Design And Optimization Of Dynamic Thermal Emission Nanostructure Based On Deep Learning

Dynamic thermal control technology has attracted wide attention due to its potential in radiation cooling,thermal switching and adaptive camouflage.However,the advanced performance of dynamic thermal control devices is still far below expectations;In addition,the inverse design of high efficiency dynamic thermal control devices faces the challenges of high computational cost and complex technical constraints.Based on machine learning,this thesis uses neural network to optimize and find the device structure with excellent performance.It proves the feasibility of using machine learning technology to carry out complex optimization in multi-dimensional space,and also provides a method for using machine learning to infer and design photonics structures with other performances.The main contents of the article are as follows:(1)Customized to the special and stringent requirement of dynamic emitters,a neural network model is developed to effectively bridge the structural and spectral spaces of the thermal emitter at different states and further realize the inverse design by coupling to genetic algorithms,which not only takes account of the broadband spectral responses at both states but also utilizes the spectral decomposition and transfer learning methods to ensure the modeling accuracy.The above method is used to optimize the metasurface of W/Si/VO2 structure,and the dynamic tunability reaches 0.8,showing an excellent performance rarely seen so far.Besides the quantitative regression,the empirical rules have also been mined qualitatively through the machining learning toolkits such as decision trees and gradient analyses,which afford beneficial guidance to the rational design.The study demonstrates the feasibility of using machine learning to obtain the near-perfect performance of dynamic emitters in multi-dimensional spaces,and also provides opportunities for designing other thermal and photonic nanostructures with machine inferences.(2)In order to further realize adaptive thermal control based on dynamic emitters,a global optimization neural network is proposed for the design of adaptive thermal control devices based on thin-film structure.The neural network is based on residual network scheme and does not rely on training data set,which combines the generated neural network with the transfer matrix method(TMM).The gradient information returned by TMM guides the generation of the multilayer structure according to the expectation.The material and thickness of each layer of the film are generated by the automatic optimization of the network,the material library provides 12 types of materials for network selection.Besides,the specific impact of key network parameters on network performance are also explored.As the verification and demonstration of the optimization ability of the network,we use three different design methods to design several multilayer devices for adaptive thermal control,the solar absorption ratio and dynamic tunability of which can reach below 0.2,and greater than 0.8,respectively.