Swelling–Shrinking Behavior and Deformation Prediction of Weakly Expansive Soils under Wetting–Drying Cycles with Different Hydraulic Coupling Paths

Clarifying the swelling and shrinking characteristics of expansive soils is crucial to addressing engineering problems in regions with such soils. This study collected deformation data of expansive soil under different hydraulic pathways during wetting–drying (W-D) cycles through laboratory experiments. The results revealed that both the W-D amplitude and vertical stress affect the volumetric change behavior of expansive soils during W-D cycles. The soil reached an elastic volumetric state after multiple W-D cycles, in which the swelling–shrinkage strain decreases with increasing vertical stress or decreasing W-D amplitude. Additionally, machine learning models were trained with different algorithms and data formats based on the time-series characteristics of soil deformation. The results demonstrate that using deformation data as sequential inputs to train sequence-based machine learning models significantly improves prediction accuracy. The gated recurrent unit model yielded optimal prediction accuracy (test set R2 > 0.92), consistently outperforming traditional artificial neural network models. Furthermore, the study highlights the critical role of experimental design in model performance, revealing that the judicious selection of test conditions is more critical for enhancing model accuracy than merely increasing the number of tests. The predictive model maintained high accuracy (R2 = 0.80) under the optimal testing scheme (42% experiment reduction). The findings provide a valuable reference for predicting the deformation of expansive soil.