Study On Resistive Switching Memory And Neurosynaptic Characteristics Based On Oxide Memristor

The amount of data that needs to be stored and processed in the era of big data is exploding.However,the traditional floating-gate-based memory and the von Neumann-based calculator have reached the technical bottleneck,and their development has gradually slowed down.Therefore,finding new storage devices to replace existing floating gate memories and developing new computing systems have become the focus in the academic and industrial fields.Memristors,especially resistive switching,are believed to be the most powerful contenders for next-generation memory and brain-like computing networks because of their advantages of high speed,low energy consumption,high integration,long retention time and the ability to simulate the function of synapses.An HfO₂-based resistive memristor is studied in this paper for nonvolatile storage,synaptic function simulation and artificial neural network.The main research contents are as follows:1.A Pt/HfO₂/Ti resistive switching was prepared by using ALD HfO₂ film as the key dielectric material,and its resistive switching characteristics under DC scanning were studied.It was found that the device can be Forming by applying a positive large voltage to the top electrode of Ti,and then show a bipolar resistive switching behavior,and the stability of negative Forming is extremely poor.The device's Forming limit current can play a good limiting role,but under small limit current,the devices will still be formed to a much lower resistance state due to the overshoot current by parasitic capacitance.Studies on devices with different thicknesses of HfO₂ show that the device's switching ratio and resistance transition parameters increase with the increase in dielectric thickness,but the device yield and stability of 10nm HfO₂ are the best.The high resistance state of the device has a negative linear relationship with the device area,and the low resistance state is independent on the device area.Combining the result that the high/low resistance state of the device is linearly related to temperature,it is inferred that the resistive switching behavior due to the formation and fracture of local oxygen vacancy conductive filaments.2.Pt/Al:HfO₂/Ti devices with a Al atomic concentration of 6.2%were prepared by inserting AL₂O₃ growth cycles during the deposition of HfO₂.Through the XPS analysis of the dielectric film,it was found that the O-binding energy of the Al-doped HfO₂ dielectric layer was increased,and the formation energy of oxygen vacancies was reduced,so that the oxygen ions in the dielectric were more easily detached from the HfO₂ lattice to form higher concentration of oxygen vacancies,resulting larger current in high resistance state and smaller storage window.While the resistance transition voltage of the device reduced,the uniformity of resistance distribution improved and the characteristic of resistance value gradual change observed,which shown the potential of multilevel storage.By changing the limit current of Set scanning and the amplitude of Set/Reset pulse voltage,a 20-level stable storage state achieved,and a 35-level of multi-value storage is valuable by adjust the number of Set/Reset pulses.3.The influence of the Crossbar structure on the device's switching parameters was studied.Although the resistance conversion voltage of the device reduced,the yield and the uniformity of the resistance distribution of the device also decreased,and the variability between devices increase.It was found that the performance of the device annealed at 500°C was improved,and the self-compliance characteristic achieved.By fitting of the Forming curves of devices with different thicknesses,it was found that as the thickness of the hafnium oxide layer increases,the conduction mechanism of the initial state device changes from P-F emission mechanism to Schottky emission mechanism.We propose that the self-compliance characteristic of the device was obtained because the high-temperature annealing process promotes the reaction between the Ti top electrode and HfO2,producing a thick layer of titanium oxide,which plays a role in current limiting and partial pressure.The oxygen vacancies generated at the HfO₂/Ti interface diffuse downward at high temperatures,making the formation and breaking of oxygen vacant filaments in the device more stable,and the performance of the device improved.The Schottky emission conduction mechanism of thicker hafnium oxide in the initial state is more obvious because the annealing process reduces the Pt/HfO₂ interface state,making the Schottky barrier effect of the interface higher.4.Crossbar arrays with different sizes were prepared and packaged.The reliability of different array test methods for array operation were analyzed.It was found that the array composed of unannealed devices could not work because of the lack of self-compliance characteristic.The floating test method cannot initialize the array due to the different Forming voltages of the devices in the series path,and it has a high probability of misoperation.The grounding method is also impossible to initialize a large-scale array due to the excessive current on the selected word line,which will cause the broken of electrode.But it is suitable for reading the resistance of the device with small voltage.The 1/3Vd test method failed to initialize the array,because the total current of the unselected word line and bit line is too large to measure by the semiconductor analyzer.The array contains 16×16devices was successfully initialized by 1/2Vd test method,but the 32×32 array failed to initialize due to the large leakage current.100 cycles of erase/write tests were performed on the 16×16 array,and it was found that the proportion of failed devices in the array increased from 12%to 31%,much higher than the failure rate of single device.5.The 1T1R arrays were prepared by connecting the memristor and the MOS tube in series,and a double-layer fully connected neural network was built.The accuracy of 8×8handwritten digital image recognition is about 95.19%.By randomly replacing the devices with failed devices,it is found that the stuck off devices have little effect on the accuracy of the network,but the stuck on devices will cause a sharply reduction of accuracy.By using the measured conductivity adjustment range and precision data of the memristor,the relationship between the recognition accuracy of the network and the number of hidden neurons is simulated.Compared with the neural network based on the precision of 32-bit floating-point,the difference is less than 1%.The power consumption of the memristor array for recognizing a picture is only 1.88n J.6.Volatile resistive switching characteristics of the annealed 20nm HfO₂ device by clockwise scanning were studied,and it was found that the device does not need a Forming operation.The operating current is much smaller than that in non-volatile cycling,and the resistance can be restored to the initial state of the device.It is also observed that the resistance of the device decreases with the increase of the device area,and the degradation rate of resistance in low resistance state increases with increasing temperature.The I-V curve fitting of the device Set and Reset process shows that the device belongs to the Schottky conduction mechanism.It is proposed that the volatile resistive switching effect of the Pt/HfO₂/Ti device is caused by the trapping and de-trapping process of trap states near the Pt/HfO₂ interface.Finally,the Pt/HfO₂/Ti volatile resistive switching was used to simulate the short-term enhancement,short-term inhibition,transition from short-term to long-term memory and STDP function of synapses.