Abstract: Molecular reinforcement of calcium silicate hydrate (C-S-H) gel through elemental doping offers a new strategy for achieving improvements in the performance of cement-based concrete. Current experimental and simulation approaches remain confined to trial-and-error methodologies, lacking the capacity for systematic optimization across vast elemental doping schemes. This study proposed a machine learning-based high-throughput screening method for elemental doping in C-S-H molecules. The innovative approach leveraged existing phyllosilicate mineral data to predict the properties of C-S-H with varying chemical composition, using molecular formation energy as an indicator to identify highly stable molecular structures. This enables batchwise and efficient screening of multi-element doping configurations, with computational costs below 1 s per sample on an 8-core processor while bypassing complex modeling procedures. The random forest algorithm demonstrates exceptional performance in high-throughput screening of elemental doping cases, achieving an root mean square error (RMSE) of 0.060 and R² of 0.980 on test samples. Prediction results confirm: 1) stability reduction due to charge imbalance, 2) decreased formation energy in highly polymerized molecules, 3) weakened chemical effects from protonation of non-bridging oxygen sites. The study investigates mono- and multi-element doping schemes involving Al, Fe, Ti, Na, Mg, Li, Cr, and H, proposing feasible Na/Al co-doping approaches to enhance C-S-H stability, and validating magnesium ions' kinetic propensity to substitute calcium sites in C-S-H. Grounded in first-principles theoretical frameworks, this method demonstrates potential for extension to inorganic non-metallic mineral systems, enabling high-throughput prediction and screening of numerous unknown materials.

Key words: C-S-H, silicate mineral, machine learning, elemental doping, high-throughput screening