High-precision integrated diffractive optical networks enabling regression analysis

Abstract

The growing demands of artificial intelligence impose tremendous challenges on computing hardware. Integrated diffractive optical networks (IDONs) have emerged as a promising candidate for next-generation computing architectures, offering ultrafast processing speed, superior energy efficiency, and inherent parallelism. However, phase-error accumulation and fabrication-sensitive metasurfaces limit their physical precision, restricting most demonstrations to classification tasks. Here, we implement a high-precision IDON by incorporating hardware-level error compensation. Cascaded metasurfaces provide high computational capacity through densely integrated optical modulation units, while integrated thermo-optic tuning arrays enable in-situ calibration to accommodate fabrication nonidealities. Numerical analysis reveals that the proposed approach suppresses the relative error of matrix multiplication from 37% to 4.7%. Leveraging this improvement, we experimentally achieve regression on the Boston Housing dataset with an R² of 0.71, comparable to that of state-of-the-art digital implementations. The demonstrated IDON represents a significant advance toward high-precision optical computing. Integrated diffractive optical networks offer fast and energy-efficient computing, but fabrication and environmental variations limit their precision. Here, the authors introduce hardware-level in-situ error compensation, enabling high-precision optical computing and regression analysis.