The material is a two-dimensional alternating-cations-in-the-interlayer (ACI) perovskite, GAMA5Pb5I16, grown with additive engineering (MACl) and isopropanol-vapor annealing so that the crystals orient vertically between the electrodes. Because the film is a mixed ionic-electronic conductor, the voltage pulses move iodide ions and vacancies along the inorganic Pb–I framework, while the organic spacer suppresses migration across it (supported by DFT). The conductance therefore changes gradually through an interfacial valence-change mechanism rather than abrupt filament formation.
Key numbers from the paper: 100 discrete conductance states with nonlinearity of 0.027 (potentiation) and 0.064 (depression) and asymmetry of 0.037; with 50 states the nonlinearity is 0.010 and 0.012. The device reproduces paired-pulse facilitation (110% index) with fast and slow relaxation times of about 5 and 30 ms, similar to biological synapses. The read energy reaches 5.5 aJ at 0.05 V and 100 ns, versus 169.5 nJ for a 1.5 V, 100 µs programming pulse. An 8×8 crossbar array showed all 64 units switching.
In simulation, the measured conductance curves were used to constrain the weights of a YOLOv9-based vision network: object detection reached mAP50 of 57.6%, 98.1% of the 58.7% floating-point baseline. The authors stress that this is a software simulation based on a single device's curves, without device-to-device variation, endurance or drift.
Why it matters: the result shows that controlling ion transport by crystal orientation can make a soft, solution-processed material behave like a precise, tunable synapse for in-sensor computing.
Paper: Z. Sun et al., "Highly Linear Perovskite Synapses for Neuromorphic Computing," Advanced Science (2026), DOI: https://doi.org/10.1002/advs.78461
No comments:
Post a Comment