Frederick Iat-Hin Tam

NWC Colloquium

Frederick lat-Hin Tam

AI for Knowledge Discovery in Tropical Meteorology: Pattern Extraction, Uncertainty Quantification, and Causality

Tuesday, April 23th, 2024

3:00 pm

NWC 1313

Recent advancements in machine learning (ML) have led to data-driven weather prediction models that rival or outperform state-of-the-art numerical weather prediction models in short— to medium-range forecasts. In order to expand the utility of ML tools in weather prediction and research, practitioners need to have high trust in these tools. The adoption of Explainable Artificial Intelligence (XAI) tools is critical as they show how the ML models make their predictions. Furthermore, it is important to show that ML models can be used to obtain new physical insights on different weather forecasting problems, such as the intensification of tropical cyclones. In this two-part presentation, I will present two studies that demonstrate the added value of ML tools in (a) discovering new knowledge on the early intensification of tropical cyclones, and (b) improving existing statistical models for TC intensity forecast. The first example [1] concerns the role of radiation in tropical cyclogenesis (TCG). Traditional diagnostic methods to isolate the effect of radiation on TCG have assumptions that remove important spatiotemporal context, making them suboptimal in studying the effect of three-dimensional or transient radiative heating. The diagnosed kinematic responses cannot be directly converted to TC surface intensity. To address this, we developed an ML feature extraction algorithm based on the Variational Encoder-Decoder (VED) architecture to predict 24-hour TC surface intensification. Applied to Typhoon Haiyan (2013) and Hurricane Maria (2017), the VED model discovers radiative patterns that are most informative to the prediction. Different aspects of these wavenumber-1 asymmetric structures will be highlighted to demonstrate that the VED model yields new insights into how the spatial distribution of cloud affects intensification rates. Furthermore, we show that uncertainties in the VED prediction differentiate intensification periods strongly coupled to radiation from periods not coupled to radiation. The second example [2] illustrates the importance of causal feature selection in statistical TC intensity models. Traditional feature selection methods, based on correlation-based metrics, may overlook confounders and keep features that are highly correlated to but not directly causing changes in TC intensity (“spurious causal links”). Models with fewer spurious links could generalize better to unseen cases than models with more spurious links. As a proof of concept, we employ the recently developed multidata M-PC1 algorithm to infer the causal relationships between TC intensity and different ERA5 predictors in a curated time series dataset of 260 Western Pacific Typhoons. ML models trained on features filtered with M-PC1 outperform models with features chosen from other traditional feature selection baselines (e.g., lag correlation, XAI-based, lasso-based, random). The causal ML model with optimal generalizability underscores the importance of inner core boundary layer moisture, hitherto underutilized in TC intensity predictions. Motivated by this, we reexamine the input feature list in the existing Statistical Hurricane Intensity Prediction Scheme (SHIPS). Our results demonstrate that causal inference and the removal of causally spurious features improve the prediction skills of SHIPS for some Atlantic hurricanes. To conclude, these examples highlight the potential of data-driven methodologies to enrich our understanding of tropical cyclones and improve operational empirical statistical schemes by uncovering new causal drivers and removing old ones. This paves the way for the broader application of these tools across different prediction problems in tropical meteorology