High‐Performance Synaptic Phototransistor Using A Photoactive Self‐Assembled Layer toward Ultralow Energy Consumption

Abstract

Artificial synapses have gained great interest in the past few years because of their importance in deep learning and image recognition. To fulfill device miniaturization and reduce the energy consumption of device operation, a series of silane‐based photoactive/conjugated self‐assembled molecules (CSAMs), including isoindigo (IID), diketopyrrolopyrrole, and benzodithiophene, are used as the charge‐trapping layers in synaptic phototransistors. The devices comprising CSAM demonstrate excellent short‐term/long‐term memory behaviors and can emulate the paired‐pulse facilitation (PPF) function. Notably, the IID‐based device shows the highest photoresponse, and this performance is highly related to the charge transfer efficiency and the photophysics lifetimes derived from the time‐resolved photoluminescence and the transient absorption characterizations. Therefore, IID produces the highest PPF ratios of 139% to blue light and 144% to green light. In addition, the energy consumption of 0.029 fJ at an operating voltage of −0.1 mV is achieved, which is the lowest value in synaptic phototransistors so far. Notably, neural networks of supervised and unsupervised learning algorithms are demonstrated in the device studied to process a pattern recognition system. Collectively, using conjugated self‐assembled materials as a charge‐trapping layer is a promising way for synaptic phototransistor applications to reduce energy consumption and fulfill the device miniaturization.

Computational Investigation of the Potential and Limitations of Machine Learning with Neural Network Circuits Based on Synaptic Transistors

“…In principle, different neurons can have different NAFs (neuron activation functions) within those realizable by the transistor types used. In recent years, synaptic transistors, both inorganic and organic, realizing nonlinear responses suitable for use as NAFs in NNs, have been actively researched. − Several groups have mimicked the behavior of human synaptic signals through gate voltage stimulation. Lee’s group improved the nonvolatility of artificial synapses by modifying the side chains of conjugated polymers .…”

“…An important issue specific to analog circuit-based NNs is noise and other instabilities. The available literature on synaptic transistor-based NN reported substantial instabilities of the NAF shape from experiment to experiment. ,, Chen’s group utilized optoelectronic synaptic transistor-based NNs to impose noise on the training data during the training process . Lee’s group introduced Gaussian-distributed random noise into the synaptic weights during the training process of spike neural networks (SNNs) .…”

“… 17 , 19 , 20 Chen’s group utilized optoelectronic synaptic transistor-based NNs to impose noise on the training data during the training process. 19 Lee’s group introduced Gaussian-distributed random noise into the synaptic weights during the training process of spike neural networks (SNNs). 20 That is, while there have been studies of the effect of perturbation to NN inputs and weights (corresponding to instabilities or noise in the resistive circuit around the transistor), the effect of the factor of instability of the shape of the NAF itself on NN performance, to the best of our knowledge, remains unquantified even though, as we demonstrate in this article, it appears to be critically important.…”

“…The perturbed shapes of the activation functions shown in the figure are within the degrees of device-to-device and run-to-run variations observed with some experimental synaptic transistors. 17 , 19 , 20 The optimal noise parameter was σ 2 = 10 –4 . The range of train/test set errors is (1301–1375)/(1321–1367) cm –1 (over different train/test splits of the data).…”