Answers are grounded in this paper and cite their sources.
Deep learning models for seismic response prediction often require extensive datasets generated through repeated time-history analyses. This study investigates uncertainty-driven active learning for training a long short-term memory model to predict complete roof-displacement histories of base-isolated structures. A numerical database comprising 1000 linear time-history simulations was developed using recorded earthquake ground motions and varying isolation parameters. Monte Carlo dropout-based uncertainty acquisition was compared with random and diversity-based sampling across 10 matched repetitions and labeled-sample budgets ranging from 100 to 300. Performance was evaluated on ground-motion records withheld from model development. The results reveal substantial variability and nonmonotonic learning behavior. Uncertainty-based acquisition did not consistently outperform the alternative strategies, and negative mean R 2 values indicated difficulties in generalizing to unseen ground-motion records. Nevertheless, Monte Carlo dropout uncertainty scores exhibited a strong positive correlation with actual prediction errors (Spearman ρ = 0.8697). These findings highlight both the diagnostic value and the limitations of uncertainty-driven sample acquisition for seismic response-history prediction.
No takes yet. Share an insight, caveat, or question.
Arcan Yanik (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: