Analysis of Lithium Niobate Memristor Devices for Neuromorphic Programability

This paper describes the fabrication and characterization process used to develop a series of lithium niobate memristors. A common approach for the development of memristor-based neuromorphic circuits is to store synaptic weight values within memristors as resistance states. Therefore, memristors for these systems must be stable, symmetric, and programmable with a significant bit resolution. In other words, a continuous resistance range must be available in these memristors to store the weight matrix produced by a learning algorithm. Furthermore, it is important to be able to iteratively program a target resistance through a number of feedback controlled voltage pulses as opposed to abruptly switching the device between two binary states. This paper describes the fabrication and characterization results for a set of six different memristor wafers. These results are used to decide which device composition is the best for neuromorphic computing applications through the properties of symmetry, reliability, stability, and programmability. Once the best device for neuromorphic programmability was selected, we also show the potential programming resolution available in this device using a voltage pulse characterization.