Spatiotemporal Computational Meta-Imager for Broadband Dual-Layer Optical Convolution.

Optical neural networks face a tradeoff between device footprint and computational expressivity: deep systems are bulky, whereas single-layer hardware lacks intrinsic depth and native negative-weight support. We experimentally demonstrate a single-layer metalens imager with "spatiotemporal intrinsic depth" for equivalent broadband dual-layer convolution. A polarization-multiplexed metasurface provides fixed Bessel-vortex channels (l = 0 to 3) as the immutable first convolution layer, where the (l = 0) channel implements a hardware-native residual connection. A metasurface-generated 3 × 3 optical vortex matrix forms the second functional layer; time-sequential liquid-crystal amplitude modulation delivers tunable positive/negative weights. This temporal encoding solves the optical sign problem and converts fabrication imperfections into compensable biases via in-situ training. Experiments verify rotational robustness for fashion-item classification, raise full-color CIFAR-10 accuracy from 57.5 % to 86.6 %, and reach 99.3 % on the SIPaKMeD medical dataset, offloading 99.7 % of convolutional multiply-accumulate operations to optics. By removing volumetric diffractive layers, this work enables low-latency hybrid optoelectronic edge intelligent vision.