Abstract
The active development of the construction industry leads to the fact that territories composed of specific soils are covered. The number of storeys and the complexity of the constructed structures are increasing. Limited financial and time resources require more efficient methods of pre-survey of the development site. In this regard, the requirements for engineering and geological research are increasing. Areas composed of swelling soils require special attention. Swelling clay rocks have the ability to change their volume when wet and thus pose a serious threat to the stability of structures and foundations. The measurement of swelling characteristics takes a long period of time and requires a significant investment of money. The purpose of this study is to establish the relationship between the physical and geological characteristics of clay soils, which require minimal resources, and their relative swelling using machine learning methods. As input data for building a relative swelling model of clay rocks, the physical characteristics of soils were used, such as water content, liquid limit, plastic limit, plasticity index, liquidity index, solid particles density, soil density, dry soil density, buoyant density, void ratio, maximum water capacity, soil's moisture level. According to these characteristics, three relative swelling models were built using the following machine learning methods: Decision tree, Gradient Boosting, Random Forest. The models were evaluated using the Mean Squared Error parameter, which measures the standard deviation between predicted and true values. As a result of evaluating this criterion, the Random Forest model was chosen to build the relative swelling model. The presence of a relative swelling predictive model will save money and time resources at the preliminary stage of engineering and geological studies