Haiyang
Yu†
ab,
Jiangdong
Gong†
ab,
Huanhuan
Wei
ab,
Wei
Huang
ab and
Wentao
Xu
*ab
aInstitute of Photoelectronic Thin Film Devices and Technology, Nankai University, No. 38 Tongyan Road, Haihe Education Park, Tianjin 300350, People's Republic of China. E-mail: wentao@nankai.edu.cn; bnuch@hotmail.com
bTianjin Key Laboratory of Photoelectronic Thin Film Devices and Technology, No. 38 Tongyan Road, Haihe Education Park, Tianjin 300350, People's Republic of China
First published on 25th March 2019
Coexisting-Cooperative-Cognitive (Tri-Co) neuromorphic electronics emulate humankind in order to interact with robotic peers and dynamic environments. To fulfill synaptic functional emulation in a single ultrasensitive and energy-efficient electronic device is of vital importance in this field. This study reports the design and fabrication of two-terminal synaptic devices that are based on mixed halide perovskites (HPs). These devices emulate fundamental synaptic principles by the successive modulation of conductance, for example, excitatory post-synaptic current, paired-pulse facilitation, spike-number dependent plasticity, spike-duration dependent plasticity, spike-voltage dependent plasticity, and spike-rate dependent plasticity. The artificial HP synapses show high sensitivity of 100 mV to external stimuli. These results demonstrate the applicability of the HP synapses in Tri-Co neuromorphic electronics and sensory-motor nervetronics.
Synapses constitute the basic unit of a brain, and synaptic plasticity constitutes the molecular basis of learning and memory. Therefore, the fabrication of a single device that can mimic this synaptic plasticity is an essential step in the construction of an artificial neural network.4,8–10 Traditional silicon complementary metal-oxide-semiconductor (CMOS) technology has been applied to emulate artificial synapses successfully, but each artificial synapse requires ten or more transistors, so they are not practical for realizing a high-density artificial neural network like a biological brain.11,12 Two-terminal and three-terminal device structures have been introduced into artificial synapses.13,14 The change in resistance in two-terminal devices is similar to the change in weights of connections between biological synapses.15 Several synaptic functions, including potentiation/depression of synaptic plasticity, spike-timing dependent plasticity (STDP), and paired-pulse facilitation (PPF) have been mimicked.16–19 Artificial neuromorphic systems that are constructed from synaptic devices have a simple structure and low energy consumption, so, in theory, they could reach a complexity that is approximately equal to a biological brain.6,20,21 These devices provide a feasible path toward building artificial neural networks and Tri-Co electronics.
Halide perovskites (HPs) have remarkable optical and electronic properties, including high light absorption coefficients, tunable bandgaps, long electron–hole diffusion lengths, and long carrier lifetimes.22–25 HPs (i.e., MAPbI3) have been widely used in solar cells, whose power conversion efficiency has exceeded 22%.26–29 However, the development of HP-based solar cells is somewhat hindered; the most relevant here is that its current–voltage curve shows a strong hysteresis, which may result from the ion migration and imbalanced extraction of holes and electrons.30–33 This phenomenon provides a gradual modulation of resistance that can be useful in synaptic electronic devices.
In this study, a two-terminal artificial synapse based on HPs was fabricated by low-temperature solution processes. Artificial synapses based on MAPbClBr2 and MAPbBr3 emulate several synaptic functions, including excitatory post-synaptic current (EPSC), PPF, spike-number dependent plasticity (SNDP), spike-voltage dependent plasticity (SVDP), spike-duration dependent plasticity (SDDP) and spike-rate dependent plasticity (SRDP). These electronic devices are ultrasensitive to external stimuli and energy-efficient.
The XRD patterns (Fig. 2a) of MAPbClBr2 and MAPbBr3 thin films slightly differed. MAPbBr3 crystallites show sharp diffraction peaks at 14.88°, 30.12° and 45.90°, which correspond to the (100), (200) and (300) planes, respectively. These results indicate that the MAPbBr3 crystals are quite pure in phase and preferred-oriented on growth. Doping of MAPbBr3 with Cl led to lattice distortion and crystal defects, so the intensity of diffraction peaks decreased for the MAPbClBr2 crystals and another peak that corresponds to the (210) plane appeared. To further confirm the existence of Cl, XPS was performed to study the distributions of surface component. The full scan XPS spectra (Fig. 2b) show carbon, nitrogen, lead and bromine in both specimens, and an additional Cl 2p peak appeared at 198 eV in MAPbClBr2; this peak indicates that the Cl had been incorporated into the crystal lattice.
