Modeling for Thermal Data Correction in Emissivity-Variant Materials: Numerical and Machine Learning Approaches

Infrared thermography is a well-established non-contact technique for surface temperature monitoring across diverse manufacturing and materials processing applications, such as thermal process monitoring, quality inspection, and high-temperature material evaluation; however, its measurement accuracy is affected by emissivity variations in multi-material environments, which distort the relationship between apparent radiometric temperature and true surface temperature. This study presents a sequential thermal correction framework to improve surface temperature estimation in heterogeneous material systems under normal viewing conditions (90° incidence angle). The methodology operates in two stages: first, apparent emissivity is predicted from infrared temperature and material-specific features using regression-based models; second, the estimated emissivity is incorporated into the temperature-correction model to estimate the reference surface temperature, validated against K-type thermocouple measurements. Controlled experiments were conducted on A36 structural steel and alumina ceramic specimens, representing multi-material combinations found in manufacturing environments, over approximately 20 °C to 700 °C. Synchronized thermal imaging and thermocouple data were acquired throughout heating. Polynomial regression and machine learning approaches were evaluated using leave-one-session-out validation. The framework achieved an out-of-session mean absolute error of 9.45 °C, supporting reliable thermal monitoring in manufacturing applications.