Hybrid intelligence models for compressive strength prediction of MPC composites and parametric analysis with SHAP algorithm

Nowadays, hybrid soft computing technics are attracting the scholars of construction materials field due to their high adaptability and prediction performances to data information. Hence, the current research aims to predict the compressive strength of magnesium phosphate cement (MPC) composites using the deep learning and machine learning based hybrid models, which is rarely seen in the literature. Data was collected from published papers, where 70% data used for training the models and 30% for testing stage. Four different hybrid models like CNN-LSTM, CNN-GRU, DTR-RFR and GBR-RFR were formulated to achieve the goals by comparing their forecasting performances with statistical parameters. Additionally, governing input variable parameters and prediction process explanation were also interpreted by SHAP algorithm under hybrid models. As is observed, all selected hybrid models presented the good corroboration to output CS data with higher accuracy results. Besides, CNN-LSTM and GBR-RFR models exhibited the superior fitness (R2 ≈ 0.99) to strength properties in relation to other three models at both phases. Average error ranges were observed very condense to ± 5%. Moreover, testing age was observed as the most influential variable to model outputs. Furthermore, it was exposed that CNN-LSTM model can well interpret the interactions of inputs to outputs and inner-working process of prediction, whereas GBR-RFR describes the dependence plot at decent level to elucidate the connections among the inputs for model outputs. However, the proposed hybrid approaches of the research might be a potential solution to optimize the mix design of MPC mixtures containing supplementary cementitious materials (SCMs) and well predict the strength characteristics of MPC matrices for real field applications by engineering practitioners.