Reconfigurable optical nonlinear activation functions using a phase-change microring for photonic neural networks

The implementation of reconfigurable nonlinear activation functions (NLAFs) is a key challenge in advancing photonic neural networks to model complex relationships. This study addresses this issue by introducing a device that exploits the phase-dependent optical properties of a Ge2Sb2Te5 superlattice (GST-SL) integrated on a silicon microring resonator. By manipulating the distinct absorption losses and refractive indices of the GST-SL across its crystallization states, we experimentally demonstrate five types of NLAFs: half-sigmoid, ELU, softplus, ReLU, and radial basis. The characterized functions exhibit a low activation power threshold of 1.76 mW and a high operational speed of up to 1 MHz. In a system-level benchmark using a three-convolutional-block neural network for MNIST handwritten digit classification, our activation functions boost the classification accuracy from 95.35% to 99.24%. These results highlight the promising performance and application potential of our device for future on-chip photonic neural networks and advanced optical computing.