The lack of a reliable method to initialize the size of the different SOC kinetic pools is recognized as a clear limitation for the accuracy of SOC stocks evolution predictions provided by these models. There were a few attempts to use results derived from soil fractionation schemes to initialize the size of the stable pool of models but the proposed methods were not validated and are time consuming. Using an exceptional soil sample set from long-term bare fallow experiments in Europe, we showed that a machine learning algorithm based on Rock-Eval 6 thermal analysis data can accurately predict the proportion of the centennially persistent SOC pool in a soil sample for a wide range of temperate soils (Cécillon et al., 2018). We hypothesized that this quick and cheap method could significantly improve SOC dynamics models initialization. We tested this hypothesis on a soil sample set from French long-term agronomic experiments (AIAL). Our results showed that our Rock-Eval-based method is suitable to initialize the AMG model in French cropped soils and that it improves the accuracy of AMG simulations.