Resistive Switching Oxide Nanocomposite Thin Films for Sustainable Artificial Intelligence

Artificial intelligence (AI) applications are increasingly affecting many areas of our lives. However, they are mostly being implemented in the conventional von Neumann computing architecture, which suffers from the memory bottleneck, i.e., a bottleneck in shuttling data between the memory and processing parts of the von Neumann architecture. This leads to immense power consumption due to AI applications, because with billions of parameters, they are very memory-heavy. New approaches to AI computing are thus necessary to enable a sustainable future.<br/><br/>One of the most promising approaches are resistive switching (RS) materials, where information is stored as the resistance state of a memory cell, rather than as charge on a capacitor as in conventional computer memory. RS materials hold promise to realise hardware neuromorphic networks, for example in the form of crossbar arrays, which make it possible to combine memory and processing into in-memory computing and thus circumvent the memory bottleneck. We developed oxide nanocomposite thin films which are based on a self-assembled phase separation on the nanoscale, and which demonstrate highly uniform and stable multi-level resistive switching. They are deposited at industry-friendly temperatures of ≤400 °C and demonstrate great promise in all standard figures of merit for resistive switching. In different materials systems, we achieve ≥10 000 switching cycles across a large number of devices, stable retention of up to ≥300 days, up to ≥500 resistance states, spike-timing-dependent plasticity, and fast switching down to 20 ns. These data will be presented together with compositional and structural analyses of the underlying films and the corresponding RS mechanisms.