Light-driven molecular reorientation for large-scale photonic in-memory computing

Photonics has the potential to transform high-speed and energy-efficient artificial intelligence. However, most photonic computing platforms still rely on costly electro-optic conversions for data input, weight loading and storage, resulting in substantial overhead during large-scale computation. Here, we present a large-scale photonic in-memory computing platform based on liquid crystal (LC)-enabled Poincaré-sphere-connected neural networks. Megabyte-level nonvolatile photonic memory is introduced through light-driven reorientation of sulfonic azo-dye and LC molecules, which also functions as the reconfigurable optical neurons. This architecture uniquely integrates optical sensing, memory, and computing in situ, allowing minimal data-movement overheads in various switchable machine vision tasks. Ultrahigh scalability is demonstrated by a 9.44-million-neuron photonic processor array that supports massively parallel optical computing, recording and retrieval, providing a memory capacity 100,000 times larger than state-of-the-art photonic in-memory computing. This work offers a novel pathway toward scalable submicrometer-precision photonic memory, thus establishing a new paradigm for physical neural networks and brain-like artificial intelligence.