Название: Nanophotonics and Machine Learning: Concepts, Fundamentals, and Applications
Автор: Kan Yao, Yuebing Zheng
Издательство: Springer
Серия: Springer Series in Optical Sciences
Год: 2023
Страниц: 188
Язык: английский
Формат: pdf (true)
Размер: 14.3 MB

This book, the first of its kind, bridges the gap between the increasingly interlinked fields of nanophotonics and Artificial Intelligence (AI). While Artificial Intelligence techniques, Machine Learning in particular, have revolutionized many different areas of scientific research, nanophotonics holds a special position as it simultaneously benefits from AI-assisted device design whilst providing novel computing platforms for AI. This book is aimed at both researchers in nanophotonics who want to utilize AI techniques and researchers in the computing community in search of new photonics-based hardware. The book guides the reader through the general concepts and specific topics of relevance from both nanophotonics and AI, including optical antennas, metamaterials, metasurfaces, and other photonic devices on the one hand, and different Machine Learning paradigms and Deep Learning algorithms on the other. It goes on to comprehensively survey inverse techniques for device design, AI-enabled applications in nanophotonics, and nanophotonic platforms for AI. This book will be essential reading for graduate students, academic researchers, and industry professionals from either side of this fast-developing, interdisciplinary field.

The study of interactions between light and materials has a long history, dating back to perhaps as early as the time even when the nature of light had not been settled. As science keeps advancing, there is one line in this study that can be traced by looking at the decreasing dimensions of the materials, from optics to photonics, and all the way down to nanophotonics. Nanophotonics studies light-matter interactions at the nanoscale, where the materials in most cases are structured into subwavelength building blocks so that exotic optical properties beyond those of bulky materials emerge. Over the past two decades, nanophotonics has attracted rapidly growing interest and become a vibrant research field that contains both fundamental and application-driven studies. Depending on the materials, geometries, sizes, and arrangements of the constituent elements, nanophotonics can be categorized into several subfields, including plasmonics, metamaterials and metasurfaces, photonic crystals, photonic-integrated circuits, and other resonant nanostructures that can perform photonic functions. These devices operate on different mechanisms, enabling unprecedented opportunities to control light at the nanoscale for unveiling new physics and achieving fascinating applications not possible with conventional techniques.
Artificial Intelligence (AI), seemingly on a totally different subject to nanophotonics at a first glance, is currently among the most promising techniques that can revolutionize the world from many aspects. The history of AI is a bit less than 80 years, with the beginning marked by the early research on neural networks in the 1940s. The popularity of AI nowadays has gone far beyond Computer Science and infiltrated many other research fields, such as physics, chemistry, materials science, and biomedicine, to name a few. After the astonishing success of the computer program AlphaGo from Deepmind defeating the top professional Go players, even the public might develop the idea that a new era, where AI is competitive with human intelligence in completing certain tasks (i.e., weak AI), has come. In science and engineering, especially those fields fitting well with Big Data, the expected tasks include materials discovery, drug creation, and so forth.

It is interesting to picture how AI would transform nanophotonics. Although surely not a panacea for all the remaining challenges, AI can potentially assist the design of nanophotonic devices. Conventional inverse design is based on a trail-and-error process, which can be extremely labor-intense. The initial guess of the solution to a design task usually relies on human knowledge, a combination of intuition, the physical insights revealed from the study of modal systems, the experience accumulated during the previous practice, and reasoning. It is then examined with simulations by solving Maxwell’s equations but unlikely to meet the desired performance in one shot. Therefore, adjustments to a handful of parameters and re-evaluation of the new designs need to be repeated until some preset criteria are reached. A variety of optimization methods have been developed to prevent this process from being almost blind. They also make it possible to search the enormous design space in a more comprehensive manner, yielding complex and non-intuitive structures that cannot be parameterized. But still, the conventional workflow requires considerable computation power and time for every design task, which could explode as the complexity of the devices and the scale of integration increase. AI or, more specifically, Machine Learning provides new solutions that work in a totally different logic. Through training, these so-called “data-driven” methods leverage many instances of known devices to improve the ability of finding optimized designs for a certain set of design tasks. The questions on whether and how AI will benefit inverse design remain fairly open. On the one hand, no clear evidence has suggested that the efforts in generating sufficiently large training sets for AI programs can be less fierce than those in the trial-and-error and optimization process. On the other hand, there certainly exist AI-related techniques, both algorithms and hardware, that can change the game in some way. In a nutshell, the application of AI in nanophotonics, including but not limited to inverse design, appears worthwhile and deserves more research efforts.

What makes the combination of AI and nanophotonics more interesting is the other side of the coin. With the explosive growth of Machine Learning in recent years, the computing hardware based on general-purpose processors becomes inefficient in implementing neural networks, raising the pressing need to develop application-specific hardware. Compared with the solutions based on electronic architectures, photonic circuits that can process coherent light signals are superior in speed and power efficiency. Some recent advances have demonstrated that specially designed nanophotonic circuits or structures can perform machine learning tasks like inference. Therefore, nanophotonics is not just fueled by AI passively; it offers improvement in return, making their relationships interactive.

Because the backgrounds of these two fields are very different, there is often a knowledge gap for people interested in this topic from either side. The goal of this book is thus to introduce the basics of nanophotonics and Machine Learning, especially Deep Learning, and to help the reader to get some sense on how they work and can be utilized to enhance each other.

Contents: