† Corresponding author. E-mail:
Project supported by the National Science and Technology Major Project of China (Grant No. 2017ZX02301007-002), the National Key R&D Plan of China (Grant No. 2017YFB0701701), and the National Natural Science Foundation of China (Grant Nos. 61774068 and 51772113). The authors acknowledge the support from Hubei Key Laboratory of Advanced Memories & Hubei Engineering Research Center on Microelectronics.
Phase-change material (PCM) is generating widespread interest as a new candidate for artificial synapses in bio-inspired computer systems. However, the amorphization process of PCM devices tends to be abrupt, unlike continuous synaptic depression. The relatively large power consumption and poor analog behavior of PCM devices greatly limit their applications. Here, we fabricate a GeTe/Sb2Te3 superlattice-like PCM device which allows a progressive RESET process. Our devices feature low-power consumption operation and potential high-density integration, which can effectively simulate biological synaptic characteristics. The programming energy can be further reduced by properly selecting the resistance range and operating method. The fabricated devices are implemented in both artificial neural networks (ANN) and convolutional neural network (CNN) simulations, demonstrating high accuracy in brain-like pattern recognition.
Artificial intelligence (AI) that allows machines to think and act like human brains has been the main focus of computer science in this century, and the application of AI has rapidly extended to various fields in the past few years, such as big data mining, large-scale visual/auditory recognition and classification, driverless cars, complex strategic games, etc.[1–4] Currently, these tasks and applications are based on computers with von Neumann architecture, using conventional central processing units and graphics processing units with off-chip memories, to implement large-scale neural network training which requires several kilowatts of power.[5] Although custom-designed neuromorphic hardware with complementary metal oxide semiconductor (CMOS) technologies could greatly reduce the energy, the power consumption is still a serious issue with the fast expansion of the network scale[6] due to high-frequency data exchange and transmission. It is found that the dynamic random off-chip weight storage access memory will consume 100 times more power than the on-chip memory.[7]
In the biological brain, memory and processing are highly combined to provide an efficient and low-power way of computation.[8] Hence, we can use new memory devices to mimic the way how our brain works. Recently, novel non-volatile devices which can store information in different resistance states and exhibit conductivity modulation based on the programming history, are promising for achieving synaptic dynamics in a compact and power-efficient manner.[9–17] Among these devices, the phase-change materials (PCMs) provide a simple way to be integrated on a large scale, showing a good application prospect. It has been demonstrated by recent impressive achievements using large numbers of synapses.[5,18–28] However, the traditional mushroom-shape PCM devices present relatively high-power consumption under the RESET operation[29] due to the melt-quench amorphization process.[30] During that process, the conductance of PCM devices changes abruptly which is undesirable for mimicking the synaptic events. The mainstream solution is to use two PCM devices in the form of differential pairs to achieve synaptic functionality (one device is used for potentiation and the other for depression).[5,31–33] But it requires long refresh operations, where all PCM devices are being reprogrammed. This not only increases the energy consumption of the system but also increases the complex of processing, thus limiting the largescale integration for phase change synapses. In recent years, superlattice-like (SLL) PCM devices have drawn many attentions in this field because they reduce the energy consumption and meanwhile still maintain good performance.[34] Although the working mechanism is still under intensive debates,[35–38] the performance of low-power consumption and the outstanding phase transition behavior in SLL-PCM provides new opportunities to improve the performance of PCM devices for memory and computing applications.[39] However, there is no systematic study on how to use and operate SLL-PCM in neural networks or the accuracy of this type of device.
In this article, we fabricated a GeTe/Sb2Te3 SSL device to mimic the synaptic events. Through interface control, the analog behavior of this artificial synapse has been optimized. In the specific operation process, we have adopted two different schemes and improved them. The benchmark tasks for handwritten digits recognition are also simulated and verified with artificial neural networks (ANN) and convolutional neural networks (CNN), respectively.
We fabricated the SSL GeTe/Sb2Te3 PCM cell with 250 nm diameter via-hole structure. First, 10 nm Ti, 100 nm Pt, and 100 nm SiO2 isolation layer were deposited in sequence. Next, electron beam lithography (EBL) and inductive coupled plasma etching (ICPE) were performed to create a 250 nm hole and expose the bottom electrode (BE) contact pad. Photolithography was then employed to pattern the functional layer. After that, we deposited the functional layers through magnetron sputtering at room temperature. 2 nm Sb2Te3 was firstly deposited on the BE, and then 4 nm GeTe was deposited on Sb2Te3. After repeating this cycle 12 times, 2 nm Sb2Te3 was lastly deposited on the top. Using this method, we obtained GeTe/Sb2Te3 SSL functional layers with good heterogeneous structure. Finally, 100 nm Pt inert top electrode (TE) was deposited. The electrical properties measurements were carried out by Agilent B1500 A semiconductor analyzer. The top Pt electrode was biased and the bottom Pt electrode was grounded during the measurements.
The two phases have remarkable resistance contrast, as displayed by the I–V characteristics in Fig.
By continuously adjusting the amplitude of the voltage pulses, more intermediate states can be obtained (Fig.
Figure
The analog weight update is used in the simulation based on superlattice-like PCM conductance modulation properties, such as the number of conductance states, Gmax/Gmin. Our network uses supervised learning method with traditional backpropagation algorithms.[40]
The programming process follows either one of the two learning rules below: an update scheme using unfixed voltage pulses and an update scheme with fixed voltage pulses, where η is the learning rate of the network,
Using the three groups of synaptic weight data from Fig.
Due to its parameter sharing mechanism and sparseness of connections, CNN has better recognition results than ordinary ANN.[41–44] The schematic diagram of our CNN network structure is shown in Fig.
Figure
Low-power consumption of our SLL device is mainly due to interface scattering and phonon microstrip transport, which results in extremely low thermal conductivity, thus significantly improving heating efficiency.[45] We roughly estimate the power consumption of the device under pulse operation through equations below:
We have demonstrated that superlattice-like PCM devices can be used as artificial synapses for neuromorphic systems. The realization of gradual conductance depression improves the symmetry of synaptic weight update, which has greatly increased the accuracy of this neural network. Different strategies of voltage amplitude schemes have been selected according to different scenarios. By properly selecting the dynamic range of the device conductance, not only the power consumption, but also the performance of the neural synapses can be optimized. The simulations of this neural network based on the above device show great accuracy of 91.7% in recognizing the handwritten digits. Our results may facilitate the design of the neuromorphic hardware systems based on superlattice-like PCM devices.
[1] | |
[2] | |
[3] | |
[4] | |
[5] | |
[6] | |
[7] | |
[8] | |
[9] | |
[10] | |
[11] | |
[12] | |
[13] | |
[14] | |
[15] | |
[16] | |
[17] | |
[18] | |
[19] | |
[20] | |
[21] | |
[22] | |
[23] | |
[24] | |
[25] | |
[26] | |
[27] | |
[28] | |
[29] | |
[30] | |
[31] | |
[32] | |
[33] | |
[34] | |
[35] | |
[36] | |
[37] | |
[38] | |
[39] | |
[40] | |
[41] | |
[42] | |
[43] | |
[44] | |
[45] | |
[46] | |
[47] | |
[48] |