Neuromorphic applications based on 2D materials ferroelectric field-effect transistors

Abstract

Drawing inspiration from biological neural networks, neuromorphic computing has emerged as a promising solution to overcome limitations imposed by traditional von Neumann architecture. Recently, two-dimensional (2D) materials have provided opportunities for achieving novel nanoelectronics and optoelectronic devices. In the meanwhile, ferroelectric materials possess spontaneous polarization that can be modulated dynamically, reversibly, and non-volatilely. Therefore, we focus on studying ferroelectric transistors based on 2D materials with different structures and investigating their potential neuromorphic applications. Firstly, we design ferroelectric-tuned synaptic transistors by integrating 2D black phosphorus (BP) with flexible ferroelectric copolymer poly(vinylidene fluoride-trifluoroethylene) (P(VDF-TrFE)). Through the nonvolatile ferroelectric polarization, the P(VDF-TrFE)/BP synaptic transistors show mobility value of 900 cm² V⁻¹ s⁻¹ with 10³ on/off current ratio and such devices can be operated with low energy consumption (~40 fJ) for each event. Reliable and programmable synaptic behaviors have been demonstrated, including paired-pulse facilitation (PPF), long-term depression (LTD), and long-term potentiation (LTP). The biological memory consolidation process is emulated through ferroelectric gate-sensitive neuromorphic behaviors. Inspiringly, the artificial neural network is simulated for handwritten digits recognition, achieving high recognition accuracy of 93.6%. Secondly, it is still a challenge to process multiwavelength images in noisy environments with simple device configuration and light-tunable biological plasticity. We demonstrate a prototype visual sensor based on ferroelectric copolymer (P(VDF-TrFE)) and 2D rhenium disulfide (ReS₂) with integration of recognition, memorization, and pre-processing functions in the same device. Such synaptic devices achieve impressive electronic characteristics, including current on/off ratio of 109 and mobility of 45 cm² V⁻¹s⁻¹. Through constructing optoelectronic device array, we achieve target extraction process with wavelength-selective capability in noisy environment, closely resembling human retina for color recognition. The above outcomes bring a notable improvement from 72 % to 96 % in image recognition accuracy. Thirdly, drawing inspiration from object motion sensitive ganglion cells, we propose OMD vision sensor with simple device structure by constructing WSe₂ homojunction modulated by ferroelectric copolymer. Operating under optically and zero power consumption mode, vision sensors can generate progressive and self-powered positive/negative photocurrents with discrete and uniform multi-states facilitated by intermediate ferroelectric modulation. This design enables a reconfigurable device to emulate long-term potentiation and depression for synaptic weights updating, which exhibit multi-levels exceeding 6 bits of 82 states resolution and uniform step of 6 pA. Such OMD devices also demonstrate non-volatility, reversibility, symmetry, and ultra-high linearity with fitted R² of 0.999 and 0.01/-0.01 nonlinearity values. Thus, vision sensors could implement motion detection by sensing only dynamic information while eliminating redundant data from static scenes. Neural network based on linear results can recognize the essential moving features with high recognition accuracy 96.8%. We also present scalable potential via a 3×3 neuromorphic vision sensor array. 2D semiconductor/ferroelectric hybrids could mature rapidly and make inroads into modern electronics by simultaneously progressing in the material, device, and system levels. It is foreseen that such a combination could continue to offer more innovative opportunities.