Design and Optimization of Cementitious Systems with Machine Learning

Abstract

Machine learning has revolutionized disciplines within materials science that have been able to generate sufficiently large datasets to utilize algorithms based on statistical inference, but for many important classes of materials the datasets remain small. After an introduction to various types of ML regression, Chapter 1 introduces a rapidly growing number of approaches to show how embedding domain knowledge of materials systems are reducing data requirements and allowing broader applications of machine learning. Furthermore, these hybrid approaches improve the interpretability of the predictions, allowing for greater physical insight into the factors that determine material properties. A survey of the modern utilization of machine learning for cementitious systems is discussed.

Chapter 2 discusses a background of cementitious systems along with a Hierarchical Machine Learning (HML) approach for improving their workability. The dispersion of cement paste induced by various hybrid polymers was explored. PEGylation of lignin derivatives has been shown to enhance emulsifying and dispersant activities. Here, the effects of anionic grafts were explored for dispersant activity within Portland cement. Kraft lignin and lignosulfonate are two important forms of purified lignin whose chemistries are characterized by low concentrations of carboxylate and high concentrations of sulfonate groups, respectively. The dispersion of cement paste by these hybrid polymers was compared with the PEGylated lignin analogues as well as a leading cement superplasticizer, poly (carboxylate ether) (PCE). Slump values significantly increased for both the PMAA-grafted lignin compared to the other analogues allowing for significant reductions in cement water content, with PMAA-grafted lignosulfonate approaching performance of the commercial PCE and suggesting that graft chemistry has a strong effect on dispersant function. Adsorption, zeta potential, and intrinsic viscosity were measured for the lignopolymer analogues to explore the interplay between lignin and graft chemistries in the mechanism of cement dispersion.

Blending metakaolin (MK), a calcined clay, into portland cement (PC) improves resulting concrete material properties, ranging from strength to durability, as well as reduces embodied CO₂ and energy. However, superplasticizers developed for PC can be inefficient or ineffective for improving the dispersion of PC-MK blends. Chapter 3 introduces a novel machine algorithm which was applied to tailor a superplasticizer to address poor flowability characteristic of 85/15 blends of MK-PC. A HML system was trained on a library of seven superplasticizers using a middle layer, which represents underlying physical interactions that determine system responses, based on polymer contributions to physicochemical forces in both the pore solution and particle surface. Synthesis of the algorithm prediction resulted in a water-soluble polymer with a high intrinsic viscosity and a resultant slump value in a cementitious paste that was comparable with leading poly(carboxylate ether) (PCE) superplasticizers. The results from this study demonstrate the importance of HML as a design tool for the molecular engineering of complex material systems.

Chapter 4 introduces alternative binder chemistries (ABC’s) in the form of calcium sulfoaluminate (CSA) cements, which have lower embodied CO₂ compared to portland cement but set rapidly, often within 15 minutes, thus limiting their application. As such, set-retarding admixtures are added to increase the length of time before setting is achieved. These admixtures are typically small organic compounds with high anionic functionality. Retardation is achieved through a complex interplay of mechanisms which involve adsorption onto calcium in the clinker and subsequent prevention of clinker dissolution, complexation with dissociated calcium in the pore solution, and adsorption onto nucleated cement hydration products inhibiting further growth. A cheminformatics based machine learning methodology for the prediction and virtual screening of set retarders for these alternative binder chemistries. Discovery of such compounds is typically achieved through extensive iterative testing that does not ensure optimal solutions. Here, the use of cheminformatics, a data-driven approach used extensively in drug discovery, is demonstrated to identify new set retarders from small datasets for calcium sulfoaluminate (CSA) cements. Based on a sparse training set of 23 molecules containing polar and anionic functional groups, the cheminformatics approach was used to develop a predictive model relating chemical structure to the retarding capability. Then structures of 500,000 compounds were downloaded from a public database, and 365 were predicted to extend set time beyond 1 h. Among these, glyphosate is a commodity chemical that was found to impart a set time of 55 minutes. This cheminformatics approach could be used to develop structure-function relationships and perform rapid virtual screening of chemical admixtures to identify novel high-performance chemical admixtures.

Despite the growing body of work relating to the development of ultra-high performance concrete (UHPC) mixes, the process of designing a UHPC mix is still a highly iterative process. By aggregating previous work on UHPC, Chapter 5 introduces a machine learning model to predict and optimize mix designs based on materials not utilized in the training set. Cement blends are represented in terms of the latent variables of particle packing, water film thickness, and equivalent cement content in order to create a generalizable model. Two rounds of training and testing of were performed utilizing an uncertainty ensemble with a ridge regression while error analysis took place through comparing the Mean Squared Error, a prediction score which ignores Bayesian probability and compares how well the mean values of the data fit to the best model; and Miscalibration area, a quantification of uncertainty in the model based on calibration techniques. The RMSE for the first iteration was 25 MPa and six blends predicted to obtain UHPC strength were tested. Two of the six blends were found to exceed 100 MPa in compressive strength, while the other four needed major blend modification in order to produce a blend of workable consistency. In the second round, three blends were selected with more tightly bound constraints to ensure mixture workability. All three blends exceeded 100 MPa, although model improvement and more informed feature selection is needed, it was shown that through the incorporation of latent variables, a generalizable model could be obtained to predict novel UHPC compositions with a disparate source of materials.

Designing Limestone Calcined Clay Cements (LC3) is a challenge because the factors that are correlated with strength are anticorrelated with workability. Chapter 6 presents a ML methodology for designing LC3 compositions for materials commonly found in North America subject to CO₂ constraints. A hierarchical machine learning approach is performed to represent cement composition as a latent middle layer which can encode any arbitrary composition from a bottom, compositional, layer. Cement blends are represented in terms of their particle packing and water film thickness in both the prediction of workability and strength, while various parameters encoding particle-particle and pore solution forces for superplasticizers in the workability model. A random forest model was utilized in the prediction of workability returning an R²=0.93 on the training set and R²=0.81 on the test set. A gaussian process regression was utilized in the prediction of strength providing a final model training score with an R²=1.0 and showing high generalizability to a test set with an R²=0.97. Analysis of the effect of changes in the compositional variables was visualized through the gaussian process regression giving a posterior probability distribution for all predictions. Finally, these models were combined with a linear model capturing the CO₂ release for every compositional variable. A genetic algorithm was performed in order to predict a Pareto front corresponding to the points of maximum strength and workability, with minimized CO₂, predicting novel blends containing various ratios in combining different sizes of the supplementary cementitious materials.

Finally, Chapter 7 presents conclusions and future directions.