| Summary: | Uniaxial compressive strength (qu) of soil stabilized with cementitious binders is a key feature for design purposes. However, its measurement requires extensive laboratory tests, which is time and resources consuming. Accordingly, aiming to make this process faster and cheaper, this paper presents a novel approach for qu estimation of soil stabilized with cementitious binders based on soft computing techniques, particularly Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs). For models training, a database comprising 444 records, encompassing cohesionless to cohesive and organic soils, different binder types, mixture conditions and curing time was compiled. The results show a promising performance in qu prediction of laboratory soil-cement mixtures, being the best results achieved with the SVM model (2 = 0.94). In addition, by averaging SVM and ANN predictions a slightly better accuracy can be achieved (2 = 0.95). Through the application of a sensitivity analysis over the fitted models, it is measured the relative importance of each model attributes, which highlighted the major effects of water/cement ratio, cement content, organic matter content and curing time, which are known as preponderant in soil-cement mixtures behaviour.
|
|---|