Optical convolution computation: principle, applications, challenges

Convolution underpins modern deep learning yet dominates end-to-end computation and data movement, exacerbating the "memory wall" in post-Moore electronic architectures. Optical convolution accelerators have re-emerged as a credible path forward, driven by the maturation of photonics that can natively exploit parallel propagation, multiplexing, and linear superposition. Despite rapid progress, the literature remains fragmented, characterized by heterogeneous physical mechanisms, incomplete CNN operator support, and inconsistent evaluation practices. This Review organizes the field through a unifying taxonomy rooted in two mathematical principles: (i) definition-based optical convolution that directly map multiply–accumulate onto optical modulation, and (ii) theorem-based optical convolution that implement transform-domain multiplication via Fourier optics and spectral filtering. We synthesize representative architectures spanning diverse physical mechanisms, clarify shared system-level signal chains in hybrid opto-electronic implementations, and discuss practical performance determinants including precision, scalability, calibration, and full-chain energy accounting. Finally, we survey application regimes by data dimensionality—from wideband RF signals to high-order tensor processing—and outline the key device-algorithm co-design challenges that will shape deployable optical convolution hardware.