Neural Network Design of Broadband Epsilon-Near-Zero Perfect Absorbers

In the past decade, optical metamaterials have demonstrated the capability to efficiently manipulate the phase, polarization, and emission of light with the usage of nanostructures – a fundamental shift from conventional refraction and propagation optics. The recent use of epsilon near zero (ENZ) materials has enhanced the capabilities and possibilities of metastructures, leading to unique optical properties such as perfect light absorption, non-reciprocal magneto-optical effects, and large optical nonlinearity. By constructing multilayer ENZ thin films and metal-oxide-semiconductor ENZ heterostructure, broadband and tunable ENZ properties and perfect absorption can be achieved [1]. However, a critical challenge in developing broadband and tunable ENZ metastructures is the complex design of ENZ multilayer stack or heterostructure involving different physical parameters such as ENZ wavelength, thicknesses, optical losses, carrier concentrations, material options, and excitation mediums and angles. Finding the optimal parameters of ENZ heterostructures is a complex issue because changing one parameter adversely affects another.<br/><br/>In this work, we demonstrate the first use of a residual generative neural network [2] to optimize broadband and tunable perfect absorption properties of ultrathin ENZ materials. The code for the neural network is written in PyTorch, a machine learning tool for python, and runs using Google’s high-performance graphics processing units (GPUs). The neural network determines the thickness of the thin film ENZ layer as well as the ENZ material parameters, such as carrier density. From here, a physics simulator returns the absorption, reflectance, and transmittance spectra of the ENZ thin film stack and is subsequently evaluated by a loss function – determining whether the network’s outputs were optimal or not. As the neural network continually trains, the global optimal thickness and materials are outputted by the network. Our initial results using the experimentally obtained ITO material properties demonstrate that the neural network can generate three-stack layers with a maximum absorption above 99.99% and a broad absorption (&gt;80%) bandwidth hundreds of nm wide. These results in machine learning designed ENZ materials and heterostructures will provide broadband and tunable perfect absorption in ultrathin thickness (&lt;10 nm), favorable for minimizing the dimensions in applications such as the absorbing layer in CMOS image sensors, smart-window layers, perfect absorbing layers for spacecrafts, and solar cell generation nanolayers, thus this will make a significant impact on the performance of imaging/display, solar/thermo-photovoltaics, and optical communication technologies.<br/><br/>References<br/>¹A. Anopchenko, <i>et al.</i>. Field-Effect Tunable and Broadband Epsilon-Near-Zero Perfect Absorbers with Deep Subwavelength Thickness, ACS Photonics 5, 2631 (2018).<br/>²J. Jiang, J. Fan. Multiobjective and categorical global optimization of photonic structures based on ResNet Generative Neural Networks. Nanophotonics, 10, 361 (2021).