Machine Learning Prediction of Threshold Switching Voltages Indicates a Filamentary-to-Homogeneous Transition in TiO 2 /TiO 2-x /TiO 2 Memristors

We report a machine-learning (ML) analysis of 763 bipolar switching cycles measured on 62 nanoscale TiO2-based memristors of two architectures: single-layer TiO2(Device A) and TiO2/TiO2-x/TiO2(Device B), the latter incorporating an engineered oxygen-deficient interlayer. Under leave-one-device-out (LODO) cross-validation, XGBoost predictsVSETwithR2=0.875, MAE = 6mV andVRESETwithR2= 0.834, MAE = 14mV on entirely unseen devices. SHapley Additive exPlanations (SHAP) identify the sweep upper limit and compliance current as dominant predictors, while the device-type indicator carries a statistically meaningful secondary weight and captures a ∼ 6× offset in mean threshold voltage (Device B: 311mV vs. Device A: 51mV forVSET). This offset is consistent with an interlayer-induced transition from filamentary to homogeneous switching, providing a data-driven indication of oxygen-vacancy engineering in metal-oxide memristors and is corroborated by independent X-ray photoelectron spectroscopy (XPS) evidence of an oxygen-vacancy-rich TiO2-xstoichiometry on the same fabrication platform.