[Paper Review] FuXi: A cascade machine learning forecasting system for 15-day global weather forecast

Lei Chen, Xiaohui Zhong+5 more|arXiv (Cornell University)|Jun 22, 2023

Climate variability and models18 citations

TL;DR

FuXi is a cascaded, pre-trained ML weather forecasting system that delivers 15-day global forecasts at 0.25° resolution, achieving ECMWF ensemble-like performance and extending skillful lead times for Z500 and T2M.

ABSTRACT

Over the past few years, due to the rapid development of machine learning (ML) models for weather forecasting, state-of-the-art ML models have shown superior performance compared to the European Centre for Medium-Range Weather Forecasts (ECMWF)'s high-resolution forecast (HRES) in 10-day forecasts at a spatial resolution of 0.25 degree. However, the challenge remains to perform comparably to the ECMWF ensemble mean (EM) in 15-day forecasts. Previous studies have demonstrated the importance of mitigating the accumulation of forecast errors for effective long-term forecasts. Despite numerous efforts to reduce accumulation errors, including autoregressive multi-time step loss, using a single model is found to be insufficient to achieve optimal performance in both short and long lead times. Therefore, we present FuXi, a cascaded ML weather forecasting system that provides 15-day global forecasts with a temporal resolution of 6 hours and a spatial resolution of 0.25 degree. FuXi is developed using 39 years of the ECMWF ERA5 reanalysis dataset. The performance evaluation, based on latitude-weighted root mean square error (RMSE) and anomaly correlation coefficient (ACC), demonstrates that FuXi has comparable forecast performance to ECMWF EM in 15-day forecasts, making FuXi the first ML-based weather forecasting system to accomplish this achievement.

Motivation & Objective

• Motivate the use of cascaded machine learning models to reduce accumulation errors in long-range weather forecasts.
• Develop FuXi, a cascade ML forecast system producing 15-day forecasts at 0.25° resolution using ERA5 data.
• Show that FuXi can match ECMWF ensemble performance and extend skillful lead times for key variables.
• Demonstrate how FuXi ensemble forecasts provide uncertainty estimates comparable to ECMWF ensemble.

Proposed method

• Use 39 years of ERA5 reanalysis data at 0.25° and 6-hourly cadence to train FuXi.
• Construct FuXi as a cascade of three specialized models: FuXi-Short (0–5 days), FuXi-Medium (5–10 days), FuXi-Long (10–15 days).
• Base FuXi architecture combines cube embedding, U-Transformer with 48 Swin Transformer V2 blocks, and a final FC layer.
• Train via two-stage process: one-step pre-training (predict single 6-hour step) and fine-tuning of cascaded models with autoregressive curriculum.
• Improve stability and reduce accumulation error using autoregressive multi-step loss during training.
• Create a 50-member FuXi ensemble by perturbing initial conditions with Perlin noise and applying Monte Carlo dropout.

Figure 1 : Overall architecture of FuXi model. a) The FuXi model consists of three components: cube embedding, U-Transformer, and fully connected (FC) layer; b) FuXi-Short, FuXi-Medium, and FuXi-Long models cascade and produce 15-day forecasts, with each model generating 5 days forecasts.

Experimental results

Research questions

• RQ1 Can a cascaded ML architecture reduce long-lead forecast accumulation errors and match ECMWF ensemble performance for 15-day forecasts?
• RQ2 What is the forecast lead time extension achievable for Z500 and T2M compared to ECMWF HRES?
• RQ3 Does a FuXi ensemble provide competitive forecast uncertainty relative to the ECMWF ensemble?
• RQ4 How does FuXi perform relative to deterministic benchmarks (HRES, EM) across multiple variables and lead times?

Key findings

• FuXi's cascaded architecture yields 15-day forecasts at 0.25° that are comparable to the ECMWF ensemble mean (EM).
• FuXi extends skillful ACC (ACC > 0.6) lead times to 10.5 days for Z500 and 14.5 days for T2M.
• FuXi outperforms ECMWF HRES in 15-day forecasts across multiple variables, with ACC and RMSE advantages increasing with lead time.
• FuXi ensemble forecasts achieve CRPS performance comparable to ECMWF ensemble within 9 days for several fields, indicating useful uncertainty estimates.

Figure 2 : Comparison of the globally-averaged latitude-weighted ACC (first and second rows) and RMSE (third and fourth rows) of the HRES (dark green lines), GraphCast (organge lines), and FuXi (light blue lines) for 4 surface variables, such as ${MSL}$ , ${T2M}$ , ${U10}$ , and ${V10}$ , and 4 uppe