Comparison of U-Net-based mapping from DIC-derived effective strain to infrared thermography data and from infrared thermography data to DIC-derived effective strain

Digital Image Correlation (DIC) and Infrared Thermography (IRT) are complementary full-field techniques used in experimental mechanics for analysing deformation and thermal response during loading. Previous work demonstrated that DIC-derived effective strain fields can be used as input for U-Net-based prediction of corresponding infrared thermography fields. In this paper, that mapping direction is compared with the reverse problem, in which infrared thermography data are used to estimate DIC-derived effective strain fields. Both directions were evaluated using the same paired experimental dataset, consisting of strain and thermal fields obtained from simultaneously recorded mechanical tests. The fields were pre-processed, normalised, resized to 256 × 256 pixels, and used for supervised image-to-image regression. The results show that both mapping directions can be learned to a certain extent using the same U-Net-based approach, but they are not equivalent. The DIC-to-IRT mapping reproduced the main thermal field distribution and showed stronger structural agreement with the reference thermographic data. The reverse IRT-to-DIC mapping achieved low average pixel-wise errors, with mean RMSE and MAE values of 0.0188 and 0.0051, respectively, but showed limited structural similarity, with a mean SSIM value of 0.0580. This indicates that thermal data contain useful information about the mechanical response, although detailed strain localisation remains difficult to reconstruct. The comparison confirms that the relationship between mechanical and thermal full-field data is direction-dependent and that thermal-to-strain estimation should be interpreted as an approximate complementary approach rather than a replacement for DIC.