Phosphates adsorption on oxides is significant interest in heterogeneous catalysts and environmental research. The CoMoS/γ-Al2O3 hydrodesulfuration catalyst which is used at the scale of petroleum refinement, uses phosphate additives from the first steps of catalyst preparation. A preparation follows the stages γ-Al2O3 calcination, wet incipient impregnation and drying (as well as activation). Conditions like oxide water interfaces are found in similar systems exist like aluminosilica in rock, sediments and soil where phosphate adsorption determines eutrophization of lakes and oceans. Oxide surfaces like of γ-Al2O3 are very complex, which renders chemical reactions with adsorbates even more intricate. However by today, improved NMR spectroscopic methods complemented with proven computational models allow to mold an atomistic model of phosphate speciation. The presented work starts from the creation of a systematic database by exploring more than 1000 geometries at different adsorption modes, adsorption sites, degrees of coverage and H-bond environment for two phosphate species. The best results by free energy agree well with recent 31P-NMR experiments at drying conditions. Liquid conditions were simulated using state of the art metadynamics methods. Our results show that the best results are kinetically plausible and condensation reactions similar to known enzymatic reactions. Machine learning methods were applied in an attempt to predict 31P-NMR chemical shifts at DFT accuracy but capturing dynamic effects. This work highlights key aspects of how experimental and computational methods can be joint to achieve a better understanding of dynamic and disordered systems.