All-optical nonlinear activation function based on nonlinear absorption in silicon and germanium waveguides

The on-chip implementation of optical nonlinear activation functions is crucial for realizing large-scale photonic neural chips. The use of nonlinear effects from two-photon absorption (TPA) allows for the realization of broadband, high-speed all-optical nonlinear activation functions. Here, we exploit the strong TPA coefficients and CMOS-process maturity of Group IV photonics to experimentally demonstrate low-loss silicon-on-insulator (SOI) and germanium-on-silicon (Ge-on-Si) waveguides as all-optical nonlinear activation units. The devices exhibit fast response frequency up to 100 and 300 MHz, respectively. Using the implemented nonlinear activation function, we investigated the performance of convolutional neural networks in classifying the MNIST handwritten digit image dataset, achieving accuracies of 97.88% (SOI) and 97.74% (Ge-on-Si). This verifies the effectiveness of our approach in handwritten digit classification tasks.