An Integrated All-Optical Multimodal Learning Engine Built by Reconfigurable Phase-Change Meta-Atoms

Optical computing is regarded as one of the most promising computing paradigms for solving the computational bottleneck and accelerating artificial intelligence in the post-Moore age. While reconfigurable optical processors make artificial general intelligence (AGI) possible, they often cannot process multimodal signals. Here, we propose an integrated all-optical multimodal learning engine (AOMLE) built by reconfigurable phase-change meta-atoms. The engine architecture can be mapped to different optical neural networks by laser direct writing for phase-change materials, enabling more efficient processing of visual and auditory information at the speed of light. The AOMLE provides a cutting-edge idea for reconfigurable optical processors with increasing demands for complicated AI models.