The advancement of nanotechnologies enables powerful control of photons by subwavelength structures. In recent years, rapid advancement ofmetasurfaceand metamaterials reveal the potential ofnanophotonicsin the applications across disciplines, from hyperspectral imaging to mathematical operations. One question emerges as: ismetasurfaces’ applications limited in deterministic spatial/spectral information or it can play more powerful roles in machine learning and dealing with uncertainties. In this talk I will review recent works on this track and introduce integrated photonicmetasurfacesystems. With lithographically defined inter-layer alignment, we demonstrate diffractive deep optical network on silicon photonic platform, towards broadband spatial pattern classification and hyperspectral imaging. The high-throughput vector-by-matrix multiplications is enabled by 103passive subwavelength phase shifters as weight elements. The integrated metasystem perform analogue optical computing tasks, from simplefouriertransformation to complicated image classification. In the presentation, we will illustrate the design principle of the foundry compatible metasystem, and its implementation of basic low loss photonic mode converters, differentiators, and image classifiers.