This study generated a numerical database using finite-element limit analysis with the Bolton model, then trained XGBoost, Random Forest and evolutionary polynomial regression to predict the Nγ bearing-capacity factor for strip footings on dense sand slopes. XGBoost test R² values were 0.994, 0.943 and 0.953 at 15°, 30° and 45° slopes, but accuracy was measured against simulations within the same modelling framework, not field data.
Key findings
- XGBoost achieved test R² values of 0.994, 0.943 and 0.953 for 15°, 30° and 45° slopes. Particle crushing strength, relative density, footing width and critical-state friction angle were influential. H/B mattered more on steeper slopes; wider footings reduced Nγ through stress-dependent suppression of dilatancy, and the numerical response approached level-ground behaviour at L/B=6.
Why this matters globally
Footings near slopes are safety-critical and detailed analysis is computationally demanding. A framework combining high predictive accuracy with interpretable equations could support preliminary screening, provided applicability limits and safety codes are strictly observed.
Thai researcher contribution
Thammasat University researchers integrated soil mechanics, limit analysis and machine learning from database generation to interpretable preliminary-design equations.
Limitations to consider
All data came from two-dimensional FELA under the Bolton dense-sand assumptions. R² reflects interpolation within the parameter space, not physical or field validation. Layering, groundwater, three-dimensional footing shapes, eccentric loads and construction uncertainty were not covered, and the model should not replace engineering judgment.