| Abstract [eng] |
Hydraulic conductivity is the ability of soil to transmit water and is measued by the rate which water can move through the porous medium. The hydraulic conductivity of soil is affected by numerous factors like soil physical properties and grain size, and has a significant role in fields like geotechnical design, contaminant migration and waste disposal. It can be determined directly in situ or through laboratory tests. Indirect methods include empirical formulas and machine learning modelling. Both utilize physical soil parameters in the determination of hydraulic conductivity. Machine learning is useful in the sense of computational capacity to process data that it can more easily find relationships between multiple parameters as well as produce new predictions by using complex algorithms, while empirical formulas require manual data processing. This thesis investigates theoretical hydraulic conductivity determination methods of Lithuanian soil samples. The primary objectives of this study are creating a database for Lithuanian soil samples and assessing theoretical hydraulic conductivity methods by comparing them to laboratory-acquired values. The study is conducted by using three empirical formulas from Hazen, Slichter and USBR and tuning seven machine learning regression models to find the best parameters to use in the determination of hydraulic conductivity. The regression models used in the study are linear and ridge regression, support vector regression (SVR), K-Nearest Neighbors (KNN), Decision tree, Random forest and Gradient boosting. The results reveal that from the three empirical formulas, Hazen‘s formula performs the best while Slichter‘s formula has the lowest correlation to actual hydraulic conductivity values. However, the overall accuracy of these empirical formulas remains low. From the six machine learning models, Random forest performed the best in multiple different tests and by using different parameters. The highest correlation is achieved by using grain size information of fine soils, grain size diameters D60 and D70, and water content. Overall, the machine learning models performed better than the empirical formulas. The results reveal that the machine learning models can adjust to the heterogenous nature of soils and find patterns between multiple soil parameters. The overall correlation of both empirical formulas and machine learning models remain relatively low, and the main reason for this might be due to the small size of the database entries. It is encouraged to keep updating the soil sample database to gain a wider resource for future investigations of soil permeability in Lithuania. |
