Advancing swelling soil prediction through TabNet-based machine learning framework

Predicting swell pressure in expansive soils remains a critical challenge in geotechnical engineering, where traditional empirical models often falter due to data scarcity and complex nonlinear interactions. This study introduces TabNet, a sparse, interpretable deep tabular learning architecture and compares it against conventional artificial neural networks (ANNs) for swell pressure forecasting using two compact datasets of 30 measurements each (untreated and treated soils) encompassing liquid limit (LL), shrinkage limit (SL), plastic limit (PL), optimum moisture content (OMC), maximum dry density (MDD), and free swell index (FSI). Remarkably, both models achieved unprecedented accuracy: TabNet delivered an R2 of 0.99, an nRMSE of 0.01, and a MAPE of 1.51%, surpassing all benchmarks in the literature despite the minuscule dataset size, a feat unattained previously. Feature importance analysis via TabNet’s mechanism pinpointed LL as the dominant predictor in untreated soils, with SL emerging as a key modulator in the network’s sparse feature selection, offering novel data-driven insights into feature contributions to swell‑pressure prediction. Even for chemically treated soils, the model maintained reliable predictions without preprocessing, though the limited data density somewhat reduced the interpretability of the results. These findings highlight TabNet’s potential for efficient geotechnical modelling, offering a robust solution where data or resources are limited.