Machine learning techniques to solve optical design challenges

Description

Data-driven approaches – propelled by the widespread success of machine learning and especially deep learning – have the potential to revolutionize optical design.

Optical system design is a large nonlinear optimization problem that is usually addressed using traditional local and global optimization methods. Traditional methods do not learn from data, and therefore must start from scratch for any new design process. A novel family of data-driven approaches, relatively unexplored, offer the ability to extract the features that make up a good design and reuse them to infer new designs – similar to how optical design practitioners develop and then exploit their intuition and knowledge.

Our research group, a pioneer in this field, wants to further develop a deep learning approach that extrapolates from previous designs to directly infer optical systems tailored to the desired specifications. Other approaches such as training an optimization algorithm may also be explored.