Nano calcium carbonate enhanced clay soils: Machine-learning prediction of unconfined compressive strength

Nano calcium carbonate (NCC) has emerged as a promising soil-stabilization additive owing to its dual role as a chemical conditioner and a microstructural filler in clay soils. However, the highly nonlinear interactions among NCC dosage, ageing duration, and initial soil strength make the prediction of unconfined compressive strength (UCS) enhancement challenging using conventional empirical approaches. This study presents a machine-learning framework for predicting UCS enhancement in NCC-treated clay soils using a literature-derived dataset of 62 samples compiled from seven independent experimental investigations. Three base regressors, namely an artificial neural network (ANN), a boosted regression tree (BRT), and Gaussian process regression (GPR), were evaluated alongside two ensemble strategies, a MAPE-optimized stacking ensemble (STACK_MAPE) and a Mixture-of-Experts model with Top-2 gating (MOE_TOP2). Generalization performance was estimated using repeated 10-fold cross-validation with 20 repeats to generate out-of-fold (OOF) predictions. STACK_MAPE achieved the lowest OOF RMSE (14.56 percentage points) and highest R² (0.9615), with a MAPE of 16.14%, while only marginally improving on the standalone ANN (RMSE = 14.74 pp, R² = 0.9605, MAPE = 16.56%). MOE_TOP2 attained the lowest mean absolute percentage error (MAPE = 14.50%) and highest Spearman correlation coefficient (SCC) of 0.9823 among all five evaluated models, indicating stronger proportional and rank-order accuracy. GPR exhibited the weakest generalization (RMSE of 27.87, R² = 0.8589) with a mean bias of +6.69 percentage points. These findings demonstrate that meta-learning strategies based on adaptive base-learner weighting generally match individual models for UCS enhancement prediction in NCC-treated clay soils, providing a reliable and practical tool for geotechnical engineering design.