Forecasting the Sudden Turning Tracks of Tropical Cyclones Using a Causality-Based and Internal Dynamics-Informed AI Model

The forecasting of sudden turning tropical cyclone (TC) tracks is a notoriously difficult task in meteorology. Recently, this has been tackled through machine learning in combination of causality analysis, with the introduction of new predictors that reflect the key internal atmospheric processes governing the TC genesis and maintenance, during the stage of covariate selection/sparsity identification. Here the presentation will be sectioned, in a sequential order, into three parts, with the final part being about a couple of real forecasts.

1. Causality ab initio and causal machine learning

We have introduced causality analysis into the current artificial intelligence (AI) algorithms, in the hope of overcoming the interpretability crisis. Incorporation of causality into machine learning, however, is challenged with its vagueness, non-quantitativeness, computational inefficiency, etc. During the past 18 years, these challenges have been essentially resolved, with the advent of a rigorous formalism of causality analysis, i.e., the Liang-Kleeman formalism, which was initially motivated from atmospheric predictability. This not only opens a new field in the atmosphere-ocean science, namely, information flow, but also has led to scientific discoveries in other disciplines, such as quantum mechanics, neuroscience, financial economics, etc., through various applications. In this talk we will provide a brief review of the decade-long effort, including some major theoretical results, e.g., invariance upon arbitrary nonlinear coordinate transformation, and a corollary that expresses in a concise mathematical formula the long-standing philosophical debate on causation versus correlation (causation implies correlation, but correlation does not imply causation). Also provided will be an introduction of the causal deep learning framework, and some representative real-world applications pertaining to this conference, e.g., the decadal prediction of El Niño Modoki, the forecasting of an extreme drought, among others.

2. New predictor apart from traditional predictors

The genesis, intensification, maintenance, and decay of a TC have been extensively studied during the past forty years. The internal dynamical processes regarding the redistribution of energy among different scales, however, have been mostly overlooked. Using the functional analysis machinery, multiscale window transform, Wang and Liang (2017) probably was among the first who did this research. They discovered how a barotropically unstable background flow may inject energy to a typhoon through canonical transfer. Recently, Rong and Liang (2022) composited the canonical transfers, both baroclinic transfer and barotropic transfer, for all the typhoons in Northwestern Pacific during 1995-2016 (201 typhoons in total), using the 0.25° resolution, 6-hourly ERA-Interim reanalysis data, and found that barotropic transfer, as well as the non-adiabatic heating, dominates the TC-scale (defined as the time scale of 1-32 days) energetics balance. For typhoon movement, the asymmetric kinetic energy balance plays an important role. The barotropic canonical transfer from the background circulation to the typhoon-scale, and the transport of typhoon-scale KE have obvious asymmetric structures. In particular, the former tends to be in the form of a dipole at the center of the typhoon. Most importantly, the typhoon moves in the direction of the dipole, i.e., the direction from the positive barotropic canonical transfer center to the negative center. In the composite pattern of the sudden northward turning cases, the shift from the southeast (positive)-northwest (negative) barotropic transfer dipole to the southwest (positive)-northeast (negative) dipole is completed 36-12 hours before the sudden turning of the typhoon. This implies that the sudden turning of a recurvatured typhoon can be predicted at least 12 hours in advance, by taking into account the canonical transfer alone..

3. Causality-based typhoon trajectory forecasting model

With the above preparation, we combine deep learning with the information flow-based causality analysis to construct a causality-based AI model for the typhoon trajectory forecasting. Based on the dynamics and predictors for typhoons as diagnosed, the longitude and latitude variabilities at a lead time of 48 hours are predicted. The candidates for the inputs include, in addition to the traditional predictors such as the subtropical high index, steering flow, vertical wind shear, sea surface temperature, to name but a few, the newly introduced canonical transfer as elaborated above. The predictors are screened out from the pool of candidates through causality analysis, and then input into the error back propagation (BP) neural network and convolutional long and short-term memory neural network (ConvLSTM), respectively, and a typhoon trajectory forecasting model is henceforth trained. One thing that merits mentioning is that the inputs may vary during the course of TC evolution; they should be selected dynamically based on the result of causal inference. So far with the cases we have examined, our causality-based ConvLSTM model proves to be promising in forecasting the recurvature of typhoon, albeit with coarse spatial and temporal resolution data. Particular good performance has been demonstrated in the forecasting of sudden turning tracks, such as those for the 2015 typhoons Linfa and Nangka (see the attached figure), whose trajectories changed abruptly, making the forecasts by US Joint Typhoon Warning Center, Japan Meteorological Agency, and China Meteorological Administration all off the mark.