Quantifying Variability, Symmetry, and Nonlinearity of Protonic Electrochemical Random-Access Memory for Hardware...

Programmable synaptic devices that operate reliably in the low-conductance regime with low variability, good linearity and symmetry are essential for low-error, high-accuracy training in crossbar-based neural network hardware. Electrochemical random-access memories (ECRAMs) operate through ion insertion and extraction and uniform lattice doping, offering deterministic conductance modulation with low variability. Despite recent progress, device variability, symmetry, and linearity have not been systematically assessed in the targeted low-conductance regime with an appropriate modulation range. To address this gap, we evaluate device metrics in protonic ECRAMs fabricated with back-end-of-line (BEOL)-compatible materials and processes. The devices consist of amorphous WO3 channels, sputtered HfO2 electrolytes, and PdHx proton reservoirs. They simultaneously satisfy multiple requirements for neural network hardware. They demonstrate low cycle-to-cycle (CtoC) variability (3%) and low device-to-device (DtoD) variability (21%). They also exhibit a large modulation dynamic range (>9) in the low-conductance regime (10-90 nS), together with good and consistent symmetry and linearity. The devices achieve endurance exceeding 109 conductance updates. These characteristics benefit from high channel conductance sensitivity, efficient proton conduction through the electrolyte, and sufficient hydrogen supply from the reservoir. Uniform lattice doping and conformal deposition contribute to low variability and good endurance. We incorporate quantified device characteristics and nonidealities into hardware training simulations. The resulting simulations yield low training error and high test accuracy on the handwritten digit classification task. The simulations also indicate that asymmetry and nonlinearity remain key factors limiting accuracy and require improvement. These results establish protonic ECRAMs as strong candidates for programmable resistors for efficient, low-error hardware AI training.