Predicting concrete strength using machine learning and fresh-state measurements

Abstract Understanding the properties of lower-carbon concrete products is essential for their effective utilization. Insufficient empirical test data hinders practical adoption of these emerging products, and a lack of training data limits the effectiveness of current machine learning approaches for property prediction. This work employs a random forest machine learning model combined with a just-in-time approach, utilizing newly available data throughout the concrete lifecycle to enhance predictions of 28 and 56 day concrete strength. The machine learning hyperparameters and inputs are optimized through a novel unified metric that combines prediction accuracy and uncertainty estimates through the coefficient of determination and the distribution of uncertainty quality. This study concludes that optimizing solely for accuracy selects a different model than optimizing with the proposed unified accuracy and uncertainty metric. Experimental validation compares the 56-day strength of two previously unseen concrete mixes to the machine learning predictions. Even with the sparse dataset, predictions of 56-day strength for the two mixes were experimentally validated to within 90% confidence interval when using slump as an input and further improved by using 28-day strength.