Rustu Orkun Karamut, Adil Binal

Artificial neural networks-based ternary charts for predicting strength and frost heaving in mountain soils

Abstract Abstract. Hydraulic conductivity (k) is a crucial parameter in hydrogeology and engineering geology, describing the rate at which water filters through porous media, i.e., soil. It can be determined directly through tests on soil samples in situ, or it can be calculated from other soil parameters using various equations and models. This study aims to compare the results of six machine learning (ML) models with those of four empirical formulas and to identify the soil parameters required for the optimal ML performance. A dataset consisting of 282 unique entries of Lithuanian soils was compiled from laboratory testing reports. Twelve features, including grain sizes and particle diameters, were used to create 4095 combinations of inputs for each ML algorithm. Prediction results were evaluated using the determination coefficient (R²) and the mean absolute error (MAE). The ML models provided more accurate predictions (R² 0.36–0.46, MAE 2.31–2.81 m/d) compared to the empirical formulas (R² 0.10–0.33, MAE 3.05–6.54 m/d). However, some ML models showed signs of overfitting. The study also revealed that each ML algorithm performs best with a customized combination of input parameters, ranging from 4 to 8, whereas the empirical formulas used in this study utilize only 1–2 parameters.

Doi https://doi.org/10.5200/baltica.2024.2.6

Keywords mountain soil; physical-mechanical properties; frost-heaving pressure; ternary diagram

Full text