Predição da resistênica à compressão e do coeficiente de permeabilidade de concretos permeáveis por meio da aprendizagem de máquinas

Abstract

Due to the limitation of drainage systems and soil sealing, the chances of flooding occurring due to increased surface runoff are high on days of peak rainfall. The increase in impermeable areas also affects the recharge of the water table, as it hinders the infiltration of water into the soil. To mitigate the aforementioned disorders, as an alternative to traditional drainage systems, pervious concrete appears, which is a porous material that allows water to penetrate between its layers. Due to its high volume of voids, pervious concrete ends up presenting inferion physical properties when compared to conventional concrete, requiring studies that seek to optimize these aspects. Machine learning methods demonstrate success in predictive studies of pervious concrete properties. By adjusting the hyperparameters of each model, it is possible to further optimize the algorithms and increase the accuracy of the predicted results. Machine learning allows a trained model to seek results according to previously desired parameters. In addition, the desirability function tool optimizes responses simultaneously. The database used was selected A database was used in the work of Zhang et al. (2020) and is composed of different proportions of pervious concrete mixtures, taking into account the water/cement ratio, aggregate diameter and aggregate/cement ratio, and their respective responses that study the two main properties of pervious concrete, compressive strength and permeability coefficient. The data were processed, using the R language, in search of the model that presented the best behavior and optimized. The algorithms were adjusted for their optimization and the desirability function was used in search of the best configuration, maximizing the responses. For the compressive strength of pervious concrete, the model that presented the best result was the cubist model which, after adjustment, presented a value of R2 equal to 0.9994 for training and 0.9987 for testing. For the permeability coefficient, the best model was svmRadial, with an R2 value equal to 0.9887 for training and 0.9719 for testing, after adjustments. The required desirabilities were D1, maximum compressive strength value and maximum permeability coefficient value, D2, target compressive strength value of 20 MPa and target permeability coefficient value of 1 mm/s, and D3, strength value at target compression 20 MPa and maximum permeability coefficient value, with at least 1 mm/s. The global desirability results were, respectively, 0.7731, 0.9618 and 0.9289, demonstrating that the statistical technique is capable of improving and directing studies, finding the best possible result for each combination, mitigating expenses related to laboratory tests.