[Paper Review] FuXi-Extreme: Improving extreme rainfall and wind forecasts with diffusion model

Xiaohui Zhong, Lei Chen+4 more|arXiv (Cornell University)|Oct 25, 2023

Climate variability and modelsEnvironmental Science10 citations

TL;DR

FuXi-Extreme uses a denoising diffusion probabilistic model to recover fine-scale details in FuXi’s 5-day surface forecasts, improving extreme rainfall and wind predictions compared to FuXi and HRES.

ABSTRACT

Significant advancements in the development of machine learning (ML) models for weather forecasting have produced remarkable results. State-of-the-art ML-based weather forecast models, such as FuXi, have demonstrated superior statistical forecast performance in comparison to the high-resolution forecasts (HRES) of the European Centre for Medium-Range Weather Forecasts (ECMWF). However, ML models face a common challenge: as forecast lead times increase, they tend to generate increasingly smooth predictions, leading to an underestimation of the intensity of extreme weather events. To address this challenge, we developed the FuXi-Extreme model, which employs a denoising diffusion probabilistic model (DDPM) to restore finer-scale details in the surface forecast data generated by the FuXi model in 5-day forecasts. An evaluation of extreme total precipitation ($ extrm{TP}$), 10-meter wind speed ($ extrm{WS10}$), and 2-meter temperature ($ extrm{T2M}$) illustrates the superior performance of FuXi-Extreme over both FuXi and HRES. Moreover, when evaluating tropical cyclone (TC) forecasts based on International Best Track Archive for Climate Stewardship (IBTrACS) dataset, both FuXi and FuXi-Extreme shows superior performance in TC track forecasts compared to HRES, but they show inferior performance in TC intensity forecasts in comparison to HRES.

Motivation & Objective

• Motivate improved extreme-event forecasting under climate change where ML models tend to oversmooth predictions.
• Develop FuXi-Extreme by coupling FuXi with a denoising diffusion probabilistic model (DDPM) to restore fine-scale surface features.
• Evaluate extreme precipitation, wind speed, and temperature against ERA5 and benchmark tropical cyclone forecasts against IBTrACS.
• Assess the trade-offs in TC track and intensity forecasts among FuXi, FuXi-Extreme, and HRES.
• Identify limitations and propose directions for extending FuXi-Extreme to upper-air variables and higher-resolution data.

Proposed method

• FuXi-Extreme combines a frozen FuXi base model with a DDPM that denoises FuXi outputs to recover fine-scale extremes.
• The DDPM is conditioned on the FuXi time step and predicts the original target values directly using an MSE loss.
• The forward diffusion adds Gaussian noise to the ground-truth surface variables; the reverse process learns to denoise conditioned on FuXi predictions.
• Training updates only the DDPM parameters while keeping FuXi fixed, using AdamW with scheduled DropPath on multi-GPU hardware.
• Evaluation uses CSI and SEDI metrics for extreme events and TC track/intensity measures against IBTrACS and ERA5 ground truth.

Experimental results

Research questions

• RQ1 Can a diffusion-based denoiser (DDPM) enhance extreme-value accuracy in FuXi-generated surface forecasts over 5 days?
• RQ2 How does FuXi-Extreme compare to FuXi and ECMWF HRES in predicting extreme TP, WS10, and T2M?
• RQ3 What is the impact of FuXi-Extreme on tropical cyclone track and intensity forecasts relative to HRES and FuXi?
• RQ4 Do evaluations based on ERA5 ground truth and IBTrACS trajectories yield consistent conclusions about model skill for tracks and intensities?

Key findings

• FuXi-Extreme achieves the highest CSI scores for extreme WS10 and TP across 5-day forecasts compared to FuXi and HRES.
• SEDI analyses show FuXi-Extreme outperforms FuXi and HRES at higher percentiles for T2M, TP, and WS10.
• For TC tracks, FuXi and FuXi-Extreme outperform HRES in 0-2 day forecasts and continue to outperform for longer lead times in track prediction.
• In TC intensity, HRES yields the best RMSE for IBTrACS-based evaluations, while FuXi and FuXi-Extreme show improved RMSE against ERA5 ground truth.
• FuXi-Extreme’s improvements are most pronounced for surface-extreme variables, while upper-air variables remain shared with FuXi in forecasts.