In this study, we employed laser-induced breakdown spectroscopy (LIBS) along with machine learning algorithms, which encompass partial least squares regression (PLSR), the deep convolutional neural network (CNN), the deep residual neural network (ResNet), and the deep residual shrinkage neural network (DRSN), to estimate the surface hardness of laser cladding layers. (The layers were produced using Fe316L, FeCrNiCu, Ni25, FeCrNiB, and Fe313 powders, with 45 steel and Q235 serving as substrates.) The research findings indicate that both linear and nonlinear models can effectively fit the relationship between LIBS spectra and surface hardness. Particularly, the model derived from the ResNet exhibits superior performance with an R 2 value as high as 0.9967. We hypothesize that the inclusion of numerous noises in the LIBS spectra contributes to the enhanced predictive capability for surface hardness, thereby leading to the superior performance of the ResNet compared to the DRSN.