Artificial Synapse with Mnemonic Functionality using GSST-based Photonic Integrated Memory

Abstract

Machine-learning tasks performed by neural networks demonstrated useful capabilities for producing reliable, and repeatable intelligent decisions. Integrated photonics, leveraging both component miniaturization and the wave-nature of the signals, can potentially outperform electronics architectures when performing inference tasks. However, the missing photon-photon force challenges non-volatile photonic device-functionality required for efficient neural networks. Here we present a novel concept and optimization of multi-level discrete-state non-volatile photonic memory based on an ultra-compact (<4um) hybrid phase change material GSST-silicon Mach Zehnder modulator, with low insertion losses (3dB), to serve as node in a photonic neural network. An optimized electro-thermal switching mechanism, induced by Joule heating through tungsten contacts, is engineered. This operation allows to change the phase of the GSST film thus providing weight updating functionality to the network. We show that a 5 V pulse-train (<1 us, 20 pulses) applied to a serpentine contact produces crystallization and a single pulse of longer duration (2 us) amorphization, used to set the analog synaptic weights of a neuron. Emulating an opportunely trained 100x100 fully connected multilayered perceptron neural network with this weighting functionality embedded as photonic memory, shows up to 93% inference accuracy and robustness towards noise when performing predictions of unseen data