Characterization and Predictive Modeling of Diatomite Mortar Performance: A Hybrid Framework Based on Experimental Analysis and Machine Learning Meta-Models

Decarbonizing the construction sector requires high-volume replacement of Portland clinker with non-calcined supplementary cementitious materials (SCMs). This study investigates white cement pastes incorporating raw Algerian diatomite—a silica-rich biogenic mineral—at substitution levels from 40% to 95% (5% increments) and a fixed water-to-binder ratio of 0.5. The target application is ultra-lightweight, multifunctional composites for non-structural uses such as decorative panels and partition elements. Increasing diatomite content progressively reduced bulk density from 1.483 g/cm3 (D40) to 0.557 g/cm3 (D95) and increased porosity. 28-day compressive strength decreased monotonically from 16 MPa (D40) to 2.4 MPa (D95) as clinker dilution intensified. Ultrasonic pulse velocity dropped from 6205 m/s to 1495 m/s, reflecting progressive pore development and confirming the material’s lightweight potential. Statistically significant strength gains beyond 28 days were recorded (+25.87% for compression, p-value < 0.05), evidencing delayed pozzolanic activity. These results confirm that raw, non-calcined diatomite is a viable SCM for eco-efficient, low-density construction systems. To overcome the extrapolation instability of purely data-driven approaches, a Meta-Avrami Hybrid Framework was developed. It anchors Gradient Boosting residual learning to a sigmoidal Avrami hydration kernel. The model achieved high predictive accuracy (R2≈0.999, RMSE≈0.010) under 10-fold cross-validation. Generalization was well-controlled, with a low overfitting gap (ΔR2=0.0226) and stable fold-to-fold performance (Std=0.0204). These metrics confirm suitability for unseen mix designs. This is particularly relevant for service-life assessment of partition panels and lightweight façade elements, where long-term performance guarantees are required. The physics-informed architecture ensures asymptotic strength stabilization up to a 10-year horizon (amplification ratios 1.03–1.05). This prevents the non-physical divergence observed in polynomial and power-law hybrids (ratios 1.36–1.70). The framework provides a reliable and interpretable tool for service-life design of sustainable low-carbon cementitious systems.