Abstract
This study investigates adhered mass formed on internal concrete-mixer surfaces, a process associated with material losses, increased cleaning demand, and reduced operating efficiency.The aim was to develop and evaluate a machine-learning model for predicting adhered mass during mixing, including mixtures containing man-made mineral waste.The dataset comprised 350 independent mixing-cycle observations from a 0.5 m 3 mixer, divided into a development subset (n = 245) and a locked test subset (n = 105).Nine variable predictors were used; shaft speed was fixed at 60 rpm and treated as an operating condition rather than a predictor.Random Forest, Gradient Boosting, and multilayer perceptron models were compared under the same leakage-free validation protocol.Gradient Boosting achieved the best locked-test performance, with MAE = 0.061 kg, RMSE = 0.078 kg, and R 3 = 0.991.Ablation and SHAP analyses identified the water-cement ratio, drive motor current I, and mixing time as the largest contributors to prediction.Tests in a 0.22 m 3 mixer were used only as a supplementary qualitative cross-mixer check, not as quantitative external validation.The proposed framework supports identification of mixture compositions and operating conditions associated with increased adhered mass.