Date of Award

2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy in Civil and Environmental Engineering

Department

Civil and Environmental Engineering

First Advisor

Sumanta Das

Abstract

The growing demand for sustainable and resilient infrastructure has accelerated the development of advanced cementitious materials and predictive frameworks capable of improving material design, performance evaluation, and infrastructure assessment. Simultaneously, recent advances in numerical simulations and machine learning (ML) have created new opportunities for accelerating materials discovery while reducing experimental and computational costs. This dissertation integrates numerical simulations and interpretable ML to predict the mechanical behavior of sustainable cementitious materials and the thermal response of concrete bridge deck systems across multiple engineering applications. Through the integration of transfer learning (TL), generative adversarial networks (GANs), finite element analysis (FEA), and SHapley Additive exPlanations (SHAP), this research establishes efficient and interpretable frameworks for predicting complex material and structural responses while providing mechanistic insights into the governing factors influencing performance.

Chapter 1 presents a comprehensive review of ML applications in alkali-activated materials (AAMs), including pastes, mortars, and concretes. The review critically examines the evolution of ML techniques from conventional regression models to advanced approaches such as neural networks (NNs), genetic programming, random forests (RF), adaptive boosting, and gradient boosting algorithms. It highlights the ability of ML models to accurately predict fresh-state, mechanical, and durability properties while identifying key challenges related to data quality, dataset limitations, model interpretability, and generalization. The findings demonstrate the considerable potential of data-driven approaches for accelerating the design and optimization of sustainable cementitious materials.

Building upon these foundations, Chapter 2 develops an interpretable ML framework for predicting the compressive strength of fly ash-based alkali-activated concretes (AACs) using data imputation and augmentation strategies. Several imputation methods were employed to address incomplete experimental records, while GAN-based augmentation was utilized to substantially increase dataset size and diversity. RF, Extreme Gradient Boosting (XGBoost), and NN models were subsequently trained and evaluated. The NN model combined with k-nearest-neighbor (kNN) imputation demonstrated superior predictive performance compared with RF and XGBoost models. SHAP analyses identified water content, SiO2 content, and curing conditions as dominant variables governing compressive strength, while SHAP violin and river plots provided further insight into feature contributions and interaction effects. The study demonstrates that integrating data preprocessing, augmentation, and interpretability techniques significantly improves predictive accuracy while providing valuable insights into composition-property relationships within AAC systems.

Chapter 3 further advances data-efficient predictive modeling through the integration of GAN-based augmentation and TL for compressive strength prediction of AACs. An experimental dataset consisting of 188 records was expanded to 4,950 statistically representative samples using a GAN framework. The augmented dataset was subsequently utilized to develop TL-based NN models and benchmarked against conventional NNs trained on augmented data. The TL framework achieved improved predictive accuracy, lower prediction error, and superior generalization while maintaining compact network architectures. SHAP analyses identified curing time as the dominant variable affecting compressive strength, followed by activator composition, precursor content, and aggregate proportions. The results demonstrate that combining TL with generative data augmentation provides an effective strategy for overcoming data scarcity while maintaining model transparency and physically meaningful predictions.

Extending the application of interpretable ML to architected cementitious materials, Chapter 4 presents an integrated FEA-ML framework for predicting and interpreting the auxetic behavior of elliptic arc cementitious composites. A database comprising 900 geometrical configurations was generated by varying primary axis length, aspect ratio, cell thickness, volume fraction, and arc angle. Approximately 20.6% of the geometries exhibited auxetic behavior, with Poisson’s ratios ranging from -0.47 to 0.20. The developed NN achieved excellent predictive performance with testing R² of 0.987, while SHAP analysis identified primary axis length and cell thickness as the dominant parameters governing auxetic response. The proposed framework enables accurate prediction, interpretation, and optimization of geometry-dependent deformation behavior in cementitious composites.

Moving from material-scale applications to infrastructure-scale systems, Chapter 5 develops an interpretable ML framework for rapid prediction of bridge deck surface temperatures under winter climatic conditions. A transient thermal FEA model was used to generate 14,723 bridge deck thermal response cases from historical weather records for five cities across the United States between 2003 and 2024. The developed NN accurately predicted bridge deck surface temperatures with robust performance on independent test data. SHAP analysis confirmed that the model preserved the underlying thermal relationships captured by FEA, identifying ambient air temperature as the dominant factor, followed by solar irradiance and wind speed. The proposed framework enables efficient prediction of bridge deck temperatures and freeze-thaw exposure while substantially reducing the computational effort associated with repeated FEA.

Collectively, this dissertation establishes a comprehensive framework integrating numerical simulations and interpretable ML for predicting the behavior of sustainable and resilient cementitious materials. By combining physics-based modeling with advanced data-driven techniques, including TL, GAN-based data augmentation, and SHAP analysis, the proposed approaches enhance predictive accuracy, interpretability, and computational efficiency across multiple material and structural applications. The findings contribute to the development of next-generation sustainable construction materials and intelligent engineering frameworks that address challenges related to resource efficiency, environmental sustainability, and infrastructure resilience.

Download

Available for download on Thursday, August 31, 2028

Share

COinS