MoTe2-based synaptic-bridge memristor for brain-inspired computing: neuromorphic performance evaluation using MLP-CNN frameworks

Neuromorphic computing provides a transformative route to overcome the power and memory bottlenecks of conventional von Neumann systems. Memristors, with electrically programmable resistance states emulating biological synapses, serve as a key enabler where precise resistance modulation facilitates diverse spatiotemporal functionalities. A scalable and solution-processed memristive layer is pivotal for large-scale fabrication with tunable material properties and integration compatibility. However, inhomogeneous filler networks and semiconductor-polymer interfaces often result in non-uniform filament nucleation, leading to unstable switching, poor retention, and degraded memory performance. Herein, a two-terminal memristive device incorporating highly crystalline 2H-MoTe2 within a PVA matrix is fabricated to explore the resistive switching driven by the formation of a percolative network. The device with a MoTe2:PVA ratio (3:1) exhibits stable bipolar resistive switching with minimal voltage variation over 125 cycles. The device further emulates essential synaptic functions (STDP, SNDP, LTP, LTD) and higher-order behaviors, such as Pavlovian learning and Morse code recognition. Their neuromorphic potential is further demonstrated and evaluated through simulations employing multilayer perceptron (MLP) and convolutional neural network (CNN) architectures for off-chip digit classification on the CIFAR-10 dataset. Collectively, these results establish MoTe2 memristors as promising enablers of non-von Neumann, in-memory computing, paving the way for next-generation neuromorphic and AI-driven technologies.