators and neuromorphic systems. The inherent device nonlinearities are advantageous, suppressing sneak‐path currents in selector‐less (1R) arrays, enabling 3D back‐end‐of‐line integration, and ensuring read stability against the voltage‐time dilemma. However, these same properties, combined with device variability, IR drop, and parasitic effects, severely challenge precise programming. Conventional amplitude‐modulated write‐verify algorithms are consequently slow, energy‐intensive, and limited by write endurance. Here, a time‐modulated write algorithm is introduced, specifically designed for analog‐switching 1R crossbars. It employs a single programming voltage, varying only the write pulse duration. The algorithm leverages a compact model to estimate an initial optimal write time, which is then refined by a dynamic gain mechanism that rapidly compensates for deviations arising from non‐ideal effects. The method's performance is validated through SPICE simulations of a physics‐based / model across various crossbar sizes and V/2, V/3 and floating biasing schemes. The results demonstrate that the time‐modulated approach significantly reduces write iterations and total programming time, while simultaneously lowering energy consumption and improving final conductance accuracy. This proposed algorithm enables faster, more energy‐efficient weight updates and restoration in memristive neural network accelerators.