Reconfigurable nonlinear activation functions in an optomechanical whispering-gallery-mode microcavity

Optical neural networks (ONNs) have attracted significant attention for artificial intelligence because of their inherent parallelism. As a key component of ONNs, all-optical nonlinear activation functions (ONAFs) are critical to neural network performance. Here, we theoretically investigate reconfigurable nonlinear activation functions realized by optomechanically induced transparency (OMIT) in a whispering-gallery-mode microcavity. The proposed system can implement distinct nonlinear activation responses, including a smooth saturating response and a ReLU-like response. These responses can be continuously tuned through all-optical operation with microwatt-level thresholds. In compact neural network simulations, the proposed activation functions improve classification performance over the linear baseline on representative binary tasks with nonlinear decision boundaries. Further evaluation on MNIST and Fashion-MNIST using a compact convolutional neural network (CNN) yields accuracies of 98.93% and 98.88% on MNIST, and 91.16% and 89.70% on Fashion-MNIST. These results support the optomechanical effect as a promising mechanism for on-chip reconfigurable nonlinear activation in photonic neural computing.