A Sb2S3-based reconfigurable diffractive optical neural network (RDONN) for on-chip integration is proposed. The RDONN can be integrated into standard silicon-on-insulator systems, offering a compact, passive, all-optical solution for implementing machine learning functions. The weights of the proposed optical chip are reconfigurable without the need to modify hardware structures or re-fabricate the chip. Its main structure consists of multilayer metalines made from Sb2S3, a low-loss phase change material. The RDONN architecture is constructed using the two-dimensional electromagnetic propagation model and implements the classification task on the Iris dataset with both intensity modulation and phase modulation inputs. This demonstrates its feasibility, with classification accuracies reaching 95.0% and 98.3%, respectively. Our model enables reconfigurable manipulation of the weights in the on-chip diffractive optical neural network, which can be used in the design and fabrication of real chips. This advancement holds significant promise for future all-optical in situ learning systems.
Read full abstract