Investigation On The Artificial Synapses Based On Ferroelectric Memristors

The development of Von Neumann computing system has been facing critical challenges due to so-called memory wall problem.As one of possible solutions,emerging electronics beyond CMOS have drawn researchers’attention tightly and became a hotspot in information technology.In particular,it has been demonstrated that memristors,whose prototype was firstly proposed in 2008,could potentially work as not only non-volatile memory devices but artificial synapses.As a result,memristors can hopefully become the foundation of in-memory computing system with high efficiency and low energy consumption.Recently,a number of energy-saving memristors have been investigated and reported.Meanwhile,artificial synapses based on memristive devices have mimicked a series of neuromorphic functions.Moreover,challenging tasks such as image recognition and logic computing were accomplished by artificial neuron networks,where memristors played significant roles.So far there are so many works focusing on implementing these functions and decreasing energy cost of memristors.However,researches on another essential issue,the energy consumption of the circuit seems insufficient.In this work,we tried to solve the problems mentioned above.Firstly,high-quality BaTi O₃ thin film was deposited by pulse laser deposition,based on which we fabricated ferroelectric memristors with the structure of Pt/BaTi O₃/Nb:Sr Ti O₃.The devices could serve as advanced memory units because of their prominent advantages,including but not limited to high ON/OFF ratio,excellent reliability,high-speed switch and low energy consumption.Furthermore,we demonstrated that the memristors can be regared as energy-efficient artificial synapses by mimicking several sorts of synaptic plasticity,especially low-energy-cost spike-timing-dependent plasticity（STDP）.Afterwards we designed an associative learning circuit based on the memristors and discussed the factors that could influence the energy consumption and the efficiency of the circuit.In the end,we realized associative learning with low energy consumption and high energy efficiency.The key results are summarized as follows:1.Employing pulse laser deposition,we gained high-quality BaTi O₃ thin film in oxygen atmosphere（15 Pa）at 650℃.According the results of X-ray diffraction and reciprocal space scanning,epitaxial growth of crystal thin film on Nb:Sr Ti O₃ substrate was achieved.With the help of atomic force microscope,flat surface of the film was confirmed and its root mean square roughness is as low as 325 pm.Having analyzed local hysteresis loops and polarization reversal via piezoelectric force microscope,we confirmed that the BaTi O₃ thin film had good ferroelectric properties.2.The Pt/BaTi O₃/Nb:Sr Ti O₃ ferroelectric memristors could work as brilliant memory units.In addition to the high ON/OFF ratio（10⁴）,the devices had a strength of superior reliability,which were strongly supported by retention, endurance and uniformity tests.Besides,we explored switch behaviors derived by writing pulses varying in amplitude and width.According to the results,fast switching can be realized.Even if the writing pulses were as short as 60 ns,the ON/OFF ratio could reach 10².Also,low energy consumption in high-speed switching was demonstrated.For example,the ON/OFF ratio in switching could exceed 10 by consuming the energy of 6.97 p J and 0.27 in SET（OFF state to ON state）and RESET（ON state to OFF state）operations respectively.3.Regarded as artificial synapses,the devices were employed to mimic synaptic plasticity.The weights of the artificial synapses could be modulated linearly.After stimulation,the weights’evolvements were corresponding to Ebbinghaus curve.By increasing the amplitude and number of pulses respectively,short-term plasticity was successfully transferred to long-term plasticity.Meanwhile,low-energy-consumption STDP was achieved by relative weak stimulus（2 V,60 ns）.It is noteworthy that the average cost（3.83 p J）was as good as advanced researches on memristors.4.Lots of energy was wasted by resistors when circuits were utilized to neuromorphic functions,which led to a problem of low energy efficiency.To deal with it,a circuit consisting of one memristor and one resistor was designed to accomplish associative learning.After analyzing and optimizing related factors,we reduced the energy consumption to 7.86 p J which was far lower than reported works.A high efficiency of 96.2%was also realized as well.