UC Irvine

The ability to engineer how light interacts with a surface is at the core of the development of materials that can passively and selectively control thermal radiation for thermal management solutions. Recent years have seen significant developments in the design of microstructured materials and metamaterial surfaces that are engineered to control matter-light interactions. This doctoral research focuses on the design and optimization of one such microstructure – micropyramids – and how micropyramids can be designed, optimized, and implemented to control thermal radiation using both advanced computational and experimental approaches. Micropyramid texturing consists of nanometer to micrometer scale pyramids with a symmetric base and can be used to modulate the optical and thermal properties of a surface. Micropyramids control optical properties by introducing localized electromagnetic confinement and geometric scattering that reduces a surface’s reflectivity. These mechanisms can be leveraged by controlling the micropyramids’ key geometric parameters and constituent material(s) – and if properly engineered – can induce and optimize anti-reflective behavior at desired broadband wavelength spectrums. While the general anti-reflective properties of micropyramids and other similar surface-relief grating structures are well studied in the field of optics, much less is known about the application of micropyramids to broadband thermal management. To investigate micropyramids for thermal management, an optimization engine derived from cost-function driven thermal analysis and the finite-difference time domain (FDTD) optical solver was developed. Using this process, several common engineering materials – such as Nickel (Ni) and Alumina – were studied and analyzed for tunable optical properties by combining them with micropyramid structures. From the analysis, the key geometric and material parameters that link to thermal-optical properties were identified. A key limitation to the cost-function driven analysis, however, is that the FDTD solver is slow and computationally expensive. To address this, a neural network method derived from deep learning was developed to act as a “surrogate” optical solver. The surrogate network can perform simulations at a rate 6 to 8 orders of magnitude faster than the FDTD simulations it was trained on, allowing for large-scale thermal optimizations of properties that would be impossible using traditional simulation methods. The network’s novel design allows for discrete material inputs, and demonstrates an exceptionally high degree of accuracy in extrapolating the optical properties of materials that the model has not been trained on. The surrogate method is further refined and improved using machine-vision and image-based methods, paving the path towards neural network derived models that can predict the optical properties of complex geometries and multi-material systems without computationally expensive simulations. While the cost-function method for thermal optimization is effective and is greatly accelerated using the surrogate neural network, it is limited in its ability to perform inverse design. Finding optical spectrums to match thermal conditions using the surrogate method requires “brute-force” optimization methods, which limits the optimization speed and scope. To rectify this, a neural network architecture was developed that inverts the problem and directly provides geometric and material solutions that best fit a desired input spectrum. This process occurs nearly instantaneously and facilitates the optimization and design of micropyramids for broadband and narrow-band applications. The inverse neural network was combined with the surrogate network and post-processing/simulation methods to form a self-learning loop that improves thermal and optical prediction accuracy as more inputs are processed. Furthermore, we utilize the network construction to develop a material search algorithm that can both search through existing materials and identify new materials that best solve for a desired input spectrum. From the analysis of micropyramids and deep-learning process, a material and geometry that improves broadband infrared (IR) thermal emission is identified and subsequently fabricated. By combining nanometer (nm)thick coatings of metal with micro and nanotextured silicon, specific control of broadband optical properties is demonstrated that would be impossible to achieve with either material on its own. While the average infrared absorptivity of untextured silicon with 20 nm of Ni is 0.11, it is shown that micro and nanotexturing can increase the average infrared absorption when combined with a 20 nm thick Ni coating to 0.46 and 0.66 respectively. The findings in this thesis will guide the design of future surrogate neural networks for microstructure design and for the optimization of micropyramids for thermal radiation control.