<p>Conducting dynamic reliability analysis for steel structures subjected to ground motions is crucial to ensure the safety of the structures amidst various sources of uncertainties. However, this task is highly time-consuming and resource-intensive because it requires performing both dynamic analysis and reliability analysis simultaneously. Therefore, this study introduces a novel adaptive method that integrates two main components: i) a modern deep learning-based prediction model for approximating the structures’ dynamic responses and ii) an uncertainty quantification based on quantile regression for assessing prediction uncertainty. An adaptive learning framework then seamlessly combines these components into a unified computation procedure for reliability analysis. The effectiveness and efficiency of the proposed approach are demonstrated through two structures with multiple components: a 2D frame structure with a non-linear damper and a 3D six-story structure with up to 52 random variables. The results clearly reveal that the proposed method can accelerate reliability analysis by approximately 60 times compared to the conventional Monte Carlo simulation. Additionally, it is one of the most accurate methods, featuring very fast inference time and requiring less training data compared to other data-driven counterparts.</p>