Machine learning-driven modeling and interpretative analysis of drying shrinkage behavior in magnesium silicate hydrate cement

Abstract

Magnesium silicate hydrate cement (MSHC), as a novel eco-friendly construction material, exhibits remarkable low-carbon advantages. However, its engineering applications are significantly constrained by poor volumetric stability, particularly manifested in pronounced drying shrinkage behavior. The drying shrinkage of MSHC are influenced by multiple complex factors, presenting challenges for traditional methods in rapid and accurate assessment and prediction. Machine learning (ML) demonstrates superior capabilities in processing high-dimensional nonlinear data, offering an efficient solution for material performance prediction. This study aimed to develop ML-based predictive models for drying shrinkage of MSHC and investigate the influence mechanisms of key parameters. Results revealed that the extreme gradient boosting (XGB) achieved optimal generalization capability, with an R² value of 0.963 on the test set. Relative humidity (RH) and Age were identified as critical factors affecting drying shrinkage. Notably, the maximum drying shrinkage occurred at 50 % RH. Further analysis demonstrated that optimizing material composition and curing conditions could significantly enhance shrinkage resistance: increasing sand-to-binder ratio (S/C) and magnesium-to-silicate ratio (M/S), reducing water-to-cement ratio (W/C) and MgO reactivity (A_MgO), incorporating dipotassium hydrogen phosphate (DKP) as superplasticizer, and maintaining pre-curing duration (PCT) around 3 days.