LFMP: physics embedded neural network for light-field multispectral thermography

Light-field multispectral radiation thermometry has emerged as a promising non-contact technique for two-dimensional surface temperature measurement. However, its performance is still limited by temperature inversion algorithms. In this work, we propose LFMP (light-field multispectral physics-embedded network), a physics-informed neural network framework designed for temperature inversion in light-field multispectral thermography. The framework explicitly incorporates Planck’s law and a reference temperature model into its architecture, thereby enforcing physical consistency and enhancing interpretability. The framework enables high-accuracy, spatially resolved reconstruction of two-dimensional temperature fields without requiring explicit emissivity modeling. Blackbody calibration experiments conducted over the temperature range of 573 K to 823 K demonstrate high accuracy, with absolute errors below 5 K and relative errors of less than 1%. In blade film cooling experiments, LFMP maintained robust performance, yielding absolute errors generally below 10 K and relative errors of less than 2% compared to thermocouple measurements. Notably, under reduced coolant flow rates (≤11 g∕s), the relative error further decreased to below 1%, with absolute errors remaining under 8 K. Compared with conventional optimization-based methods, LFMP demonstrates improved temperature accuracy and smoother spatial distributions in the tested cases, highlighting its potential for thermal diagnostics in engineering applications. © 2025 Chinese Laser Press