Bearing capacity and safety factors of ring foundations in spatially variable soils: a hybrid FELA-machine learning approach

Ring foundations efficiently support axisymmetric structures like silos, storage tanks, chimneys, and cooling towers with reduced concrete volume compared to solid circular foundations. Their ultimate bearing capacity is strongly affected by asymmetric loading and spatial variability in soil properties. This study presents a hybrid machine learning framework to predict normalized bearing capacity factors and safety factors for ring foundations in spatially variable cohesive-frictional (c-φ) soils. Datasets were generated via finite element limit analysis (FELA) in OptumG2, incorporating random field modeling of effective cohesion (c’) (lognormal, depth-linear mean: µ c’ = 10 kPa, k = 1 kPa/m, COV = 0.3) and friction angle (φ’ up to 40°). Four loading configurations (complete, internal half, middle half, external half), inner-to-outer radius ratios of 0–0.9, rough/smooth interfaces, and nine anisotropic scale-of-fluctuation cases (horizontal: 10–60 m; vertical: 2–8 m) were analyzed. A Convolutional Neural Network processed random field images and parameters, achieving R² > 0.95 and relative errors < 3%. External half and complete loading gave highest capacities; internal half reduced them by up to 40%. Spatial variability in c’ and φ’ contracted failure envelopes by 10–25%, emphasizing reliability-based design. The framework dramatically reduces computational cost versus traditional FELA/Monte Carlo while retaining accuracy, supporting preliminary design, optimization, and risk assessment.