Xgboost-based prediction of compressive strength loss ratio in graphene nanoplatelet-reinforced concrete subjected to freeze–thaw cycles

The freeze–thaw (F–T) durability of concrete is a decisive performance criterion for infrastructure in cold climates. Repeated freezing and thawing of pore water gradually damages the microstructure, causes mass loss, and undermines load-bearing capacity. Among the nano-engineered remedies proposed in recent years, graphene nanoplatelets (GNPs) have attracted particular interest. Their two-dimensional barrier morphology restricts water ingress, refines the pore network, and densifies the hydration matrix. The relationship between GNP dosage, F–T exposure, and the resulting compressive strength loss ratio (Df), however, is markedly nonlinear; an optimum dosage window exists, beyond which agglomeration begins to undermine performance. Conventional empirical models struggle to describe this behavior, which makes a data-driven approach attractive. This study applies an Extreme Gradient Boosting (XGBoost) regression framework to predict Df in GNP-modified concrete under rapid F–T cycling. The dataset draws on the experimental program of Chen et al. (2019), spanning seven GNP dosages (0–0.4 wt% of cement) and eight F–T exposure intervals (0–200 cycles) per Chinese standard GB/T 50082-2009. Seven physically meaningful descriptors were used as inputs, among them GNP content, F–T cycle count, water-to-cement ratio, 28-day compressive strength, slump, a GNP × cycles interaction term, and the absolute deviation from the optimum dosage. The model was trained with 400 boosting rounds, a learning rate of 0.05, and a maximum tree depth of 4. The model achieved R² = 0.985 and RMSE = 0.89% on the test set, with a five-fold cross-validated R² of 0.852 ± 0.145. Slump, cycle count, and GNP content jointly explained over 89% of the model's predictive behavior. The partial dependence analysis pointed to an optimum near 0.05 wt% GNP, a dosage at which Df reached its lowest values regardless of cycle count. At the 200-cycle mark, predicted Df dropped from 74.8% in plain concrete to 20.5% in the optimum mix, a reduction of roughly 73% in strength loss.