2D Heterostructure‐Based Nanofluidic Ion‐Channel Synapses for Physical Event–Memory Bifurcation and Trajectory‐Memory Encoding

ABSTRACT Biological neural circuits achieve efficient spatiotemporal processing by operating fast event transmission and slow memory formation in parallel. This work presents a 2D nanofluidic bimodal synapse (2D‐NBS) that physically bifurcates raw analog inputs into a fast event component and a slow memory tail within a single ionic‐transport architecture. The device integrates an electrode/electric‐double‐layer (EDL) interface with a cation‐selective 2D nanofluidic ion channel (2D‐NIC) formed in a WS ₂ ‐ or Al ₂ O ₃ ‐on‐multilayer graphene heterostructure. By engineering ionic diffusion resistance and effective EDL capacitance, the relaxation time constant is tuned across a wide range from sub‐millisecond levels to over 7 ms, enabling trajectory‐history‐preserving representations without external spike generation or preprocessing. Exploiting the trajectory‐memory‐encoded representations from 2D‐NBS, convolutional neural network (CNN)‐based recognition improves from 86.8% to 90.3% on the Weizmann action dataset and from 90.4% to 95.0% on the LSA64 dataset for Argentinian Sign Language. Multivariate urban fine particulate matter (PM _2.5 ) forecasting is demonstrated using the 2D‐NBS as a physical reservoir, achieving a mean squared error (MSE) of 3.33 × 10 ⁻⁴ at ∼162 × lower energy per step than a 1941‐parameter vanilla long short‐term memory (LSTM) of comparable accuracy. These results establish the 2D‐NBS as a physics‐driven, sensor‐side feature‐transform primitive that acts as a compact multi‐timescale physical reservoir, separating event and memory components at the sensor front‐end before a lightweight digital readout.