Experimental parameter study of thermally tunable Mach-Zehnder switches for photonic neuromorphic computing

Photonic neuromorphic computing has attracted great interest in recent years for artificial intelligence applications because it enables highly parallel computing with low latency and low power consumption. Neuromorphic integrated photonics can be realized with networks of Mach-Zehnder interferometers (MZIs), which should be reproducible and stable and have a large tuning range. Here, we experimentally investigate the suitability of MZI-titanium heater configurations for optical neuromorphic computing. Specifically, we rigorously assess their heating efficiency and switching rate for various distances between the MZIs and the heaters and various heater widths. Moreover, we study the influence of different MZI footprints between 6200 μm² and 30000 μm². We complement our experiments with LUMERICAL simulations and establish a highly predictable and reproducible fabrication process compatible with complementary metal oxide semiconductor (CMOS) technology, which makes them ideal for future applications in photonic neuromorphic computing.