Neural-network-based multi-spectral thermometry and emissivity reconstruction in cavity high-temperature environments.

Abstract

This study presents a neural-network-assisted framework for accurate radiation thermometry in high-temperature cavities, overcoming challenges from unknown emissivity and multi-reflection effects. The method combines Monte Carlo ray-tracing with deep learning through: (1) physics-informed training data including diffuse/specular reflections, (2) alternating neural networks for decoupled temperature/emissivity prediction, and (3) full multi-reflection modeling. Validated with zirconia in a graphite cavity within the 1273-1673 K temperature range and 2-16 µm spectral range, it achieves 0.7% temperature error (9 K, compared to real temperature) and 0.05-0.1 emissivity error in 2-16 µm, outperforming first-order methods (neglecting multiple reflections) by 5%-27% (peak at 2 µm) in emissivity reconstruction. The framework maintains <1% error, with only 10 spectral channels and tolerates 1% intensity noise (<1.8% variation), enabling reliable thermometry in low-emissivity materials like alloys and ceramics where conventional methods may fail.