Augenstein, Yannick 1,2 1 Institut für Theoretische Festkörperphysik (TFP), Karlsruher Institut für Technologie (KIT) 2 3D Matter Made to Order (3DMM2O), Karlsruher Institut für Technologie (KIT)
Abstract (englisch):
Inverse problems are ubiquitously encountered throughout science and engineering.
Where the forward problem answers the question of what the output for a given input looks like, the inverse problem tries to answer the opposite: given a set of outputs, what were the inputs? While the forward problem is typically uniquely defined and can be solved through numerical modeling, the inverse problem is generally ill-posed, making its direct solution intractable. Inverse design is a class of methods that aim to solve the inverse problem, at least to a "good enough" approximation, by computational optimization of a mathematically defined objective function.
Topology optimization, in particular, is a gradient-based method for inverse design.
The method has gained popularity in photonics in the past decade and has led to the creation of devices with non-intuitive designs and exceptional performance. This thesis applies topology optimization to designing various nanophotonic devices, from two-dimensional structures that manipulate and guide surface waves to fully free-form and three-dimensional devices such as fiber-to-chip and grating couplers. ... mehr We find that while topology optimization and additive manufacturing via methods such as 3D laser nanoprinting ideally complement each other, creating fabrication-ready free-form nanophotonic devices presents unique challenges. We identify the issue of structural integrity and develop a method for coupled mechanical and electromagnetic inverse design, demonstrating that this approach can yield more feasible devices for fabrication.
Lastly, we focus on the issue of computational cost - topology optimization typically involves many iterations of computationally expensive numerical simulations, which can limit the extent to which the design space can be explored. We develop a framework for the inverse design of nanophotonic devices via a neural operator-based surrogate solver and apply it to optimize free-form electromagnetic scatterers. As these surrogate solvers are trained on data obtained from numerical simulations, we discuss the trade-offs in terms of generality and accuracy and examine the problem settings in which such trade-offs can be feasibly made.