Demonstration of designable emission control: from numerical simulation to machine learning

Abstract

Controlling thermal emission in both spectral and angular domains is essential for enabling high-performance systems in thermophotovoltaics, radiative cooling, infrared detection, and thermal camouflage. This work presents a comprehensive computational and machine-learning-based framework for the forward and inverse design of thermal emitters using multilayer epsilon-near-zero (ENZ) materials, particularly gradient Si-doped InAs structures. By integrating physics-based simulations with deep learning, the approach enables efficient exploration and optimization of complex photonic structures that were previously inaccessible through conventional methods. ☐ Two major categories of thermal emission behavior are investigated: omnidirectional and directional. For omnidirectional thermal absorption, gradient ENZ structures are designed and experimentally demonstrated to achieve polarization-independent high absorptivity across a wide angular range with tunable spectral control for both TE and TM polarizations. For directional thermal emission, optimized structures exhibit broadband, high-emissivity peaks in targeted angular windows, covering spectral ranges from 7–10 μm and 10–15 μm, enabled by careful doping and thickness control across layers. These designs are validated using Fourier Transform Infrared (FTIR) spectroscopy and shown to match numerical predictions. ☐ To accelerate design and enable inverse retrieval, Machine Learning (ML) models are developed. A 1D forward model capable of accurately predicting mid-infrared spectral responses at a fixed incident angle from structural parameters is demonstrated, achieving near-perfect agreement with TMM simulations while accelerating prediction speed by over 500 times. Beyond forward prediction, this work establishes a powerful inverse design framework, enabling one-to-many retrieval of structural configurations from desired spectral responses. This allows for the identification of multiple viable designs that meet optical targets, from which the most fabrication-compatible or performance-optimal structure can be selected. ☐ This work is further extended to a 2D ML model capable of predicting full-angle optical spectra. Despite the significantly higher output dimensionality, the model maintains high accuracy and generalizability, while offering over 30,000 times speedup over conventional simulations. An accompanying 2D inverse model is also developed, enabling the extraction of optimal design candidates for full-angle target spectra with high fidelity and efficiency. This inverse design capability significantly expands the solution space and offers new avenues for intelligent photonic structure design. ☐ Overall, this thesis establishes a unified framework that integrates physical simulations, nanofabrication, optical characterization, and machine learning to design high-performance thermal emitters. The methodologies demonstrated here are not limited to mid-infrared ENZ systems but can be generalized to other spectral regimes and materials systems. Ultimately, this work establishes a scalable platform for data-driven inverse design, integrating physics-based simulation with modern machine learning to accelerate innovation in materials science.