SWELLING CHARACTERISATION AND RISK CLASSIFICATION OF BLACK COTTON SOIL USING EMPIRICAL CORRELATIONS AND MACHINE LEARNING

Abstract

Black cotton soils (BCS) of Deccan Trap region of India are characterized by the presence of high proportion of montmorillonite clay mineralogy and exhibit high swell-shrink characteristics leading to distress in roads, foundations, embankment and lightly loaded structures. In Indian geotechnical practice, the Free Swell Index (FSI) determined by comparing the sedimented volume of oven dried soil in distilled water with kerosene is the standard index of expansiveness of the soil. The FSI test is relatively simple to perform in a well-equipped lab, but the Atterberg limit data (liquid limit, plastic limit, plasticity index and shrinkage limit) are readily available from the earliest stages of site investigation. Therefore, it would be beneficial to establish reliable relationships between these index properties and FSI and provide a complementary assessment tool for geotechnical engineers that would not replace the actual FSI test. This study aims to fill that need with a two-part investigation. During the first phase, the black cotton soil was taken from Shajapur district, Madhya Pradesh and mixed with sodium bentonite at 0%, 2%, 4%, 6%, 8% and 10% of the dry weight of soil to prepare six soil mixtures (M1 – M6). Each mixture was tested for Atterberg limits (LL, PL, PI, SL) and FSI as per IS:2720 and IS:9451 respectively. The Pearson correlation analysis showed that all four Atterberg limit parameters were highly correlated with FSI: LL (r = +0.9998), PI (r = +0.9997) and SL (r = −0.9981). Five empirical regression equations were developed, ranging from single-variable Atterberg limit models — FSI = 2.299·PI − 22.315 (R² = 0.9995) and FSI = 2.037·LL − 61.289 (R² = 0.9996) — to dual-Atterberg-limit models incorporating PI and SL (R² = 0.9996) and LL and SL (R² = 0.9997). The single-variable PI and LL models were found to be the most robust models in this small dataset by leave-one-out cross validation (LOOCV), with RMSE values of 0.53% and 0.47%, respectively. The second phase involved training and testing three machine learning classifiers (Random Forest (RF), Support Vector Machine (SVM), Multilayer Perceptron (ANN)) on a compiled database of 186 Indian BCS samples from ten states to III predict the IS:1498 swelling risk class (Low, Medium, High, Very High) from LL, PI, SL, optimum moisture content (OMC), and maximum dry density (MDD). The class imbalance issue was addressed using stratified 5-fold cross validation. Random Forest was the best model in terms of balanced accuracy (0.860, CV accuracy = 0.880 and Cohen's κ = 0.792), indicating its appropriateness for categorical swelling risk assessment. The derived empirical equations can be used to estimate FSI from Atterberg limit index properties for Shajapur-type Malwa Plateau BCS and the Random Forest classifier can be used to map the swelling potential of BCS in a geographically wider Indian context. Both are designed to be complementary rapid assessment tools to support preliminary decision making in the field by the geotechnical engineer, rather than replacing direct measurement of FSI where site conditions require it.