Enter the effective focal length, f-number and half field of view desired for your lens design project. Our deep learning framework will infer a selection of lens designs tailored to those specifications.
For every query made, we filter the lens structures that fit the desired specifications and pass on the specifications to our trained deep neural network, which infers a number of lens designs for every selected lens structure. Then, out of all designs from a given structure, we showcase the one with the lowest RMS spot size and provide a quick overview of the lens aberrations. The Code V (.seq) and Zemax (.zmx) files of the designs are available for download.
Our deep neural network was trained to optimize the RMS spot size of the designs based on the C, d and F wavelengths in the visible spectrum, by extrapolating from 150 reference designs across 80 different lens structures. Though we cannot completely prevent ray failures or overlapping surfaces, the vast majority of inferred designs should be devoid of those. The glass materials are fictitious materials, but our training objective favors glass materials that are close to the ones in the popular Schott catalog, as well as reasonable diameter-to-thickness ratios.
For personal communication regarding LensNet, please contact Geoffroi Côté at
The authors gratefully acknowledge funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) and Sentinel North.