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

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 the Fourier Forecasting Neural Network (FourCastNet), a weather surrogate model, within a variational data assimilation framework using partial, noisy ERA5 global reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF). 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. Significance Statement Data-driven weather models have recently shown substantial success in acting as surrogates of physics-based numerical weather prediction models. However, this success has been measured using high-resolution reanalysis data, which are not available in real time. This work assesses data-driven weather models’ use in more practical settings, where we observe noisy, partial measurements of a high-dimensional weather state across time. Our experiments show that integrating FourCastNet into a data assimilation algorithm, 3DVar, using low-resolution observations, allows us to 1) accurately estimate high-dimensional weather states across a long time horizon and 2) effectively use these estimates for initializing forecasts, particularly for 3) the extreme event Typhoon Mawar in 2023. This work provides evidence that data-driven weather models can successfully perform in realistic settings.