Multi-Angle Focusing Metalens Designed via KAN Deep Learning for Expanding the Reception Field of MEMS LiDAR

Abstract

The small receiving field of view of the detector often limits the scanning angle of the MEMS LiDAR. This paper designs and experimentally verifies a multiangle focusing metalens to tackle this challenge, which can focus light at various incident angles near its optical axis, unlike a conventional focusing metalens. To simplify the design process, we propose a novel neural network architecture that integrates Kolmogorov–Arnold Network (KAN) modules, Multilayer Perceptron (MLP) modules, Convolutional Neural Network (CNN) modules, and Transformer structures. This architecture enables accurate and rapid predictions of the phase and transmission efficiency for various meta-atoms at different incident angles, achieving prediction accuracies exceeding 95%. Considering fabrication constraints, the designed neural network can generate a library of 600 simple geometric meta-atoms in less than 1 min. Using this library, we simulated and experimentally validated the multiangle focusing metalens. Results show that with a diameter of 2 mm and a focal length of 1 mm, the metalens can effectively focus light at incident angles from 0 to 30°, with a maximum focal length shift of only 20 μm, one-tenth of that of conventional metalenses with the same parameters. This demonstrates that the designed metalens can significantly enhance the receiving field of view for MEMS LiDAR detectors and similar systems.