Acoustic wave control is an important issue in living environment. Designing metasurface on scatterers is expected to control the sound field. However, an effective method to design the metasurface for large regional control is still lacking. Here we propose a machine-learning optimized method to solve problem of designing metasurface. According to the relationship between sound pressures at multiple points, convolutional neural network (CNN) is used to establish the mapping from local sound field to phase gradient of metasurface, which is further optimized by another CNN. The machine-learning method on designing metasurface has higher accuracy than the genetic algorithm. Using the machine-learning optimized method, not only the phase gradient of the metasurface can be obtained according to sound field, but also regional control of local sound field can be realized. For example, we can realize 8.37 dB intensification and 1.50 dB weakening of sound field at a square with a half-wavelength side. The metasurface designed by our proposed method is expected to realize noise reduction in large space, opening an avenue to achieve complex wave manipulation.

声波控制是生活环境中的重要问题。在散射体上设计超颖表面有望控制声场。但是，仍然缺乏有效的方法来设计用于大型区域控制的超颖表面。在这里，我们提出了一种机器学习优化的方法来解决设计超曲面的问题。根据多点声压之间的关系，使用卷积神经网络（CNN）建立从局部声场到超表面的相位梯度的映射，并通过另一个CNN对其进行优化。设计超表面的机器学习方法比遗传算法具有更高的精度。使用机器学习优化方法，不仅可以根据声场获得超表面的相位梯度，而且可以实现局部声场的区域控制。例如，我们可以在半波长边的正方形处实现8.37 dB的增强和1.50 dB的声场减弱。通过我们提出的方法设计的超颖表面有望在大空间中实现降噪，从而为实现复杂的波操纵开辟了道路。