Speaker
Description
Direct Laser Writing (DLW) is the basis of maskless photolithography for the fabrication of high-resolution diffractive optical elements (DOEs) and coded structures for long-term storage media. Traditional DOE design relies on analytical calculations, electromagnetic modeling, and iterative parameter optimization. For free-form geometries, the configuration space grows rapidly, which leads to significant computational costs and complicates the search for structures that simultaneously satisfy optical and technological requirements.
Inverse design based on machine and deep learning allows replacing repeated iterative search by building a relationship between the target optical response and the geometry of the element. Such methods (tandem networks, generative-adversarial networks, conditional diffusion models) are successfully used in integrated photonics and metasurfaces. At the same time, they have not been adapted to the technological context of DLW for optical information storage.
The development of deep learning methods and models for automated inverse design of DOE geometry taking into account the constraints of the real recording process is a relevant task. The following stages were proposed: (1) selection of a DOE class compatible with DLW equipment and formalization of target characteristics (diffraction efficiency, angular intensity distribution) as optimization problems; (2) construction of a direct diffraction model and formation of a dataset of “geometry - optical response” pairs; (3) formalization of DLW technological limitations as model constraints or components of the loss function, which combines the deviation of the optical response from the target with a penalty for violating technological limitations; (4) comparison of a tandem network, a generative-adversarial network, and a conditional diffusion model with a physically informed component in terms of response reproduction accuracy, speed, and variety of solutions relative to classical methods; (5) manufacturing prototypes on the DLW installation, measuring their optical response, and correcting the direct and inverse models in a closed loop of “design → manufacturing → measurement → correction”.
The proposed approach is expected to substantially reduce the design cycle of new DOEs by replacing repeated manual optimization with automated generation and subsequent verification. It can also expand the range of accessible geometries through free-form design and provide a reproducible set of data and models for further development, in particular for increasing the recording density of long-term optical storage media.
| Type of presence | Presence online |
|---|