Machine Learning Designed and Experimentally Confirmed Enhanced Reflectance in Aperiodic Multilayer Structures
Prabudhya Roy Chowdhury, Krutarth Khot, Jiawei Song, Zihao He, David Kortge, Zherui Han, Peter Bermel, Haiyan Wang, Xiulin Ruan
Abstract
Abstract Machine learning (ML) methods have gained widespread attention in optimizing nanostructures for target transport properties and discovering unexpected physics. Here, an ML method is developed to discover and experimentally confirm binary CeO 2 ‐MgO aperiodic multilayers (AMLs) for thermal barrier coatings with significantly enhanced reflectance. The effect of varying AML design parameters like total thickness, average period, and randomness in layer thicknesses, on the spectral and total reflectance is demonstrated. Introducing aperiodicity in layer thicknesses is shown to lead to a broadband increase in spectral reflectance due to photon localization. Since the number of possible AML structures increases exponentially with total thickness, a Genetic Algorithm optimizer is developed to efficiently discover AMLs with enhanced reflectance, for total thicknesses of 5–50 µm. Surprisingly, all the optimized structures show an odd number of layers with a CeO 2 layer at both ends, deviating from the traditional way of designing binary superlattices with paired layers. The optimized AML and a reference periodic superlattice of 5 µm thickness are fabricated by Pulsed Laser Deposition and characterized by optical reflectance measurements. The fabricated AML, despite considerable fabrication uncertainty, still enhances the reflectance to 48% from 40% of the reference superlattice, validating the effectiveness of our ML‐based optimization process.