Predicting the settlement of a shallow foundation on soil is a challenging engineering challenge since soil is a heterogeneous medium and because many of its effective properties are involved in the geotechnical behaviour for the soil-foundation system. The heterogeneity in soil properties is taken into consideration as our understanding of soils continues to grow. Because of this, the current study methodology has also changed from a deterministic to a probabilistic one. This paper examines the use of two probabilistic soft computing techniques, namely Functional Network (FN) and Multivariate Adaptive Regression Splines (MARS), to investigate the reliability of shallow foundations according to settlement criteria. These models can be used to regular design work because they are straightforward and trustworthy. Furthermore, FN and MARS were evaluated for suitability in predicting shallow foundation settlement while taking various soil properties into account. The models’ performance was evaluated using various fitness measures, such as β, VAF (variance account factor), RSR (standard deviation ratio), and RMSE (root mean square error). When compared to functional network (FN) (RMSE = 0.0017, VAF = 98.5127, RSR = 0.1416, NS (Nash-Sutcliffe efficiency) = 0.9799, RPD (relative percentage difference) = 7.0622), the MARS model performed better across the board in terms of all fitness characteristics MARS (RMSE = 0.0002, VAF = 99.9645, RSR = 0.0187, NS = 0.9998, RPD = 53.0204). The findings demonstrate that the MARS approach is a dependable soft computing method for solving non-linear issues such as the settling of shallow foundations on soils.