Prediction models for fly ash/slag 28-day activity index: Development and performance comparison

To address the long testing cycle and high cost associated with traditional admixture activity index testing, and realize rapid evaluation of admixture activity for high-strength concrete, the 28-day activity indices of fly ash and slag are targeted as prediction objectives in this paper. Two machine learning prediction models, random forest (RF) and artificial neural network (ANN), were utilized separately, with 14 input features were selected. SHAP value analysis was used to quantitatively interpret the influence mechanism of core features. Results indicate that R 2 values of RF and ANN models for fly ash were 0.944 and 0.920, respectively, while those for slag were 0.942 and 0.879. Specific surface area and 30 μm residue were identified as core influencing factors for the two admixtures. Additionally, the derived feature exhibited a negative correlation with the 28-day activity index of both two admixtures. RF model was more suitable for fly ash, with a coefficient variation of predicted values of 0.7599. ANN model demonstrated better prediction stability for slag, with a standard deviation of predicted values of 0.9009, a coefficient variation of 0.7400, and an interquartile range of 1.14. Using a specific surface area of fly ash below 911.2 m 2 /kg and a 30 μm residue of slag below 8.55% as parameter thresholds could effectively achieve targeted regulation of the 28-day activity index of admixtures. An efficient machine learning method and theoretical basis are provided by this paper for the rapid prediction of admixture activity and the engineering optimization design of high-strength concrete.