In the realm of stochastic nonlinear response analysis for large and intricate structures， the Monte Carlo simulation meth‐ od stands out as a pivotal approach. However， its widespread practicality is hampered by its exorbitant computational costs. To sur‐ mount this challenge， researchers have endeavored to develop the active learning‐based Gaussian process surrogate model algo‐ rithm. Despite its promise in reducing computational expenses， the optimization strategy associated with active learning necessi‐ tates further refinement to meet the exacting demands of engineering applications. For this purpose， we introduce a search function endowed with ‘intelligent’ attention capabilities. This function is meticulously crafted to concentrate on exceedingly high‐risk one‐sided tail events in engineering scenarios. By incorporating this search function， we have engineered an algorithm that surpass‐ es existing methodologies. Our algorithm finds successful application in the analysis of complex adhesive anchoring structures with‐ in subway tunnel rings and linings. Compared to conventional methodologies， our algorithm exhibits a remarkable 30% reduction in the estimation error of single‐tailed probabilities. This advancement facilitates a more precise estimation of the one‐tailed proba‐ bility distribution governing the stochastic response of complex structures. Consequently， it enhances the precision of assessing the occurrence probability of extreme events. These findings yield invaluable insights for decision‐making processes in pertinent engi‐ neering domains and insurance sectors.