Data Assimilation with Machine Learning Surrogate Models - A Case Study with FourCastNet

Aug 5, 2026

Upload by Oliver

Modern data-driven surrogate models can give accurate short-term weather predictions but often produce inaccurate, nonphysical long-term forecasts. This study examines online weather prediction with machine-learning surrogates augmented by partial, noisy observations. The work shows filtering estimates can stay accurate over long horizons even when surrogate dynamics become unstable and measurements are sparse. As an applied case study, FourCastNet is integrated into a variational data assimilation framework using partial noisy ERA5 data, yielding year-long accuracy and strong initial conditions for forecasting tasks, including extreme-event prediction.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages23

Reading Time58 min

AccessFree to download

Introduction

Numerical weather prediction and computational cost

Weather surrogate models and practical observation limits

Data assimilation background and motivation

Extreme event prediction and operational relevance

Case study setup and objectives

Variational assimilation with partial noisy ERA5 data

FourCastNet as a surrogate model

Evaluation goals for forecasting and extremes

arXiv :2405 . 13180v1 [ ee ss . SP] 21 May 2024 Data Assimilation with Machine Learning Surrogate Models: A Case Study with FourCastNet Melissa Adrian∗1, Daniel Sanz-Alonso 1 , and Rebecca Willett 1,2 1 Department of Statistics, The University of Chicago, Chicago, IL 60637, USA 2 Department of Computer Science, The University of Chicago, Chicago, IL 60637, USA Abstract Modern data-driven surrogate models for weather forecasting provide accurate short-term predictions but inaccurate and nonphysical long-term forecasts. This paper investigates online weather prediction using machine learning surrogates supplemented with partial and noisy observations. We empirically demonstrate and theoretically justify that, despite the long-time instability of the surrogates and the sparsity of the observations, filtering estimates can remain accurate in the long-time horizon. As a case study, we integrate FourCastNet, a state-of-the-art weather surrogate model, within a variational data assimilation framework using partial, noisy ERA5 data. Our results show that filtering estimates remain accurate over a year-long assimilation window and provide effective initial conditions for forecasting tasks, including extreme event prediction. 1 Introduction Numerical weather prediction (NWP) at an operational scale relies on large-scale systems of partial differential equations to model atmospheric dynamics. However, these physics-based models are computationally expensive to simulate, particularly when operating at a high resolution. This computational burden plagues both weather and climate models alike [Tollefson, 2023], leading research focus to shift towards cheaper alternatives: data-driven machine learning surrogate models for weather forecasting. Weather surrogate models have been analyzed and evaluated extensively using high-resolution accurate datasets, usually ERA5 reanalysis data [Hersbach et al. , 2020] . However, in practical weather forecasting settings, we must provide high-fidelity forecasts given only sparse observations, often contaminated with measurement errors. Because major advances in high-resolution data-driven global weather modeling have only been made in the past few years, substantial work has yet to be done to analyze its utility in settings of operational interest, including both data assimilation and forecasting with these sparse and noisy observations. Data assimilation is an operational task with a long history in weather forecasting, rooted in seminal works, such as Richardson [1922], Kalman [1960], and Gandin [1966] . Preliminary data assimilation methods were developed specifically for the vast problem that characterizes weather settings: inferring a high-resolution representation of the atmosphere with only (1) sparse, noisy measurements throughout the globe and (2) a time evolution mapping of atmospheric states. Data assimilation produces high-dimensional representations, also referred to as analyses, which provide a detailed view of historical global weather patterns. These analyses can be used for numerous purposes, most notably to initialize forecasts based on current and historical observations. However, producing analyses using physics-based NWP models is computationally expensive, especially for longtime horizons. Consequently, cheap-to-evaluate surrogate weather models have enormous potential to expedite this data assimilation process. The potential of weather surrogate models to accelerate extreme event prediction has recently received attention at the government level, with an April 2024 U.S. Executive Office report calling attention to its potential widespread operational use [The President’s Council of Advisors on Science and Technology, 2024] . This report stated that in the U.S. in 2023 alone, the economic damage due to extreme weather events totalled to $92.9 billion from 28 weather disasters, and the frequency of these extreme events is expected to continue to increase in the co

FAQs

Why do machine learning weather surrogate models struggle in long-term forecasts?

They can become unstable over long horizons and may generate inaccurate, nonphysical long-term trajectories, despite good short-term predictive performance.

How does the paper keep filtering estimates accurate over long time horizons?

It combines machine-learning surrogates with partial and noisy observations within a filtering/variational data assimilation framework, and provides both empirical and theoretical justification for long-horizon accuracy under sparsity.

What is the role of FourCastNet in the case study?

FourCastNet is embedded into a variational data assimilation pipeline using partial, noisy ERA5 observations, producing analyses that support year-long assimilation accuracy and improved forecasting initialization, including for extreme events.

Aug 5, 2026

Upload by Oliver