SOLID-STATE REPROGRAMMABLE ANALOG RESISTIVE DEVICES FOR ELECTRONIC NEURAL NETWORKS

Electronic neural network hardware implementations, particularly with full parallelism and self-learning capabilities, have generated a need for compact “variable-analog memory” elements as “synapses.” We report on electrochemically reprogrammable, solid-state resistor elements as analog memory synapses, which are based on thermally evaporated, electrochromic tungsten oxide (WO3) thin films. The electrical resistance (connection strength) of such tungsten oxide films is reversibly modified by electrochemical cathodic injection into/anodic depletion of cations (H+, Na+, etc.) from a solid-state electrolyte using a third control gate electrode. Two device configurations, one with chromium oxide and the other with sodium tungstate thin film electrolytes as hydrogen and sodium ion sources, respectively, are fabricated and tested. A critical component for the device operation is a silicon oxide blocking layer between the WO3 and the electrolyte layer which provides a nonlinear barrier and thereby increases the stability of the device. The electrical resistance of the device can be modified to, and stabilized at, any desired value over a wide dynamic range of four orders of magnitude. The details of the fabrication, programming characteristics, and operation of the devices are reported in this paper. © 1990, The Electrochemical Society, Inc. All rights reserved.