All-Optical Nonlinear Activations Using Saturable Absorbers on a Generic InP Platform for Photonic Neural Networks

We experimentally demonstrate all-optical nonlinear activation functions using standalone saturable absorbers, as well as saturable absorbers integrated into microring resonators and Mach-Zehnder interferometer structures on a monolithic InP platform. These configurations enable programmable nonlinear responses and support both passive and active operation modes. The measured activation characteristics exhibit softplus- and ReLU-like behaviors, suitable for neuromorphic computing. When incorporated into a convolutional neural network for MNIST digit classification, the system achieves over 97% accuracy, highlighting the potential of SA-based photonic components for compact, low-power, and scalable optical neural networks.