Programmable optical activation functions for KANs via feed-forward integrated microring resonators

Activation functions are indispensable in optical neural networks (ONNs). However, existing solutions often require high optical power for nonlinearity or suffer from limited tunability that restricts deployable architectures. Achieving diverse and tunable activation functions remains a significant challenge. In this Letter, we propose a programmable activation function unit based on a feed-forward microring resonator with a tunable coupling coefficient on a silicon photonics platform. By leveraging the high degrees of freedom in tunability, the device can generate a diverse series of activation functions spanning both monotonic and non-monotonic profiles. This versatility extends its applicability from classical multilayer perceptrons to emerging architectures like Kolmogorov-Arnold Networks (KANs), which require learnable activation function shapes. We perform a detailed modeling analysis and experimentally validate the device's flexible generation capabilities. Furthermore, we construct a physics-aware KAN framework based on the device's solution space, achieving a classification accuracy of 98.1% on the MNIST dataset. We envision that this device will provide enhanced design freedom and pave the way for the monolithic integration of versatile ONNs.