Machine Learning-Enhanced In-Situ Combustion: Correcting Carbonate-Derived CO2 Emissions and Evaluating Mineral-Based Carbon Capture

Abstract In-situ combustion (ISC) is a well-established enhanced oil recovery (EOR) technique, traditionally used for high-viscosity oil reservoirs. However, its application to low-viscosity oil reservoirs presents challenges due to the high ignition temperature required and insufficient heavy oil fractions to sustain combustion. This study investigates ISC feasibility in a U.S. reservoir by conducting three combustion experiments—two wet and one dry—using rock samples from three different wells to capture reservoir heterogeneities in combustion performance. X-ray diffraction (XRD) analysis showed that around 45 wt% of the reservoir rock consists of dolomite and calcite, leading to significant carbonate decomposition at elevated ISC temperatures, which increases CO2 emissions and complicates traditional ISC stoichiometric models. Recognizing that existing analytical models fail to account for CO2 from rock decomposition, we developed an integrated approach combining combustion tube experiments with thermogravimetric analysis (TGA) and machine learning (ML). A total of 5,748 TGA data points were collected for calcite and dolomite decomposition at multiple heating rates. Two ML models—Random Forest and XGBoost—were trained (80% of data) and tested (20%) to predict CO2 release from carbonates during ISC. Additional TGA experiments on pure calcite and dolomite confirmed that carbonate decomposition begins as low as 350–400°C, significantly earlier than the commonly reported 550–600°C, especially at higher heating rates (15–20°C/min) under air injection. XGBoost outperformed Random Forest, achieving a higher test R² of 0.986 and a lower RMSE of 0.5044 compared to R² of 0.9846 and RMSE of 0.5289 for Random Forest. Applying ML-based corrections to ISC models significantly improved the accuracy of combustion parameter calculations. The updated model corrected the main combustion reactions, accurately estimating energy generated by crude oil burning (~15,000 BTU/lbm) and energy consumed by carbonate decomposition (~3,000 BTU/lbm). Since CO2 emissions limit ISC applicability, we also explored real-time carbon capture by installing natural mineral filters composed of olivine, dolomite, and ultramafic rock at the combustion tube outlet. These filters were exposed to flue gases for 6.7 hours, achieving measurable carbon uptake of 2.39–7.81 mg carbon per 100 mg sample. Despite the absence of XRD peaks, our results strongly indicate the formation of amorphous MgCO3 or a related carbonate phase on the filter surfaces, as supported by Energy Dispersive X-ray Spectroscopy (EDS), Thermogravimetric Analysis (TGA), and Fourier Transform Infrared (FTIR) spectroscopy results. It underscores the role of mineral-based filtration in ISC carbon mitigation, relying on surface interactions rather than bulk mineral carbonation. These findings highlight the significant impact of rock-derived CO2 on combustion stoichiometry and demonstrate the potential of mineral-based carbon capture for reducing ISC emissions. Implementing zero-emission ISC strategies could enhance the feasibility of this technology for carbonate-rich light oil reservoirs.