Inverse Design Optimization and Fabrication of Nanophotonic Devices

Hadibrata, Wisnu

doi:https://doi.org/10.21985/n2-z3yd-bc10

Work

Inverse Design Optimization and Fabrication of Nanophotonic Devices

Public

Download PDF

With the ever-increasing demand for more complex functionalities and miniaturization of photonic devices, the design of such devices requires a whole new different approach, deviating from classical photonic design approaches. Conventionally, the design of a typical photonic device starts with a prior knowledge, intuition, and practical experience of specific behavior of a general structure. The designer then fine-tunes the geometrical parameters to achieve closely matching functionalities. This approach has seen a long list of successes in real life applications throughout the years. Generally, the intuition-based direct search is often employed for structures with a few design parameters to tune, due to the constraint of computational resources. As the number of design parameters increase, so does the requirement for the computational calculations. Therefore, the continuously increasing complexity of the device design makes the conventional intuition-based method prohibitive.Inverse design aims to solve this problem by providing a much more efficient alternative to device design via optimization schemes. The final device design is obtained via optimization by minimizing (maximizing) the objective of the desired functionalities with respect to the design parameter space. This potentially saves computational resources by significantly reducing the number of simulations to be performed to arrive at optimal design. In this dissertation, we explore different new and upcoming approaches in design optimization including objective-first and adjoint inverse design. Additionally, we will explore the use of deep neural network in the context of inverse design for nanophotonic devices. This dissertation begins with a short overview of an objective-first inverse design. Using the method, a metalens and an on-chip spectrometer were designed, fabricated, and characterized. As a proof-of-concept, the metalens was utilized in a homemade two-photon polymerization setup. A compact spectrometer combining random scattering medium based on tailored disordered modes and an inverse designed mode decomposer was also fabricated. Both devices were successfully fabricated using two-photon polymerization (TPP) direct laser writing. In addition to the objective-first method, in which the optimizations are strictly derivative free, an adjoint method is also introduced. In this method, the gradient of the objective function with respect to the full space of design parameters can be calculated using only two full-field simulations. Two device examples were optimized using the adjoint method, namely a circular polarizer and a beam deflector. For experimental demonstration, a circular polarizer was fabricated using the same TPP method. Lastly, we attempt to utilize deep learning approach to construct an inverse design network for predicting optimal design parameters. We first train a forward network to accurately predict optical responses of cylindrical meta-atoms. Then, we formulate an inverse model to retrieve the design parameters for an arbitrary desired optical response. We show that by implementing a physics-based data-preprocessing approach, the models are still capable of prediction outside the training spectral range. We implement a similar approach to a study of DNA-assembled programmable atom equivalent (PAE) metasurfaces. We train a forward network to predict the S-parameters of the PAE system, from which variety of optical responses can be calculated. Unlike the previous approach, we then use a global and gradient descent optimization method to retrieve optimal design parameters for various desired optical response.

Creator