Neuromorphic Computing With Silver Doping Chalcogenide Memristor

The computing systems combining information storage and processing are the key technology to meet the urgent demand for high-performance computing at "Age of Big Data" and break through the limitation of von Neumann bottleneck. The biologic brain is a kind of efficient and intelligent information systems combining storage and processing. By the learning, study and mimicry to the brain, the researchers can implement the neuromorphic computing and build the new architecture of computers. Since the birth of the memristor, it has been extensively interested by the industrial and academic circles because of its superior characteristics, such as high speed, low energy consumption, high density and good endurance. Particularly, with memristive theory perfection and device performance optimization, the memristor shows the significant competitive advantage in the neuromorphic computing field. After the memristive synapse and neuron were realized, the study on memristive neuromorphic computing is transferring from the "cell level" devices to the "network level" associative learning and artificial neural network.This thesis focuses on the silver doping chalcogenide memristive materials, devices and application in neuromorphic computing. Some main content and achievements such as fabrication of memristor, materials and devices properties, associative learning, episodic memory and pattern recognition are summarized as follow:The AgGeTe films are prepared by magnetron sputtering. And two kinds of silver doping chalcogenide memristor Ag/AgGeTe/Ta and Ag/AgInSbTe/Ta are fabricated by magnetron sputtering, ultraviolet lithography and lift-off processes. Besides the test methods of memristor and neuromorphic computing circuits are studied.In materials study, the AgGeTe is used to study that Ag doping affects the crystallization and phase transfer temperature of the chalcogenide. The results of DSC and TEM prove the enhancement of stability of amorphous state in Ag doping chalcogenide materials; the results of Raman and XPS provide the physics mechanism of above phenomenon from the change in local order; the result of ab initio simulation analyzes the change in bond angle of Ge atom and number of the n-rings and gives another explanation to above phenomenon from the nucleation model. Besides, with Raman and XRD, we study the phase transfer temperature of AgGeTe. Ag doping can enhance the stability of amorphous state, contribute the formation of conductive filament and benefit the multiple-valued property.In device study, with DC and pulses test, Ag/AgGeTe/Ta and Ag/AgGeTe/Ta memristors reveal the stable hysteresis loop and multiple-valued property. Besides, the high-speed switch of devices also is found by the high-speed pulses test. The multiple-valued and gradual resistance tuning are most important properties of analogue memristor, which is helpful for neuromorphic computing circuits. The high-speed switch can improve the computing speed and decrease the energy consumption, which is helpful for memristive logic circuits.In associative learning, combining the biologic associative learning model and the Ag/AgInSbTe/Ta memristor behaviors, the simple and effective associative learning circuit and array are designed. By the pulses test and Hspice simulation, the circuit can implement the learning and forgetting function in Pavlov' Dogs. Moreover the circuit has the strict temporal relations in learning process, namely only when CS is before US, the circuit will implement the associative leaning. And the learning speed and time window can be controlled by the shapes of the US and CS.In episodic memory, with the multiple-valued property of chalcogenide memristor, a simple episodic memory circuit is build. By the pulses test and Hspice simulation, the episodic-like memory recollection process in cuttlefish is mimicked.In pattern recognition, based on the template matching principle, the "Matcher" which can compute the similarity between the template and target information is designed and verified. This model can recognize the difference in the feature number of the template and target and their similarity degree, and then it is used in the recognition of the 3×3-pixel target images with different shapes and grey-scale. Besides, referring to the associative learning array, a memristive neural network which can storage three nine-features templates is designed, and with Hspice software, the leaning and test processes of'v'?'n'?'z' are simulated. At last, combining the neural network and "Matcher" computing neuron, a memristive template matching network is designed. Because the main body of this network only contains resistors and memristors, it will be benefit for the simplification of artificial neural network.