Solitary-wave-based deep learning for compressive strength estimation in cementitious materials

Abstract

This study presents a novel deep learning (DL)-based approach for predicting the compressive strength of cementitious materials using highly nonlinear solitary waves (HNSWs) as input data. The proposed method leverages convolutional neural networks (CNNs) to classify compressive strength of mortar by transforming continuous measurements into discrete categories. Four different formats of HNSW signals are explored to evaluate the impact of signal preprocessing on model performance. Multiple mode testing is implemented to enhance the robustness of predictions, using multiple signals from the same class to reduce variability and stabilize results. The DL models were tested on datasets varying by water-to-cement (w/c) ratios and hydration time, achieving superior performance through signal slicing and frequency-domain transformations. Notably, the model achieved high prediction accuracy, with R² values up to 0.989 and RMSE as low as 0.930 MPa, demonstrating its reliability for predicting compressive strength in cementitious materials. Comparative analyses with benchmark architectures such as AlexNet, GoogleNet, and ResNet-18 highlight the effectiveness of the tailored CNN model, which consistently outperforms these benchmarks, especially in multiple mode testing.