Machine Learning Approaches for Improving Construction Materials and Pavement Systems

Abstract

The construction industry is faced with unprecedented challenges of population growth, urbanization, and worldwide climatic change. These necessitate the development of optimized construction and pavement materials that are cost-effective, sustainable, resilient, and long-lasting. The conventional experimental design approaches to designing construction materials and assessing pavement performance are time-consuming, expensive, and unable to investigate the complex interactions of the different factors influencing their performance. Therefore, this dissertation research applies machine learning (ML) techniques to predictive modeling and optimization and forms a data-driven strategy for material selection and performance prediction. This dissertation is focused on four primary studies, each showing the application of ML in different applications in construction materials and pavement engineering. The first study uses ML algorithms like Categorical Boosting (CatBoost) to predict the unconfined compressive strength of cement-treated soil for deep mixing. Deep-mixed soil plays a crucial role in reinforcing weak clay soil, which is a common construction problem. The results demonstrated that ML models can enhance the accuracy of forecasts significantly, relative to nonlinear regression, and soil stabilization interventions can be more effective. The second study examined how crumb rubber particle size affects the strength of rubberized concrete. As demand for eco-friendly materials increases for buildings, rubberized concrete is a cost-effective prospect by recycling tires at the end of their life. Uncertainty of the rubber particle size, however, affects the mechanical properties of the concrete. To remedy this, the study used the Synthetic Minority Over-sampling Technique (SMOTE) to over-sample and CatBoost to build a predictive model. The results indicated that particle size distribution optimization will improve concrete strength and could be useful information in the future design of materials for sustainable development, although the amount of crumb rubber was found to have greater impact on the predicted strength than the size of the rubber particles. The model also indicated that a 40 MPa concrete can be attained if the rubber addition is less than 10% and the water-to-cement ratio is less than 0.5. The model validation results showed that its prediction performance was limited within the water-to-cement ratio range of 0.35 to 0.5, with particularly lower accuracy at 0.35 due to insufficient consideration of the curing process effects and low amount of data in that range. The third study employed ML to predict the concrete strength with steel furnace slag (SFS) aggregates. SFS is an industrial by-product of steel production that can be used to enhance the performance of concrete and decrease the amount of waste in the environment. However, the blended nature of SFS, such as basic oxygen furnace (BOF) slag, electric arc furnace (EAF) slag, and ladle metallurgy furnace (LMF) slag, has been established in literature to confer various concrete characteristics. On the basis of synergy between SMOTE and Light Gradient Boosting Machine (LightGBM) K-fold model, the present work could effectively predict compressive strength of concrete with SFS aggregates. The model showed that concrete with SFS aggregates can have statistically significant increases in compressive strength over plain concrete. Feature analysis also showed that the amount of SFS aggregate was more dominant in the prediction than the SFS type. The model was valid up to a 30% replacement of aggregate but indicated that more data and/or development of the model to incorporate additional input variables are required. The last study employed supervised and unsupervised ML models to predict the International Roughness Index (IRI) of concrete pavements from the Long-Term Pavement Performance (LTPP) database. Not unexpectedly, the models concluded that material properties, traffic load, and climate conditions were crucial to IRI prediction. Hierarchical clustering models were constructed to group the data into four clusters, where each cluster contains data with similar traffic volume, climate, pavement age and IRI values, which suggests that, with this limited data at least, pavement performance could be regionally divided for pavement design purposes. The findings conclude that ML models can be used effectively for the prediction of IRI, hence facilitating effective proactive pavement maintenance planning, effective resource planning, and enhanced infrastructure management practices. This dissertation puts into perspective the application of ML towards system and construction material optimization. Through ML-based data analysis, this dissertation improves prediction, decreases dependence on expensive experimental tests, and supports the construction of sustainable infrastructure. The results increase material selection process improvements, lessen the environmental impact, and increase road maintenance methods for more sustainable and less expensive construction practices.