Novel hybrid deep-long short-term memory and machine learning algorithms for the crack-healing estimation of eco-friendly engineered cementitious composites

Abstract

Self-healing concrete addresses cracks caused by environmental factors and tensile strain, helping to extend its lifespan. While Engineered Cementitious Composite (ECC) is particularly important for self-healing, developing a mix that includes supplementary materials instead of relying solely on cement can help reduce environmental impacts. Therefore, accurately predicting its self-healing capacity is crucial. The goal of this study is to develop two new hybrid models that combine Long Short-Term Memory (LSTM) with Gaussian Process Regression (GPR) and Least-Squares Boosting (LSBoost) models, respectively, as well as their individual algorithms to predict the crack-healing ability of eco-friendly ECC. The fly ash, silica fume, limestone powder dosages, and crack width before self-healing (CW-B) were used as input to predict the crack width of ECC after self-healing (CW-A). The ANOVA analysis revealed that CW-B had the most significant impact on CW-A, accounting for 91.84% of the variation. The individual models-LSTM, GPR, and LSBoost-estimated the CW-B with accuracies of 94.6%, 75.4%, and 64.4%, respectively. However, the performance of GPR and LSBoost improved with their hybrid usage alongside LSTM, achieving accuracies of 95.7% and 90.9%, respectively. The LSTM-GPR hybrid model outperformed all others, as evidenced by its narrow range of point-by-point residuals and Taylor plot. Additionally, the LSTM-LSBoost model, despite being slightly less accurate, still demonstrated acceptable predictive capability. These findings indicate that the novel hybrid LSTM-GPR model is the most effective for the self-healing ability prediction of eco-friendly ECC, achieving higher accuracy and lower error rates compared to actual outcomes.