Symmetric and Energy-Efficient Conductance Update in Ferroelectric Tunnel Junction for Neural Network Computing

The rapid development of artificial intelligence requires synaptic devices with controllable conductance updates and low power consumption. Currently, conductance updates based on identical voltage pulse scheme (IVPS) and nonidentical voltage pulse scheme (NIVPS) face drawbacks in terms of recognition accuracy and energy efficiency, respectively. In this study, a mixed voltage pulse scheme (MVPS) for tuning conductance is proposed to simultaneously achieve high recognition accuracy and high energy efficiency, and its superiority is experimentally verified based on high-performance Au (or Ag)/PbZr0.52Ti0.48O3/Nb:SrTiO3 ferroelectric tunnel junction (FTJ) synaptic devices. The MVPS-based neural network simulation achieves a high recognition accuracy of approximate to 92% on the CIFAR10 dataset with better energy efficiency and throughput than those of NIVPS. In addition, high-precision experimental vector-matrix multiplication (with a relative error of approximate to 1.5%) is obtained, and the simulated FTJ synaptic arrays achieved a high inference energy efficiency of approximate to 85 TOPS W-1 and a throughput of approximate to 200 TOPS, which meets the requirements of artificial intelligence in low-power scenarios. This study provides a possible solution for practical applications of FTJ in neural network computing. A mixed voltage pulse scheme is proposed to achieve symmetric and energy-efficient conductance updates. Using this scheme with a high-performance Au/PbZr0.52Ti0.48O3/Nb:SrTiO3 ferroelectric tunnel junction (FTJ), simulated artificial intelligence tasks reveal an impressive training accuracy-energy ratio (approximate to 1.23% J-1), along with a high inference energy efficiency (approximate to 85 TOPS W-1), which lays the groundwork for incorporating FTJs into neural network computing applications. image