This paper discusses the estimation of adsorption energies for reaction intermediates for a given metallic surface and molecule. Regression models are learned from DFT data available in the literature in a two step approach. First, metallic surfaces are characterized by a principal component analysis (PCA) followed by a suitable orthonormal rotation to find a set of species that can be used as descriptors for the metallic surface. Then, different machine learning techniques are considered for the regression using the previous descriptors for the metallic surface and molecular descriptors such as the number and type of bonds for the adsorbate. With the available data, CH3, CO2 and CH2 were found to explain 93% of the total variance, thus were used as surface descriptors. Threeof the tested models were found to adjust similarly well to validation data.