Prediction of soil water retention curve based on physical characterization parameters using machine learning

This paper explores the potential of machine learning techniques to predict the soil water retention curve based on physical characterization parameters. Results from 794 water retention and suction points obtained from 51 different soils were used in the algorithm. The soil properties used are the percentages of gravel, sand, silt, and clay, the plasticity index, the porosity, and the relation between the volumetric water content and total suction. The data were used as input for machine learning estimators to predict the volumetric water content of a soil with specified physical characterization parameters and suction, the techniques of artificial intelligence were developed in python. Results show that an extremely randomized trees’ estimator can reach a coefficient of determination of 0.99 in the training dataset, with a coefficient of 0.90 in the cross-validation and testing dataset, which measures the generalization capacity. Furthermore, a continuous function can be obtained by fitting a model such as Cavalcante & Zornberg, or van Genuchten, or Costa & Cavalcante (bimodal) to the predictions of the machine learning for use in numerical methods. These results indicate that the proposed machine learning estimator can become an interesting alternative to estimate the soil water retention curve in engineering practice. This work is in progress and the predictions can be improved with the addition of new data. Know how to participate at the end of the paper.