Thermally enhanced resistive switching in interface-engineered AgHfO3-x/Al2O3 memristors via trap depopulation for neuromorphic computing.

Thermal instability remains a fundamental bottleneck in oxide-based resistive random-access memory (RRAM), where elevated operating temperatures invariably increase the high-resistance-state (HRS) leakage current and collapse the ON/OFF switching window, severely limiting reliable deployment. Here, an Au/AgHfO3-x/Al2O3/FTO (fluorine-doped tin oxide) memristor is reported in which this paradigm is inverted. The HRS current anomalously decreases with increasing temperature, yielding a reversible ∼25-fold enhancement in the ON/OFF ratio from 300 K to 400 K. This counter-intuitive behavior originates from trap-controlled space charge-limited conduction (SCLC) at the chemically sharp 5 nm Al2O3 interlayer, where thermal energy depopulates shallow trap states, suppressing hopping-assisted leakage without disrupting filamentary conduction in the low-resistance state. The optimized heterostructure achieves a room-temperature ON/OFF ratio of 349 ± 53, endurance exceeding 103 bipolar switching cycles, retention over 25 000 s, and ten discrete analog conductance states spanning six orders of magnitude. Device-derived conductance states are directly used as synaptic weights in software-level simulations of handwritten-digit classification and image denoising on the MNIST benchmark. The handwritten-digit classification achieves 97.74% accuracy via transfer learning without implementation in a physical crossbar array or its associated peripheral circuitry. These results establish interface-engineered trap control as a design principle for thermally self-correcting memristors, a class where elevated temperature enhances rather than degrades performance, with direct relevance to neuromorphic hardware in thermally demanding environments.