Knowledge-Guided Optimization of Thermally Stable Cement Formulations Using Machine Learning Surrogates for Geothermal and CO2 Storage Wells

Summary Ensuring the long-term mechanical integrity of wellbore cement systems under extreme thermal and chemical environments is essential for the safe and reliable operation of geothermal wells and carbon capture and storage (CCS) projects. Conventional Portland cement systems frequently suffer from strength retrogression, debonding, and durability loss when exposed to high temperatures, aggressive brines, and cyclic thermal loads typical of these applications. With this study, we present an integrated machine learning (ML)-assisted optimization framework for the data-driven design of thermally stable cement formulations tailored to geothermal and CCS environments. A comprehensive data set of more than 360 laboratory cement tests was compiled, encompassing compressive strength, elastic properties, curing conditions, and detailed compositional information for a wide range of additives and cement classes. Five ensemble learning models—random forest, gradient boosting, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost)—were trained as surrogate predictors of compressive strength and temperature suitability. CatBoostRegressor demonstrated the best generalization performance, achieving a test coefficient of determination (R2) of 0.9873 and root mean square error (RMSE) of 9.92°C, with robustness confirmed through cross-validation and formulation-based grouped evaluation. A constrained nonlinear optimization scheme was then developed to generate candidate cement recipes using these surrogate models. The optimization explicitly incorporates operational and compositional constraints, including minimum cement content, cement dominance, additive limits, and a minimum predicted compressive strength of 4,000 psi. Domain knowledge is integrated at the optimization stage through a transparent material-property dictionary that guides the search toward thermally robust and mechanically feasible blends. An ablation study demonstrates that this knowledge-guided objective improves feasibility and interpretability relative to purely data-driven optimization. The resulting optimized formulations, particularly those containing silica flour, metakaolin, slag, and selected nanomaterials, achieve predicted compressive strengths above 7,000 psi and temperature suitability exceeding 300°C. The proposed framework provides a reproducible and computationally efficient approach for rapid screening and design of high-performance cement systems, reducing reliance on empirical trial-and-error testing. The methodology is readily extensible to additional performance metrics and offers a practical decision-support tool for improving well integrity in demanding geothermal and carbon dioxide (CO2) storage applications.