<p>The problem considered in this paper is the probabilistic stability analysis of an active trapdoor beneath spatially random clay exhibiting anisotropic behavior. The adaptive finite element limit analysis (AFELA), with Monte Carlo simulations, is employed to evaluate the probability of active failures, while several novel machine learning frameworks are integrated with random field theory (RFT) to study the effects of cover depth ratio, anisotropic ratio, coefficient of variation, and dimensional correlation length. Using a dataset of 1152 data points produced by random adaptive finite element limit analysis (RAFELA), several advanced machine learning models are developed using several ML algorithms, namely Random Forest (RF), Gradient Boosting Machine (GBM), and eXtreme Gradient Boosting (XGB). The primary objective is to predict the probability of failure (<i>PoF</i>) for trapdoors by considering several practical values of factor of safety (<i>FoS</i>). The results indicate that an increase in coefficient of variation of undrained shear strength (<i>COV</i>_<i>suc</i>) and cover depth ratio (<i>C/W</i>) leads to a higher <i>PoF</i>, while smaller correlation lengths (Θ_<i>suc</i>) increase failure probability. Among the three models, XGB achieves the highest accuracy, demonstrating its effectiveness in capturing complex failure trends. With the novel soft-computing methods developed for predicting the stability of trapdoors beneath spatially random anisotropic clays, this research on random fields represents a significant advancement over traditional deterministic design approaches using the factor of safety.</p>