On-Chip All-Optical Nonlinear Activation Function for Photonic Neural Network via Two-Dimensional Ti3C2 (MXene) in Near-Infrared

  Adir Hazan [1]  ,  Barak Ratzker [2]  ,  Zhang Danzhen [3]  ,  Aviad Katiyi [1]  ,  Nahum Frage [2]  ,  Maxim Sokol [4]  ,  Yury Gogotsi [3]  ,  Alina Karabchevsky [1]  
[1] School of Electrical and Computer Engineering, Ben Gurion University of the Negev, Israel
[2] Materials Engineering Department, Ben Gurion University of the Negev, Israel
[3] Materials Science and Engineering Department, Drexel University Philadelphia, USA
[4] Materials Science and Engineering Department, Tel-Aviv University

The utilization of integrated photonics in Neural Networks (NN) offers a promising alternative approach to microelectronic and hybrid optical-electronic implementations due to improvement in computational speed and power efficiency for machine-learning tasks. However, until now the non-linear activation function is still fulfilled electronically resulting in slower computational speed and noisy signals.

Here we report on Ti3C2 MXene all-optical devices for implementing the neuron's nonlinear activation function optically. We studied two configurations: 1) saturable absorber made of MXene thin film, and 2) silicon waveguide with MXene flakes overlayer. 

We utilize unique light-matter interactions in 2D material - MXene, in the nonlinear regime, and experimentally show the nonlinear activation function on a chip. The results presented here are generated through a combined approach of experimental measurements and photonic NN emulation. Our results show that classification accuracies of a common handwritten digit recognition task rival that of a software-based non-linear activation function.