Concrete performance assessment and quality control are described using workability, mechanical, and durability characteristics of concrete. Machine Learning (ML) offers computational and data-driven means for predicting and assessing concrete performance. Using ML in concrete performance prediction has shown superiority over traditional methods (i.e., holistic procedures and standardized methods) because it saves time and cost and provides sustainable construction while achieving reliable and accurate predictions. Previously, most studies used ML on small data (<1000 observations) associated with laboratory-produced concrete. Models developed on laboratory concrete have deficiencies for their implementation in the performance assessment of industrially produced or field-placed concrete since they do not consider the uncertain and varying conditions in the field. This research provides another attempt to predict and assess industrial concrete performance in addition to a few previous studies on industrial concrete modeling.

This study comprehended the use of ML in concrete performance assessment. It used eleven well-known and conventional ML methods to predict three key performance characteristics of industrial concrete: (1) 7-day compressive stre...[ Read more ]

Concrete performance assessment and quality control are described using workability, mechanical, and durability characteristics of concrete. Machine Learning (ML) offers computational and data-driven means for predicting and assessing concrete performance. Using ML in concrete performance prediction has shown superiority over traditional methods (i.e., holistic procedures and standardized methods) because it saves time and cost and provides sustainable construction while achieving reliable and accurate predictions. Previously, most studies used ML on small data (<1000 observations) associated with laboratory-produced concrete. Models developed on laboratory concrete have deficiencies for their implementation in the performance assessment of industrially produced or field-placed concrete since they do not consider the uncertain and varying conditions in the field. This research provides another attempt to predict and assess industrial concrete performance in addition to a few previous studies on industrial concrete modeling.

This study comprehended the use of ML in concrete performance assessment. It used eleven well-known and conventional ML methods to predict three key performance characteristics of industrial concrete: (1) 7-day compressive strength (f_c7), (2) 28-day compressive strength (f_c28) and (3) fresh concrete slump. Big data with 12107 observations of industrially produced concrete are used to predict strengths, whereas slump is predicted using 3359 observations. This research further investigates the concrete maturity effect in predicting the strength. Finally, weather and atmospheric conditions during the concrete mixing and placing are also investigated in predicting the strength. A combination of research methodologies is followed to structure and organize the different parts of the thesis into a study of ML in concrete performance prediction.

First, a systematic and quantitative review of past studies is conducted to highlight the use of ML for the investigation and assessment of concrete performance. Text-mining-based modeling is applied to literature data, and informative analysis and topic modeling are performed. The output of the review highlights two main application areas of ML in concrete performance assessment: (1) the prediction of concrete properties; with compressive and shear strengths as majorly investigated properties, and (2) performance monitoring and assessment of concrete structures; with bridges and dam as two majorly investigated concrete structures.

Second, the f_c7 and f_c28 are predicted using mixture proportioning of seven concrete constituents (water, cement, coarse aggregate, fine aggregate, fly ash, superplasticizing admixture, and water reducing admixture) as predictors. It is found that non-linear models perform better than linear models. Moreover, the random forest (RF) model is found as the best among all other models and gain highest goodness of fit (i.e., highest determination coefficient (R²)) and smallest error in prediction (i.e., smaller root mean squared error (RMSE) and mean absolute error (MAE)). Compared to previous studies, the models of this study significantly improve the prediction of strength. For example, for RF, R² is improved by 17% for fc₂₈ prediction. Besides, this study confirms that data visualization is useful in learning about and summarizing the data, understanding the relationships of the variables, and making pre-modeling assumptions. Moreover, feature importance analysis indicates that cement, water reducing admixture, fly ash, and fine aggregate are the top influencing parameters in f_c7 and f_c28 prediction.

Third, the fresh concrete slump is predicted using 3599 observations of mixture proportioning of seven concrete constituents. A multiclass classification approach is adopted to predictively classify the slump into one of the eight characteristic classes (from 25 mm to 200 mm). Extreme gradient boosting (xgboost) and RF are found to be the two best ones that perform excellently well after a comprehensive comparison of the performance of the seven ML models against a metrics of accuracy, Kappa, Matthews correlation coefficient, logLoss, receiver operating characteristic plot, precision-recall plot, and the area under the curve corresponding to the two plots. Compared to past studies based on laboratory concrete with only 30 to 250 observations, models developed in this study better reflect the uncertainties in the industrial production of ready-mix concrete for actual job site applications and, therefore, are more directly relatable to the practice in the concrete industry.

Forth, the role of concrete maturity in the prediction of compressive strength is investigated and discussed. Most previous studies used mixture proportions and curing time to predict the strength, and they are limited to considering the partial effect of concrete maturity. Besides mixture proportions, this study takes curing time and early age strength as potential predictors to represent the concrete maturity effect in modeling. In this regard, three modeling scenarios are designed and evaluated. Moreover, this study considers both laboratory (1030 observations) and industrially (12,107 observations) obtained data for analysis. Comparison of modeling scenarios shows that considering concrete maturity improves the predictions, and the results are statistically significant for both data types. For example, with a 95% confidence and p-value < 0.05 improvement in R² range from 18.4% to 62.2% for industrial concrete and 0.19% to 15.3% for laboratory concrete modeling.

Fifth, the weather and atmospheric conditions during concrete mixing and placing is investigated in the prediction of f_c7 and f_c28 . In this regard, open-sourced data related to fourteen meteorological parameters are gathered using web-scraping of the Hong Kong Observatory (HKO) website. RF model is used for predictive modeling. After considering meteorological parameters, predictions are significantly improved, and R² is increased by 58.08% and 57.93% for prediction of f_c7 and f_c28, respectively. The feature importance analysis indicates that mean temperature, solar radiation, relative humidity, and evaporation are the top influential meteorological features in predicting strengths.

This study would shed useful insights for using ML in performance assessment of industrial concrete. More importantly, achieving higher accurate prediction of industrial concrete strengths and slump this study paves the way towards timely, cost-effective, and sustainable performance assessment of industrial concrete.