Column-streaming optical convolution with programmable MZIs processor

Programmable photonic processors based on Mach–Zehnder interferometers (MZIs) provide a versatile platform for optical linear computing, yet convolution on such processors is still commonly implemented through high-dimensional matrix expansion or multi-processor decomposition, both of which incur substantial hardware and control overhead. Here, we present a column-streaming optical convolution scheme that maps an N × N convolution kernel directly onto a single programmable N × N MZIs processor. By transposing the kernel to match the column-wise input order, the proposed scheme decomposes two-dimensional convolution into a sequence of low-dimensional linear operations, followed by backend shift-add accumulation of the output. Using a fabricated silicon photonic processor, we experimentally demonstrate accurate matrix configuration and optical convolution under different kernels, with root-mean-square errors below 0.061 relative to the electrical reference results. We further construct a hybrid optoelectronic convolutional neural network for MNIST handwritten digit classification and achieve an inference accuracy of 92%. These results establish a compact route to kernel convolution on programmable photonic processors and highlight the potential of column-streaming photonic dataflow for integrated optical neural computing.