A strategy for locating the rare, high-impact members of a huge simulation catalog when evaluating each member exactly is prohibitively expensive. The motivating problem is coastal climate-risk assessment: out of a large downscaled catalog of tropical cyclones, only a few produce destructive storm tides, but confirming which ones requires an expensive hydrodynamic simulation apiece, so brute-force Monte Carlo evaluation of the whole catalog is intractable, and climate non-stationarity keeps shifting the target distribution. The approach couples two ideas. First, a machine-learning surrogate is trained to cheaply predict storm-tide severity from storm features and shown to generalize to unseen climate scenarios with good precision and recall. Second, an informative online (active) learning loop bootstraps from a minimal set of fully simulated storms, then repeatedly uses the surrogate to pick the most informative next storms to simulate and retrain on, concentrating the expensive simulations where they most sharpen the identification of the rare destructive cyclones. The result is near-perfect retrieval of the rare extremes using only a small fraction of the simulations, giving a scalable, scenario-transferable workflow. The contribution is the surrogate-plus-active-sampling recipe for rare-extreme discovery under expensive evaluation and distribution shift.
LASSE (Learning Active Sampling for Storm tide Extremes) tackles a computational bottleneck in coastal climate-risk assessment: identifying which tropical cyclones, out of very large downscaled storm…