Reprogrammable All-Optical Ultra-Compact Nonlinear Activation Function for Neuromorphic Computing on Indium Phosphide Membrane on Silicon

In this work we present the characterization and numerical study of a reprogrammable all-optical ultra-compact nonlinear activation function for neuromorphic computing, implemented on the Indium Phos- phide Membrane on Silicon (IMOS) platform. The device consists of a Mach–Zehnder coupler (MZC) followed by a microring-resonator (MRR)-nested Mach–Zehnder interferometer (MZI). The nonlinear behavior is driven by the MRR, while three independently tunable degrees of freedom - the MZC phase shifter, the MRR reso- nance, and a post-MRR phase shifter - enable dynamic reprogramming of the transfer function. By exploiting ultra-compact phase shifters, the architecture achieves a footprint of 0.054 mm2 and a −5 dBm triggering power. We experimentally characterize the resonant response and tunability of the device, reporting a thermo-optic tuning efficiency of 0.9598 nm/mW. Without any transfer-function optimization, the device reproduces a Leaky Rectified Linear Unit (ReLU) response, as commonly found in modern artificial neural networks when computa- tional efficiency and sparsity are required. Biasing the phase shifters allows the device to reproduce additional transfer functions such as Radial Basis Functions (RBFs) and sigmoidal shapes. These results highlight the potential of compact, reprogrammable photonic structures on the IMOS platform to realize nonlinear activation functions for neuromorphic computing.