Deep-learning reconstruction of aspheric microlens arrays from intensity maps

Designing complex aspheric surfaces is challenging, particularly in achieving precise control over the light path. Our research addresses this by using a deep-learning approach to reconstruct dielectric aspheric microlens arrays. We formulate this as an image-to-image translation task, training a neural network to convert measured light intensity maps, obtained with a rendering engine by illuminating the sample from multiple angles and wavelengths, into spatial maps of the sample's surface. Utilizing physically accurate rendering software and a microlens array template, our method achieves highly accurate reconstructions. These results demonstrate that deep convolutional networks can stably reconstruct complex aspheric surfaces from intensity-only measurements, suggesting a practical route toward data-driven surface metrology for microlens arrays and related refractive elements, and highlighting the potential of deep learning to advance aspheric and freeform optics.