Tropical cyclone (TC) risk assessment models and probabilistic forecasting systems rely on large ensembles to simulate the track trajectories, intensities, and spatial distributions of damaging winds from severe events. Given computational constraints associated with the generation of such ensembles, the representation of TC winds is typically based on very simple parametric formulations. Such models strongly underestimate the full range of TC wind field variability and thus do not allow for accurate representation of the risk profile. With this in mind, this study explores the potential of machine learning algorithms as an alternative to current parametric methods. First, a catalog of high-resolution TC wind simulations is assembled for the western North Pacific using the Weather Research and Forecasting (WRF) Model. The simulated wind fields are then decomposed via principal component analysis (PCA) and a quantile regression forest model is trained to predict the conditional distributions of the first three principal component (PC) weights. With this model, predictions can be made for any quantiles in the distributions of the PC weights thereby providing a way to account for uncertainty in the modeled wind fields. By repeatedly sampling the quantile values, probabilistic maps for the likelihood of attaining given wind speed thresholds can be easily generated. Similarly the inclusion of such a model as part of a TC risk assessment framework can greatly increase the range of wind field patterns sampled, providing a broader view of the threat posed by TC winds.

No abstract