TR2019-096

Machine Learning Methods for Predicting the Field Compressive Strength of Concrete


    •  DeRousseau, M.A., Laftchiev, E., Kasprzyk, J.R., Balaji, R., Srubar III, W.V., "Machine Learning Methods for Predicting the Field Compressive Strength of Concrete", Construction And Building Materials, DOI: 10.1016/​j.conbuildmat.2019.08.042, Vol. 228, December 2019.
      BibTeX TR2019-096 PDF
      • @article{DeRousseau2019dec,
      • author = {DeRousseau, Mikaela A. and Laftchiev, Emil and Kasprzyk, Joseph R. and Balaji, Rajagopalan and Srubar III, Wil V.},
      • title = {Machine Learning Methods for Predicting the Field Compressive Strength of Concrete},
      • journal = {Construction And Building Materials},
      • year = 2019,
      • volume = 228,
      • month = dec,
      • doi = {10.1016/j.conbuildmat.2019.08.042},
      • url = {https://www.merl.com/publications/TR2019-096}
      • }
  • Research Area:

    Control

Abstract:

This research analyzes the compressive strength behavior of field-placed concrete (herein termed field concrete) as a function of mixture constituents. Compressive strength prediction of field concrete is inherently different and more challenging than that of laboratory concrete and merits its own analysis. In this work, we employ both field- and laboratory-obtained data to train and test machine learning models of increasing complexity for compressive strength prediction. This training and testing scheme enables determination of the best-performing model specific to field concrete. In this work, the random forest machine learning model for predicting field compressive strength generated the best performance; the RMSE, MAE, and R2 values were 730 psi, 530 psi, and .51, respectively. The methodological reasons for varying model performance are also examined. Finally, the ability of machine learning models trained on laboratory concrete data to predict the compressive strength of field concrete mixtures is evaluated and compared to those models trained exclusively on field concrete data. The analysis shows that the hybridization of field and laboratory data for building predictive models is a promising method for reducing common over-prediction issues caused by laboratory concrete models that are used in isolation