Machine Learning Prediction of Flexural Behavior of UHPFRC

To evaluate the possibility of predicting the flexural behaviour of UHPFRC, four analytical models were developed, based on artificial neural networks (ANN), to predict the first cracking tension or Limit of Proportionality (LOP), its corresponding deflection (δ_LOP), ultimate strength or Modulus of Rupture (MOR), and its corresponding deflection (δ_MOR) of UHPFRC under bending test. The models that were composed of an input level, one output level, and four hidden levels were developed through the R platform. The input level applied the most significative Principal Components (PC) of a large dimension of input dataset. To avoid overfitting K-fold validation and l² regularization was used. After the models were created, an improvement based on assembling of models by incorporating the predicted values in the dataset of features. The results indicated that the developed assembling models have a good accuracy for the prediction of the behaviour of UHPFRC under three or four points bending test, even when containing supplementary cementitious materials and hybrid mixture of fibers.