Knowledge-Guided Machine Learning for Optimizing Thermally Stable Cement Formulations in CO2 Storage and Geothermal Wells

Abstract Ensuring the mechanical integrity of cement systems under extreme thermal and chemical environments is critical for the long-term success of geothermal wells and CO2 storage operations. Conventional Portland cement often fails to maintain durability and zonal isolation when exposed to high temperatures, aggressive brines, and cyclic thermal loads typical of these settings. This study presents a knowledge-guided machine learning (ML) framework to optimize cement formulations specifically for high-temperature geothermal and carbon capture and storage (CCS) applications. A dataset of over 360 laboratory cement tests were compiled, covering compressive strength, Young's modulus, Poisson's ratio, test temperature, and a wide range of compositions and additives. Five ensemble learning models, Random Forest, Gradient Boosting, AdaBoost, XGBoost, and CatBoost, were trained to predict cement strength and thermal response. Among them, CatBoostRegressor achieved the highest accuracy, with an R2 of 0.9873 and RMSE of 9.92 °C. A constrained nonlinear optimization algorithm was then applied to identify cement blends that satisfy operational constraints, including a minimum compressive strength of 4000 psi and enhanced thermal resistance. The optimized formulations, particularly those incorporating high percentages of silica, graphene, metakaolin, and slag, achieved compressive strengths exceeding 7000 psi and predicted thermal stability above 300 °C. The proposed framework offers a robust, data-driven approach for rapidly designing high-performance cement systems, reducing reliance on trial-and-error testing, and improving well integrity in challenging subsurface energy environments.