On the development of memristive devices for electroforming-free and analog memristive crossbar arrays
Memristive devices can reversibly change their resistance by applying an electrical voltage or current. These thin-film devices have the potential to serve as central components in novel neuromorphic circuits, similar to synapses in the human brain. Unlike traditional neuromorphic systems, they enable a state-based and non-volatile weight between two neurons. This comes very close to the natural model of the human brain, where information is stored and processed together. The aim of this thesis was the development of novel memristive devices and the integration into crossbar arrays. An essential requirement was an analogous resistance change, which allows continuous changes in resistance. It was found, that devices with a combination of tunnel and Schottky barriers are best suited for this purpose. These double barrier devices show an analogous and homogeneous resistance change. As a reference system, filament-based memristive devices have been developed that alter their resistance due the migration of silver. Since the formation of filaments is almost random, they have a significantly higher device variability and very few states between the off- and on-state. Only the high quality of the double barrier component allowed the circuit integration without the need to individually adjust circuit parameters for each memristive device. Due to the non-linear switching characteristics and the advantageous I-V characteristics, the devices were integrated into a space-saving crossbar architecture, which increased the packing density tenfold. Due to the simultaneously simplified electrical connection, it was possible to realize a circuit for pattern classification with 180 memristive devices. The construction of an automated measuring system enabled the characterization of a large number of devices. The development of database-supported measurement and evaluation programs facilitated the analysis of the device and switching properties.