Author : Dianjing Liu
Publisher :
ISBN 13 :
Total Pages : 0 pages
Book Rating : 4.:/5 (126 download)
Book Synopsis The Application of Machine Learning for Designing and Controlling Electromagnetic Fields by : Dianjing Liu
Download or read book The Application of Machine Learning for Designing and Controlling Electromagnetic Fields written by Dianjing Liu and published by . This book was released on 2021 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: Machine Learning is the study of computer algorithms that improve automatically through experience. In contrary to rule-based artificial intelligence which produces pre-defined outcomes based on manually coded rules, machine learning algorithms aimed at building models and making decisions based on the sampled data, and without explicitly programmed to do so. Recently, deep neural network-based machine learning algorithms achieved great success in many applications including image recognition, speech recognition, natural language understanding, etc, while their potentials in other domains are to be explored. In this thesis, we explore the application of deep-learning-based algorithms for designing and controlling electromagnetic fields. Firstly, we design the nano-scale structure of the optical medium to change its interaction with the electromagnetic field. This process is called inverse design and is a common problem in nanophotonics. Since an optical property can be achieved by more than one structure, the same design request can have multiple candidate solutions. This issue is called non-uniqueness and it fundamentally makes the direct training of an inverse design neural network hard to converge. We propose a deep-learning-based approach to overcome the non-uniqueness issue and train a neural network as an inverse design toolbox. Once the model is trained, it generates a design for input requests in a fraction of a second without needing any iterative optimization. Another application in photonics is the spontaneous development of the imaging system and the neural network. Typically in deep learning algorithms, the inputs to the neural networks are handcrafted representations of the data. For example, a fully connected neural network requires manually created feature vectors as the inputs. Compared with the fully connected network, the convolutional neural network can process the raw pixel values (i.e., the digital image) and therefore requires less feature engineering. However, these digital images are collected by sensory functions (usually a camera) which are also designed by human intelligence. Here we set up a reinforcement learning agent with the ability to develop a sensory function by itself. We show that although the agent does not have a functional visual sensor to observe the environment at the beginning, it is able to automatically develop parabolic imaging optics and detect a clear visual representation of the environment. Finally, we apply machine learning algorithms for the controlling of electromagnetic fields. A reinforcement learning agent controls the electromagnets to manipulate the spatial distribution of the magnetic field. We demonstrate that this field manipulation is able to levitate and control a magnetic object. The reinforcement learning agent develops the control strategy from experiences and under the guidance of the rewards. The trained agent shows good control skills, and is faster, and has less overshoot compared with the traditional PID controller.
