Integrated hybrid machine learning techniques and multiscale modeling towards evaluating the influence of nano-material on strength of concrete

Nano-material impact on high-performance concrete strength represents one of the major hot research areas in recent years. However, the currently available studies are often limited by small sizes of their datasets, by incomplete multiscale modeling with subsequent suboptimal prediction accuracies and failure to obtain complete understanding of nano-material effects. In this regard, the current study provides an integrated approach combining Generative Adversarial Networks, Finite Element Analysis, Molecular Dynamics, and Long Short-Term Memory. GANs are used to generate synthetic data that enables augmentation of the experimental dataset by 50–100%, which consequently enhances robustness and training capability. The ensemble learning model includes Random Forest and Gradient Boosting Machines trained on this augmented dataset that improves predictions by 15–20% in accuracy while bringing down the Root Mean Squared Error. Multiscale modeling through MD simulation captures nanoscale material interactions and provides the necessary parameters for FEA, which in turn predicts the macroscale structural behavior, and this has seen an increase of 10–15% in compressive strength and 20–25% improvements in model accuracy. LSTM networks are applied for the predictive maintenance, thus forecasting the performance degradation over time while providing SHapley Additive exPlanations (SHAP) of the black-box model explanation. The proposed framework realizes the potential to decrease RMSE in predicted strength degradation by 5–10% and points out that nano-materials are one of the important contributors explaining up to 35% of the variance in strength predictions. It has thus integrated data augmentation with multiscale modeling and explainable AI towards a more accurate and interpretable analysis in respect of the role of nano-materials in enhancing concrete strength and durability.