The surface roughness and morphology of the MAPbClBr2 and MAPbBr3 thin films were recorded using AFM and SEM (Fig. 2c–h). Both thin films had a smooth surface with a root-mean-square (RMS) roughness of 15.9 nm for MAPbClBr2 and 7.39 nm that for MAPbBr3, and no obvious separate grains formed during the film growth. SEM cross-sectional images determined that the MAPbClBr2 films were 95.5 nm thick and the MAPbBr3 films were 109.1 nm thick.
Synaptic plasticity is the ability of synaptic connections to strengthen or weaken in response to external stimuli over time, which is the biological basis of signal transmission between neurons. Synaptic plasticity can be mainly divided into short-term plasticity (STP) and long-term plasticity (LTP).34 STP is recognized as the key element to extract valuable information efficiently while the brain is processing a huge number of stimuli.35 EPSC is caused by a temporary depolarization of the post-synaptic membrane potential, and it is one of the most common forms of STP.
To emulate this excitatory response in our artificial synapse, a synaptic device based on MAPbClBr2 was fabricated to mimic the biological synapse. The top and bottom electrodes served as the pre-synaptic and post-synaptic membranes, respectively, and the active layer of MAPbClBr2 acted as the synaptic cleft. A small constant voltage bias (0.1 V) was applied to the bottom electrode as a reading voltage, and a transient pulse with an amplitude of −0.5 V was applied to the top electrode to set up a temporary electric field across the HP thin film. The external pulse caused an increase in EPSC, which then decayed to its initial state (Fig. 3a). This phenomenon may be ascribed to the migration of halides and ion vacancies in response to the applied electrical field. When the input pulse arrived, the halide ions in HPs would migrate in an orderly fashion under the electric field force; as a result, the HP's conductivity increased substantially. Therefore, the artificial synapses showed memristive behavior, triggering a transient non-linear current that is analogous to EPSC in biological synapses. When the external stimulus was removed, the halide ions diffused back to their equilibrium positions under the effect of a concentration gradient, so the HP's electrical conductivity decreased to its initial state. The halide ion migration and the vacancy-assisted diffusion have dominant effects on the control over the multiple resistance levels of HP-based synaptic devices.
During the biological synaptic transmission, the amount of Ca2+ that is delivered to the pre-synaptic membrane increases if the pre-synaptic neuron is stimulated twice in quick succession. This result is an increase in the quantity of neurotransmitters that are released. The result is PPF, by which the absolute value of the second EPSC is larger than the first. PPF is a significant form of STP that exploits the dynamic increase in the release of neurotransmitters; this increase sensitizes the synapses to external stimuli and has applications in several processes, including signal processing and sound-source localization.36
PPF was demonstrated in our artificial synapse based on MAPbClBr2 (Fig. 3b). EPSCs, triggered by a pair of consecutive input pulses, were observed. The second EPSC had 12.5% higher amplitude than the first EPSC. The migration of the halide ions in HPs caused an increase in the conductance of the synaptic device, and the ions tended to quickly return to their positions after the first external stimulus. However, the second input impulse arrived before all of the ions could diffuse back to their equilibrium sites. Therefore, both migration distance and quantities of the halide ions were increased, so the second EPSC spike was larger than the first.
Hebbian theory suggests that the repeated stimuli between two neurons can increase their synaptic transmission efficacy and enduringly improve the neural connections.37 For instance, SNDP is a learning rule, which indicates that a series of external stimuli can alter the quantity of neurotransmitters released into a synapse. To mimic SNDP in our synaptic device based on MAPbClBr2, a train of impulses with a fixed amplitude of −0.5 V and interval of 68 ms was applied to the metal top electrode as action potentials. As the number of pulses increased from one to eleven, the amplitude of EPSC spikes increased, but at a decreasing rate (Fig. 3c). The trend is a result of an increase in the number of migrated halide ions and vacancy defects in the electronic device; these increases caused an increase in the conductance of the electronic device. Initially, only a few pulses were used to trigger EPSC, so a later pulse could drive many halide ions and yield a huge increase in conductance. Nevertheless, as the number of pulses further increased, the diffusion of halide ions decreased because they were being driven into an increasing concentration gradient.
