Machine learning inference tasks are extensively utilized across various industries. However, the escalating need for computing power and the corresponding energy consumption challenges have become more and more pressing. By capitalizing on the low-loss nature of electromagnetic wave propagation within media, optical analog computing can be realized through the careful design of media structures. This approach holds the promise of minimizing power consumption while simultaneously boosting computational bandwidth. In the context of edge computing, machine learning inference tasks impose exceptionally stringent requirements on both real-time performance and energy efficiency. Although optical analog computing may exhibit slightly lower accuracy, its rapid computational capabilities are well-suited to the strategies adopted by edge computing devices. These devices enhance inference speed by employing high quantization levels. Building on this foundation, the research team has introduced a design methodology that seamlessly integrates scattering media. This approach not only simplifies the manufacturing process and enhances structural stability but also facilitates efficient machine learning inference on chips through the implementation of intelligently crafted media superstructures.