End-to-end all-optical in-sensor computing system using photonic integrated circuits

Abstract

Photonic sensors play an increasingly important role in chemical analysis, but conventional systems rely on sequential wavelength scanning and extensive electronic post-processing, leading to large data redundancy, high latency, and excessive energy consumption. Here, we introduce an end-to-end all-optical in-sensor computing system based on photonic integrated circuits that merges sensing and computing in the optical domain. The system integrates a photonic waveguide sensor and a microring weight bank on a single chip for linear processing, while nonlinear activation is implemented using an erbium-doped fiber amplifier. By performing spectral compression, this system enables real-time optical-domain computing with up to 6-bit precision. The system achieves a classification accuracy of 94.2% across 27 liquid-mixture classes and concentration prediction for mixture chemicals. Compared with conventional photonic sensing systems, the proposed architecture reduces inference latency by a factor of 4.48 and energy consumption by a factor of 11.47. These results establish a compact, low-redundancy, and energy-efficient paradigm for intelligent photonic sensing at the edge.