Accurate Depth-Resolved Temperature Profiling via Thermal-Radiation Spectroscopy: Numerical Methods vs Machine Learning

TL;DR

This paper compares numerical and machine learning methods for accurately retrieving depth-resolved temperature profiles from thermal-radiation spectra, demonstrating the superior robustness of deep neural networks on synthetic and experimental data.

Contribution

It introduces a deep neural network approach for thermal-radiation spectroscopy, outperforming traditional numerical solvers in accuracy and robustness.

Findings

DNN method outperforms numerical solvers on synthetic data

DNN provides more accurate temperature profiles on experimental spectra

Deep learning offers a robust solution for depth-resolved temperature profiling

Abstract

We present and compare three approaches for accurately retrieving depth-resolved temperature distributions within materials from their thermal-radiation spectra, based on: (1) a nonlinear equation solver implemented in commercial software, (2) a custom-built nonlinear equation solver, and (3) a deep neural network (DNN) model. These methods are first validated using synthetic datasets comprising randomly generated temperature profiles and corresponding noisy thermal-radiation spectra for three different structures: a fused-silica substrate, an indium antimonide substrate, and a thin-film gallium nitride layer on a sapphire substrate. We then assess the performance of each approach using experimental spectra collected from a fused-silica window heated on a temperature-controlled stage. Our results demonstrate that the DNN-based method consistently outperforms conventional numerical techniques on both synthetic and experimental data, providing a robust solution for accurate depth-resolved temperature profiling.