Emissivity assessment of sandblasted surfaces: Methodology for decoupling roughness and surface chemistry effects using Physics-guided Neural Networks

This study presents a physics-guided neural network model for predicting the spectral emissivity of sandblasted titanium surfaces. The proposed approach integrates Agababov’s model with a multi-parametric description of surface roughness and accounts for chemical contamination induced by corundum (Al 2 O 3 ) particle embedding. Fifteen Ti-6Al-4V samples were prepared using five abrasive grain sizes and three projection pressures, producing surfaces with arithmetic mean roughness ranging from 0.53 to 3.72 μ m. Emissivity measurements were performed in the spectral range of 2–23 μ m with a spectrometer. The proposed neural network architecture consists of two parallel branches. The first branch estimates the roughness factor using eight surface descriptors. The second branch predicts the evolution function of the material based on the parameters of the sandblasting process. A Fourier Amplitude Sensitivity Testing (FAST) analysis was employed to identify the most influential roughness parameters, revealing that the arithmetic mean peak curvature (Spc) and the developed interfacial area ratio (Sdr) exhibit the highest sensitivity indices. The model achieves an average prediction error below 2.5 ± 1.37 % on all tested surfaces, outperforming existing analytical approaches that rely on a limited number of roughness parameters and do not account for chemical contamination effects. Analysis reveals that corundum contamination contributes 35–50% of the increase in total emissivity depending on the sandblasting conditions. This finding emphasizes the necessity of accounting for chemical modifications in radiative property models for functionalized surfaces.