MULTISCALE SIMULATION AND MACHINE LEARNING-ASSISTED PERFORMANCE PREDICTION FOR CEMENTITIOUS COMPOSITES

Gideon Arthur Lyngdoh, University of Rhode Island

Abstract

Concrete is the most widely used construction material in the world for building infrastructures, bridges, and tunnels. However, cement production contributes to 5-7% of the global CO_2 emissions, and thus, there is a significant need for innovative sustainable alternatives without compromising on the mechanical properties. Most of the approaches to address these concerns so far rely on the trial and error-based experimental response. Hence, robust multiscale simulation-based design approaches are needed to be developed for fundamental understanding as well as the comprehensive design of these new sustainable innovative materials. However, performing the simulations with millions of degrees of freedom could be daunting and requires high computation time. Thus, building data-driven models by training through the experimental/simulated data can serve as an efficient alternative to stimulate the design and development of new materials compositions for desired performance requirements. To address the poor sustainability credential-related issue, this thesis first evaluates the performance of geopolymers which are synthesized through alkaline activation of aluminosilicates. Analogous to calcium silicate hydrates (C-S-H) gel in cement paste, sodium aluminosilicate hydrate (N-A-S-H) gel is the primary binding phase in geopolymers formed via alkaline activation of fly ash. In this thesis, a realistic molecular structure of N-A-S-H geopolymer gels, inspired by the traditional calcium silicate hydrates gel, is proposed using molecular dynamics (MD) simulation. In contrast to the existing N-A-S-H model—where water is uniformly distributed in the structure— a layered-but-disordered structure is presented where water molecules are incorporated in the aluminosilicate network's interlayer space. The developed structures are further validated using experimental observations. To address the durability performance of geopolymers, the dynamics of confined water and its interplay with alkali cations in disordered N-A-S-H gel are evaluated using reactive force field molecular dynamics. This is achieved by exploiting the evolution of mean squared displacements and the Van Hove correlation function. It is observed that the Si/Al ratio significantly influences the diffusion of confined water and sodium. Increased conversion of the Si–O–Na network to Si–O–H and Na–OH components with an increase in water content helps explain the alkali-leaching issue in fly ash-based geopolymers observed macroscopically. Moreover, the fracture properties of the disorder N-A-S-H gel are explored via molecular dynamics simulations. The simulated fracture toughness values of N-A-S-H are validated with the experimental results obtained using the nanoindentation technique, where the principle of conservation of energy is implemented to evaluate the fracture toughness from the load-penetration depth responses. Afterward, to address the issue related to the high computational demands of multiscale simulations, this thesis synergistically integrates multiscale simulations, experiments, and machine learning to predict the performance of various multiphysics responses of a wide variety of cementitious materials. Firstly, an ML model is developed using high throughput MD simulation that mapped the elastic properties with its chemical composition in C-S-H. The simulation results reveal that the influence of the silicate network on all the elastic constants of C–S–H is significantly higher than that of water and CaO content. Secondly, the strain-sensing ability of a nanoengineered self-sensing cementitious composite is predicted by synergistically integrating a validated FE analysis-based multiscale simulation framework with ML. The developed model predicts the strain-sensing response efficiently. Next, an ML-based model was developed for 3D printed cementitious auxetic cellular composites. With the advent of 3D printing, auxetic cellular cementitious composites (ACCCs) have recently garnered significant attention owing to their unique mechanical performance. Here, the prediction of Poisson's ratio using ML approaches is developed by synergistically integrating an FE analysis-based framework with ML. Using SHAP, it is established that the volume fraction of voids is the most influential parameter in inducing auxetic behavior. In the end, an efficient ML model was developed to evaluate the non-linear composition-strength relationship in traditional concrete. Here, the ML model is trained using the experimental data available from the literature. However, the adopted dataset suffers from incompleteness because of missing data corresponding to different input features. To address the incompleteness in the dataset, different data imputation approaches are implemented for enhanced dataset completeness. The imputed dataset was leveraged to develop a complete ML model for concrete strength predictions. Besides, SHAP was implemented to evaluate the relative sensitivity of various performance descriptors, which can help toward the development of various high-performance concretes with unexplored compositional domains. Overall, validated performance prediction tools and various fundamental insights presented in this thesis help forward viable strategies toward the design and development of durable, resilient, and yet sustainable construction materials for next-generation civil infrastructure.