Multilevel weight optimization of photonic synapses based on slot-ridge waveguides with multi-block GST

Neuromorphic computing has been a candidate to work like human brains to deal with big data energy-efficiently beyond the von Neumann architecture. As an important part, artificial synapses based on waveguides with phase-change materials (PCMs) play a critical role in neural networks. However, because of demanding conditions of triggering intermediate phase states to achieve multilevel weights it is relatively difficult to obtain accurate multilevel weights. In this work, we designed a photonic synapse based on slot-ridge waveguides with multi-block Ge2Sb2Te5 (GST) to precisely and easily get multilevel weights by triggering the specific number of GST islands between crystalline and amorphous states. The properties of the GST material were analyzed by Raman spectroscopy and spectroscopic ellipsometry and the nonlinearity factor of photonic synaptic weights was optimized for the number of GST islands. The recognition tasks of MNIST and the optical recognition of NIST were performed by multilayer perceptron, and the optimal accuracy of the optical recognition of NIST was 92.7 % for 5 GST islands. The accuracy of MNIST was improved by 2 % for 5 GST islands compared with that of 2 GST islands. Besides, the electro-thermal phase transition conditions of crystallization and reamorphization were also obtained by simulation. This work will open the way to precisely and easily achieve photonic synapses with multilevel weights in the future.