Generative Deep Learning for Inverse Design: Photonics and Beyond
Author(s)
Zhu, Dayu
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The capacity to manipulate the light waves is largely governed by the availability of photonic materials and structures at our disposal. Over the past two decades, the exploration of artificially structured photonic media, notably metamaterials and metasurfaces, represents a central theme in optical science. However, the intricate nature of light-matter interaction at the subwavelength scale results in the design of meta-structures, to date, still largely relying on tedious trial-and-error process with iterative parametric sweeping. This conventional practice, along with existing optimization schemes in the current stage, falls short when it comes to the discovery and design of highly complicated meta-structures or multiplexed optical functionalities.
Addressing this limitation, we introduce a generative deep learning methodology that revolutionizes the inverse design paradigm in photonics. Machine learning, especially deep learning burgeons into an essential asset for the scientific community, fostering the integration of artificial intelligence (AI) into cross-disciplinary research. Very recently, generative deep learning stands out in the inverse design topics, for its capability to produce counterintuitive and high degree-of-freedom candidates, which has the potential to surpass the capacity of human-centric design and traditional optimization process.
In this work, with a state-of-the-art generative model, we are able to achieve on-demand design of the unit cells of metasurface, i.e. meta-atoms, with free-form patterns and unconventional functionality. Further, incorporating evolutionary strategies enhances our ability to efficiently architect meta-molecules composed of distinct meta-atoms, as well as a new opportunity to explore the fundamentals of light-matter interaction. Moreover, we have developed a versatile deep learning framework for the comprehensive inverse design of multi-layer, multifunctional meta-systems, which is hardly reachable by alternative methodologies. Besides, our comprehensive machine learning framework is also extendable to a breadth of inverse problems across diverse fields. We have applied our methodology into non-line-of sight imaging, and the inverse retrieval of the onset time of acute stroke, showcasing its potential as a transformative tool in optical imaging, medical science, and beyond.
Sponsor
Date
2023-12-08
Extent
Resource Type
Text
Resource Subtype
Dissertation