Explainable AI-Driven Multi-Objective Optimization of Eccentrically Loaded Reinforced Concrete Strip Foundations

Abstract

The design of eccentrically loaded reinforced concrete strip foundations traditionally relies on analytical bearing-capacity theory and trial-and-error sizing, which often result in overconservative dimensions and inefficient use of materials. This study proposes an explainable, end-to-end Artificial Intelligence (AI) framework that simultaneously minimizes construction costs and maximizes safety while strictly enforcing the middle-third eccentricity constraint. The framework operates in four tiers. Latin Hypercube Sampling (LHS) generates a dataset of 10,000 design scenarios from analytical models. Data generation is followed by training eXtreme Gradient Boosting (XGBoost) surrogate models to predict ultimate bearing capacity, cost, factor of safety, and settlement. SHapley Additive exPlanations (SHAP) are then employed to assess the surrogate models in the context of soil mechanics. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) uses these surrogates as millisecond-scale fitness evaluators to identify the Pareto front, from which the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) selects the knee-point design. The XGBoost surrogates reproduce the analytical responses with near-unity accuracy (  ≥ 0.996), enabling full optimization at negligible computational cost. SHAP confirms that the friction angle and cohesion govern bearing capacity exponentially, while width and thickness dominate cost, verifying that the model learned the correct physics. The Pareto front quantifies a safety–cost trade-off spanning factors of safety from 3.0 to 7.4. The TOPSIS optimum favors a wider base and a thinner footing, cutting cost by about 26% while increasing the factor of safety by about 22% relative to conventional design. The study provides a decision-complete, transparent pipeline that converts foundation design from manual trial-and-error into an automated, explainable, and efficient practice. Furthermore, the framework is generalizable to other foundation types and to environmental objectives such as embodied carbon.