Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

3DTCR: A Physics-Based Generative Framework for Vortex-Following 3D Reconstruction to Improve Tropical Cyclone Intensity Forecasting

Jun Liu, Xiaohui Zhong, Kai Zheng, Jiarui Li, Yifei Li, Tao Zhou, Wenxu Qian, Shun Dai, Ruian Tie, Yangyang Zhao, Hao Li

Abstract

Tropical cyclone (TC) intensity forecasting remains challenging as current numerical and AI-based weather models fail to satisfactorily represent extreme TC structure and intensity. Although intensity time-series forecasting has achieved significant advances, it outputs intensity sequences rather than the three-dimensional inner-core fine-scale structure and physical mechanisms governing TC evolution. High-resolution numerical simulations can capture these features but remain computationally expensive and inefficient for large-scale operational applications. Here we present 3DTCR, a physics-based generative framework combining physical constraints with generative AI efficiency for 3D TC structure reconstruction. Trained on a six-year, 3-km-resolution moving-domain WRF dataset, 3DTCR enables region-adaptive vortex-following reconstruction using conditional Flow Matching(CFM), optimized via latent domain adaptation and two-stage transfer learning. The framework mitigates limitations imposed by low-resolution targets and over-smoothed forecasts, improving the representation of TC inner-core structure and intensity while maintaining track stability. Results demonstrate that 3DTCR outperforms the ECMWF high-resolution forecasting system (ECMWF-HRES) in TC intensity prediction at nearly all lead times up to 5 days and reduces the RMSE of maximum WS10M by 36.5% relative to its FuXi inputs. These findings highlight 3DTCR as a physics-based generative framework that efficiently resolves fine-scale structures at lower computational cost, which may offer a promising avenue for improving TC intensity forecasting.

3DTCR: A Physics-Based Generative Framework for Vortex-Following 3D Reconstruction to Improve Tropical Cyclone Intensity Forecasting

Abstract

Tropical cyclone (TC) intensity forecasting remains challenging as current numerical and AI-based weather models fail to satisfactorily represent extreme TC structure and intensity. Although intensity time-series forecasting has achieved significant advances, it outputs intensity sequences rather than the three-dimensional inner-core fine-scale structure and physical mechanisms governing TC evolution. High-resolution numerical simulations can capture these features but remain computationally expensive and inefficient for large-scale operational applications. Here we present 3DTCR, a physics-based generative framework combining physical constraints with generative AI efficiency for 3D TC structure reconstruction. Trained on a six-year, 3-km-resolution moving-domain WRF dataset, 3DTCR enables region-adaptive vortex-following reconstruction using conditional Flow Matching(CFM), optimized via latent domain adaptation and two-stage transfer learning. The framework mitigates limitations imposed by low-resolution targets and over-smoothed forecasts, improving the representation of TC inner-core structure and intensity while maintaining track stability. Results demonstrate that 3DTCR outperforms the ECMWF high-resolution forecasting system (ECMWF-HRES) in TC intensity prediction at nearly all lead times up to 5 days and reduces the RMSE of maximum WS10M by 36.5% relative to its FuXi inputs. These findings highlight 3DTCR as a physics-based generative framework that efficiently resolves fine-scale structures at lower computational cost, which may offer a promising avenue for improving TC intensity forecasting.
Paper Structure (17 sections, 6 figures, 1 table)

This paper contains 17 sections, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Results
  3. Constructing physically high-resolution ground truth
  4. Overall tropical cyclone intensity forecast performance
  5. Reconstruction of fine-scale 3D structures of TCs
  6. Three-Dimensional Structural Characteristics
  7. Multi-scale Spectra and Harmonic Analysis
  8. Ablation Study: Effectiveness of Two-Stage Training and Latent Space Domain Adaptation
  9. Computational Efficiency and Adaptive Region Reconstruction Strategy
  10. Discussion
  11. Materials and Methods
  12. Data
  13. 3DTCR model architecture
  14. Time-Aware Latent Domain Adaptation via MMD to Mitigate Extreme Underestimation
  15. 3DTCR model architecture and training
  16. ...and 2 more sections

Figures (6)

  • Figure 1: Scatter plot comparison of maximum 10-m wind speed (WS10M) predictions across different methods. Panels (a--i) display the correlation between predicted and observed (IBTrACS) values for 2,242 test samples pooled across all forecast lead times from the 2024 Northwest Pacific dataset. Solid lines denote the global linear regression fit with 95% confidence intervals, contrasted against the ideal 1:1 reference (dashed lines). The rightmost column illustrates the marginal distributions via box plots overlaid on kernel density estimates (KDEs), while the inset values represent the global Mean Bias and RMSE metrics.
  • Figure 2: Evaluation of maximum 10-meter wind speed forecasts across lead times.(a) RMSE versus forecast lead time, with numbers in parentheses denoting sample sizes. (b) Box-and-whisker plots of forecast bias (model minus observation), where boxes represent the interquartile range (IQR) and horizontal lines indicate the median. Both panels compare standard baselines—FuXi (blue), ERA5 (red), ECMWF-HRES (orange), and WRF (green)—against the proposed 3DTCR framework: pre-trained models using ERA5 (beige) or FuXi (pink) inputs, and the final SFT model (purple).
  • Figure 3: Typhoon KONG-REY multi-variable field reconstruction at 18-h forecast lead time. Reconstruction comparison of the TC vortex structure through 2-meter temperature (T2M, K), 10-meter u/v wind components (U10M, V10M, m/s), mean sea level pressure (MSL, hPa), and 10-meter wind speed (WS10M, m/s), initialized at 12 UTC October 29, 2024. Rows from top to bottom: Interpolation (FuXi), 3DTCR reconstruction, and WRF 3km simulation (ground truth).
  • Figure 4: Multi-level wind speed profiles and spectral analysis for Typhoon KONG-REY. Evolution of wind speed characteristics across different atmospheric levels: 200 hPa (first row), 500 hPa (second row), 700 hPa (third row), 850 hPa (fourth row), and 10-meter height (fifth row). Each row shows: wind speed profiles along distance (first column), probability density functions of wind speed (second column), wind speed power spectra (third column), and kinetic energy spectra (fourth column). Results are compared among ERA5 reanalysis (red), FuXi forecast (blue), 3DTCR model (purple), and WRF 3km simulation (green). Power and kinetic energy spectra are displayed in log-log scale, revealing the multi-scale characteristics and energy cascade of the typhoon circulation.
  • Figure 5: Ablation analysis of 3DTCR using different training strategies. Comparison of Critical Success Index (CSI) scores for maximum 10-m wind speed (WS10M) at lead times of 1 to 5 days, as an indicator of extreme intensity forecasting skill. Panels (a--f) correspond to six increasing intensity thresholds ranging from 13.9 to 32.7 m/s. The bar groups illustrate the FuXi baseline alongside 3DTCR variants trained using different optimization strategies (Pre-training, SFT, and E2E) and multi-task loss settings (with or without MMD).
  • ...and 1 more figures