Session: 01-03: AI for Energy Sustainability III

Paper Number: 142244

142244 - Comparison of Different Inverse Models to Predict Optical Refractive Indices From Diffuse Reflectance Measurements at High Temperatures 

Abstract: 

The optical properties of materials are essential for real life applications including concentrated solar power (CSP) plants, radiative cooling, and additive manufacturing applications. Diffuse optical spectroscopy has been widely used to noninvasively characterize reflectance for particulate media like a powder bed. However, the retrieval of the optical properties of a target with the inversion of a measured reflectance spectrum is challenging. Using forward models such as diffusion theory or Monte Carlo simulations to iteratively optimize the solution is the existing way, but the nonlinear and multidimensional dependencies (spectral and temperature dependent optical properties and morphological parameters including particle size parameter, solid volume fraction, and particle size distributions) make this process burdensome and computationally expensive. Therefore, the reliable temperature-dependent optical properties are still limited, especially for high temperatures up to 1000K. Herein, we consider the different inverse models of determining temperature-dependent and spectral optical material properties, specifically the refractive index, based on the measured diffuse reflectance for a packed bed with particulate media. Firstly, data-driven forward and inverse models have been developed by applying a decision tree based learning algorithm due to the benefits of explainability and interpretability. Training datasets are initially obtained from physics-based simulations including Mie theory and path-length Monte Carlo ray tracing, based on which large datasets can be generated. Room temperature, constant particle size, varying wavelengths from 0.2 and 14 micrometers and ten common materials including three big categories (ceramic, metal and salt) are assumed. The inverse model predicts the spectral refractive index as a function of wavelength and the measured diffuse reflectance of sample materials. We provide the constraint to separate the materials with zero or nonzero imaginary refractive indices. Preliminary results show good performance of the inverse spectral reflectance prediction from decision tree models compared to path-length ray tracing simulations with an average root mean squared error (RMSE) of 1.42 and 3.42 for materials with nonzero imaginary part and 0.04 and 0 for materials with zero imaginary part. The corresponding R² values are all above 0.98. Secondly, diffuse reflectance measurements data at the range from room temperature to 1000K for alumina, silica and ACCUAST powders (good particle candidates for CSP) are used for training and validation data sets for inverse prediction models at high temperature. The predicted optical properties at room temperature are validated between the training datasets from simulations, experiments and also compared with the well-known material datasets. After that, the optical properties are predicted at selected high temperatures and then the temperature-dependent trends are obtained and connected. Thirdly, we also compare with other inverse models like neural networks in terms of accuracy, global constraint and dataset needed. Overall, these inverse prediction models are computationally efficient and output results within minutes compared to computationally-heavy ray tracing calculations. Additionally, these inverse predictions can offer a sensible view of determining the optimal morphology and optical design of participating media given target performance metrics and material property constraints.

Presenting Author: Zijie Chen University of Michigan-Ann Arbor

Presenting Author Biography: Zijie is a fourth year PhD student working on reduced-order modeling for radiative transport .

Authors:

Zijie Chen University of Michigan-Ann Arbor
Bingjia Li University of Michigan-Ann Arbor
Rohini Bala Chandran University of Michigan- Ann Arbor