Low-power-consumption and excellent-retention-characteristics carbon nanotube optoelectronic synaptic transistors for flexible artificial visual systems

Low-power-consumption and excellent-retention-characteristics flexible optoelectronic synaptic devices have become the key units in the advancement of neuromorphic computing systems. In this work, we firstly utilized three photosensitive pyridine-based polyfluorene derivatives to selectively isolate semiconducting single-walled carbon nanotubes (sc-SWCNTs) from commercial SWCNTs and successfully constructed low-power-consumption (98.71 aJ) and excellent-memory-characteristics (Up to 1100s) optoelectronic synaptic SWCNT TFT devices for flexible artificial visual systems (The recognition accuracy up to 97.06 %) without adding any other photosensitive materials in SWCNT TFTs. As-prepared optoelectronic synaptic TFT devices showcase excellent electrical properties with exceptional uniformity, enhancement-mode and high on-off ratios (Up to 106), low operating voltages (-2 V to 0 V), and small subthreshold swings (SS, 75 mV/dec). More importantly, they can simulate not only excitatory postsynaptic currents (EPSCs) and paired-pulse facilitation (PPF, up to 272 %) with the power consumption as low as 98.71 aJ per optical spike under light-pulse stimulation but also the traditional Pavlovian conditioned reflex and artificial visual memory system with excellent memory behaviors (Up to 1100s). Through an in-depth analysis of their working mechanism, we successfully emulated long-term potentiation (LTP) and long-term depression (LTD) phenomena, achieving a 97.06 % accuracy rate in the MNIST (Modified National Institute of Standards and Technology database) recognition task. Furthermore, employing these TFTs, we successfully constructed a five-layer convolutional neural network that operates without any external storage and computational units, validating its image recognition capabilities on the Fashion-MNIST dataset with an accuracy rate of 90.58 %, closely approaching the ideal scenario of 91.25 %. These findings provide a robust technological foundation for the development of highly efficient and flexible artificial visual systems in the future.