Neuromorphic Computing And Stochastic Computing Based On Physically Transient Memristors

In the human brain,there are a large number of neurons and synapses,which constitute a system with low power consumption and high energy efficiency computing power.It can simultaneously store and calculate information,as well as learn and process multidimensional information,such as language understanding,image recognition,abstract reasoning,and other complex functions.At present,various types of electronic devices such as thin film transistors(TFTs),phase change memory,ferroelectrics,and resistive random access memory(RRAM)are used to develop neuromorphic electronic.Among them,the artificial synapse array based on RRAM(memristor)has the analog computing,parallel computing,low power consumption,compute-in memory characteristics,which is one of the most promising candidates for the next generation of the neuromorphic electronic to break through the traditional von Neumann bottleneck in the future.On the other hand,decay factors to the brain and other parts of the central nervous system can kill or disable neurons.When a large number of neuronal cells die during migration and differentiation,neurons lose their connection and biological synapses are damaged,leading to the disappearance of memory ability in the brain.Thus,emulating the death phenomenon of biological synapses based on electronic devices has high potential in security neuromorphic computing and neuro-medicine application.In addition,the memristors show random dynamic characteristics,which is conducive to the randomness of the switching voltage of the device and the varitions of device to device.Thereby,synaptic devices based on biodegradable and physical transient memristors have been researched for secure neuromorphic computing.Additionally,artificial neuron based on memristor has been demonstrated chaotic features for random walk algorithm computing.Then,physically transient true random number generator(TRNG)has been designed based on threshold switching devices in the application of monte carlo computing.The main research contents are as follows:1.The concept of physical transient synapse device based on physical transient memristor is proposed.The W/MgO/ZnO/Mo structure memristor shows physical transient characteristics,analog switching characteristics,and synaptic functions including long-term plasticity,as well as spike time-dependent plasticity(STDP).Additionally,the device can be completely degraded in deionized water within 7 minutes.2.W/MgO/Mg/MgO/W physical transient synapse device was fabricated by nano-film doping technology,exhibiting low power consumption characteristics with single synaptic spike power consumption of 560 p J.The devices show multilevel conductance states and high-precision tunable function of conductance weights under the activation of voltage pulses with various pulse frequencies and amplitudes.3.The W/MgO/ZnO/Mo physical transient memristor based optical synapse has multilevel conductance states with optical sensory ability.Besides,various conductance weight distributions can be obtained under the red,green and blue lights.Thereby,an optical convolutional neural network can be built by the optic synaptic devices for the 32×32 colored image recognition with high recognition accuracy of 90%.4.The Cu-Ta/IGZO/TiN Schottky diode was rearched based on the Cu-Ta alloy electrode.What is important that the Schottky diode shows threshold switching behaviors with chaotic characteristics under positive sweeping voltage after the forming process.Therefore,a stochastic neuron circuit with LIF model can be implemented by the threshold switching device,applying for solving the global optimal solution with the random walk algorithm.5.The W/Ag/MgO/Ag/W device shows threshold switching feature with intrinsic randomness,which can be applied for a physical transient true random number generator.Importantly,RESET programming process is not required of the TRNG device compared to the traditional true random number generator.Besides,the true random number generator can be used to calcaulate the numerical value of pi with Monte Carlo mehotd,and the calculation result is consistent with the real solution of pi.