Abstract
<title>Abstract</title> <p>Complex-valued convolutional neural networks can jointly represent amplitude and phase information, making them well suited for wave-based data such as radar echoes, communication signals, spectral responses, and electromagnetic fields. To accelerate such networks in the photonic domain, the computing system must natively load, transform, and read out complex-valued signals while maintaining stable matrix programmability under device imperfections. Here we propose, fabricate, and demonstrate a self-configuring complex-valued photonic convolution accelerator (SCPCA), which combines a native complex-valued photonic computing architecture with a measurement-driven self-configuration algorithm. The architecture uses an IQ modulator, a cascaded Mach--Zehnder interferometer mesh, and coherent MMI-assisted balanced detection, thereby preserving complex-valued photonic signals across the full computing chain. The self-configuration algorithm directly optimizes the experimentally measured end-to-end complex-valued matrix response, enabling adaptation to the cascaded MZI architecture without device-by-device pre-calibration and compensating for practical nonidealities. Experiments show that SCPCA reaches convergence in six iterations on average and achieves an effective mapping precision of approximately 5 bits. We further apply SCPCA to polarimetric synthetic aperture radar image recognition, achieving 84.98\% test accuracy on the AIRSAR San Francisco dataset, close to the 86.00\% electronic baseline. These results provide a practical route toward native complex-valued photonic acceleration with low calibration burden.</p>