estimation of fiber pullout curve from cement base matrix using machine learning

Pull-out tests and corresponding curves help analyze fiber pull-out behavior, which significantly influences the tensile and flexural properties of fiber-reinforced concrete. artificial intelligence techniques provide an efficient means to predict fiber pull-out curves. in this study, four models based on convolutional neural networks (cnn), long short-term memory networks (lstm), and extreme gradient boosting (xgboost) are used to predict fiber pull-out curves: cnn1d, cnn2d, lstm, and xgboost. a dataset of 502 experimental samples was compiled, primarily from laboratory studies conducted by one of the authors of this paper. fiber aspect ratio, loading rate, type and quantity of cement, amount of: binder, silica fume, sand, gravel, quartz, superplasticizer, water, water-to-cement ratio, water-to-binder ratio, curing age, fiber embedded length, end-type of the fibers, the pitch of spirals, number of twists, fiber inclination angle, fiber length and diameter, and concrete compressive strength are among the used input parameters. to represent the expected curve, the model output comprises 1000 pairs of data (pull-out force vs slip). among the tested models, xgboost demonstrated superior performance with the lowest mean absolute error (8.39) and highest r² value (0.71), making it the optimal choice for predicting fiber pull-out curves. the parameters “embedded length”, “number of twists”, “silica fume”, and “pitch of spirals” were more important and influenced the model and prediction accuracy, as can be seen from the evaluation of the feature importance graph.