In a biological synapse, the increase in the duration or strength of a stimulus causes an increase in the quantity of chemical neurotransmitters that are released from the pre-synaptic vesicles into the synaptic cleft. Analogously, changes in the pulse width and amplitude of input signals yielded emulation of both SDDP and SVDP in the artificial synapse based on MAPbClBr2 (Fig. 3d). When the pulse duration was increased under a constant pulse amplitude, the EPSC peaks increased at first but at a decreasing rate. This phenomenon occurred because the total number of mobile halide ions was finite. During the initial stage, only a small portion of the halide ions was driven to move under an external electric field when the pulse duration was short. With the increase in the pulse duration, the number of migrated halide ions increased and the residual number of mobile halide ions decreased, so the effectiveness of an increase in the pulse duration declined. Similarly, when the pulse voltage was increased with constant pulse duration the EPSC peaks also increased.
Energy consumption was calculated as E = Ipeak × V × td, where Ipeak is the peak of EPSC, V is the pulse amplitude, and td is the pulse duration. The minimal energy consumption was estimated to be 5.8 pJ per synaptic event for the −0.1 V pulse. This energy consumption could further be reduced by decreasing the pulse duration.
An artificial synapse based on pure MAPbBr3 was fabricated using identical synthesis processes and device structures as for the artificial synapse based on MAPbClBr2. A series of consecutive pulses evoked a higher EPSC in the artificial synapses based on MAPbClBr2 (Fig. 4a) than that in artificial synapses based on MAPbBr3 (Fig. 4b). The presented EPSCs were the last EPSC peaks, which were also the largest per train of stimuli. The number and voltage of pulses both influenced the overall current amplitude. As the pulse number and voltage increased, EPSC also increased; these changes emulated SNDP and SDDP, respectively. The proportional increase decreased with an increase in the pulse number, but increased as the pulse voltage increased. The EPSCs were smaller in the artificial synapses based on MAPbBr3 than in the artificial synapses based on MAPbClBr2 under the same test conditions. The difference might have occurred because chloride ions (Cl−) have a smaller diameter than bromide ions (Br−).38 Small ions can migrate and diffuse more easily and freely than large ions can. Another reason may be that Cl− in perovskite halides have a relatively lower activation energy Ea than Br−,39–41 so Cl− can move more readily than Br− along the grain boundaries in HP thin films, resulting in a larger increase in the conductance of the HP film. However, both artificial synapses responded to a 0.1 V stimulus, which was comparable with that of a biological synapse.
Both devices also showed SDDP (Fig. 4c and d). Prolonged impulse duration initially caused a significant increase in EPSC, but the proportional increase declined as the duration increased. This kind of nonlinear increase indicates that the multi-resistance states could be realized by finely adjusting the pulse number, voltage and duration.
As an extension of Hebbian learning rules, SRDP describes how the firing rate of pre-synaptic spikes affects synaptic plasticity, which is closely related to associative memory, learning and forgetting behaviors of the brain.42 Both of these HP-based synaptic devices mimicked SRDP after the application of ten consecutive impulses with different pulse intervals (Fig. 4e). The current recorded here was the EPSC triggered by the tenth input pulse. EPSCs were different in the devices, but in both devices they decreased as the pulse interval increased. When the pulse interval was small (i.e., rate was high), the back diffusion of the halide ions was inhibited by the second impulse, so the conductance increased. The suppression of the back diffusion would weaken as the pulse interval increased, so the enhancement effect became inefficient.
To investigate whether LTP could be emulated in the artificial synapses based on MAPbClBr2, 30 consecutive pulses were applied to the device. EPSC increased continuously as the pulses were being applied, but decayed to the initial state within 100 ms (Fig. 4f). Although this HP-based artificial synapse exhibits poor memory retention, this ultrasensitive characteristic provides potential applications in sensory nervetronics to mimic a peripheral nerve system.43,44
Footnote |
† These authors contributed equally. |
This journal is © the Partner Organisations 2019 |