Soils contain large amounts of carbon stored as organic carbon and carbonates. These carbon pools can contribute to climate regulation and are of primary importance in ensuring proper soil functioning. However, their accurate quantification remains a complex task. Rock-Eval ® thermal analysis has emerged as an alternative to classic dry combustion and wet methods due to its ability to simultaneously provide organic and inorganic carbon measurements on the same subsample. However, it has been observed that Rock-Eval ® systematically underestimates the soil organic carbon (SOC) while overestimating the soil inorganic carbon (SIC). In this technical note, we propose a validated correction of both SOC and SIC based on a machine-learning model and using a diverse dataset of 240 soil samples. We show that the proposed correction significantly increases the accuracy of the Rock-Eval ® method when compared to reference SOC and SIC values and applied to the dataset used for training and testing and that it can be successfully applied to data originating from different Rock-Eval ® machines without changing the routine analytical protocol. The transferability of the model allows for its future implementation in the Geoworks software so that Rock-Eval ® machines can routinely provide accurate SIC and SOC measurements.