Prediction of interparticle parameters of MICP-treated calcareous sand based on the CNN-BiLSTM-attention and XGBoost hybrid models

Microbially induced calcium carbonate precipitation (MICP) has attracted significant attention as a sustainable and environmentally friendly soil improvement technique. The discrete element method (DEM) is among the most commonly used numerical approaches for analyzing the mechanical properties of MICP-cemented materials. To address the challenge of rapidly and accurately calibrating microscopic interparticle parameters in DEMs, this study introduces a novel hybrid modeling framework. The framework integrates a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism (CNN-BiLSTM-attention) with extreme gradient boosting (XGBoost). It is designed as an alternative approach for calibrating interparticle parameters in MICP-treated calcareous sand. The framework was first trained, and the microscopic parameters were predicted using 207 sets of unconfined compressive strength (UCS) test data obtained from DEM simulations. Compared with conventional CNNs, the proposed hybrid model reduced the mean absolute error ( MAE ) by 25.5% and the root mean square error ( RMSE ) by 61.1%. Finally, the predicted microparameters were applied in the DEM calculations, and the results were compared with experimental results. The stress–strain curves and failure morphology obtained from both the experiments and the DEM simulations strongly agreed. These findings demonstrate that the proposed method enables fast and accurate determination of interparticle microscopic parameters for DEM simulations. This approach provides strong support for the application of MICP-treated calcareous sand in marine development and construction projects.