Inferring the solution space of microscope objective lenses using deep learning.

Area of Science:

• Optical Engineering
• Computational Optics
• Machine Learning Applications

Background:

• Lens Design Extrapolation (LDE) is a data-driven method for creating new optical systems based on reference designs.
• Existing LDE frameworks face challenges in generating diverse lens structures and handling novel lens sequences.

Purpose of the Study:

• To develop a deep learning-enabled LDE framework for generating a wide variety of microscope objective lenses (MOLs).
• To address the one-to-many mapping challenge in LDE and enable extrapolation to unseen lens sequences.

Main Methods:

• Formulated LDE as a one-to-many problem to generate varied lenses for given specifications and sequences.
• Quantified MOL structure using marginal ray slopes to improve training objectives.
• Utilized a dataset of 34 reference MOLs to train the deep learning model.

Main Results:

• Generated designs across 7432 lens sequences, demonstrating significant structural diversity.
• Inferred designs accurately captured the structural diversity and performance of the reference dataset.
• Successfully extrapolated to lens sequences not present in the training data.

Conclusions:

• The enhanced LDE framework effectively generates diverse MOLs with varied lens sequences.
• The approach overcomes limitations in one-to-many mapping and extrapolation to novel lens sequences in optical design.
• This work advances data-driven optical design by enabling the creation of structurally diverse and performant lens systems.