Mechanical properties of frozen soils (e.g., triaxial compressive strength, ${\sigma }_{tc}$ and Young’s modulus, $E$ ) are important in tunnel, shaft, or open pit excavation projects. Although numerous attempts have been made to develop indirect methods to estimate unfrozen soils’ ${\sigma }_{tc}$ and $E$ values, this has not been done with frozen soils given the difficulty of preparing and conducting relevant laboratory tests. In this study, the accuracy of artificial neural network (ANN), adaptive neuro fuzzy inference system (ANFIS), and support vector machine (SVM) models, developed to predict ${\sigma }_{tc}$ and $E$ for frozen sandy soils, was compared. To the best of the authors’ knowledge, no study has predicted frozen soils’ ${\sigma }_{t\mathrm{c}}$ and $E$ using these methods. Eighty-two poorly graded sandy soil samples from an urban subway borehole in Tabriz, Iran, were used to develop these models. It was found that temperature, confining pressure, strain rate, and yielding strain improved the accuracy of ${\sigma }_{tc}$ and $E$ prediction. Results indicate that SVM can successfully be used in predicting the ${\sigma }_{tc}$ and $E$ of frozen soils.

Accurate and reliable groundwater level prediction is an important issue in groundwater resource management. The objective of this research is to compare groundwater level prediction of several data-driven models for different prediction periods. Five different data-driven methods are compared to evaluate their performances to predict groundwater levels with 1-, 2- and 3-month lead times. The four quantitative standard statistical performance evaluation measures showed that while all models could provide acceptable predictions of groundwater level, the least square support vector machine (LSSVM) model was the most accurate. We developed a set of input combinations based on different levels of groundwater, total precipitation, average temperature and total evapotranspiration at monthly intervals. For each model, the antecedent inputs which included H_t-1, H_t-2, H_t-3, T_t, ET_t, P_t, P_t-1 produced the best-fit model for 1-month lead time. The coefficient of determination (R²) and the root mean square error (RMSE) were calculated as 0.99%, 1.05 meters for the train data set, and 95%, 2.3 meters for the test data set, respectively. It was also demonstrated that many combinations the above-mentioned approaches could model groundwater levels for 1 and 2 months ahead appropriately, but for 3 months ahead the performance of the models was not satisfactory.