Harnessing Data-driven Neural Weather Models for Climate Attribution: A Case Study of the Oroville Dam Atmospheric River Episode of February 2017

In the last decades, meteorological forecasting has been dominated by Numerical Weather Prediction (NWP) models. NWP numerically solves a set of differential equations describing the dynamics of the atmosphere over a spatio-temporal 4D grid. Despite their impressive success in reproducing global meteorology, they require enormous computational resources to produce high-resolution ensemble forecasts. Recently, purely data-driven weather models have been proposed as an alternative, cost-effective solution to the computationally demanding NWP in the realm of meteorological forecasting. These tools train a machine learning model to fit a function between consecutive atmospheric states, typically using global reanalysis data (e.g., ERA5). Once trained, they perform inference in auto-regressive mode producing long rollouts. The very latest data-driven weather models developed based on deep learning —transformers [1] and graph neural networks [2],— have reported results which are comparable to the state-of-the art dynamical operational tools or NWP at short lead times (< 5 days), such as the IFS model, with a small fraction of the computational requirements. These benefits make data-driven tools attractive to be used for various meteorological applications, particularly where large ensembles are required.

Here we demonstrate the suitability of data-driven weather models for climate attribution tasks. We focus on the extreme atmospheric river event of February 2017, that caused flooding and severe impacts along the California coast, in particular damage to the Oroville Dam main and emergency spillway, prompting concerns of structural failure and leading to emergency evacuation [3]. This episode was also the subject of study in a recent climate attribution analysis with the MPAS-A dynamical model [4], and therefore allows for a proper comparison between studies. We find that ensemble forecasts based on our data-driven neural weather model accurately reproduce the atmospheric river intensity in the present climate. To assess the role of climate change on the event we follow [4] and design a “pseudo-global warming” experiment where the climate signal —as described by an ensemble of GCMs,— is added to the temperature fields of the initial condition. By feeding the data-driven neural model with these fields we find an increase in the intensity of the atmospheric river similar in magnitude with the one described by the dynamical model in [4]. Therefore, data-driven models lead to qualitatively and quantitatively comparable results with their counterpart dynamical model but with a considerable reduction in both computational times and resources needed (e.g., to produce a 10–day, 32-member ensemble takes only 1 minute on a cluster of 32 P100 GPUs, while IFS requires about 1 hour). These findings provide promising avenues for application of data-driven AI tools that can accelerate the climate attribution of extreme events across the world, potentially providing real-time attribution while public attention is captured.

[1] Pathak, J., Subramanian, S., Harrington, P., Raja, S., Chattopadhyay, A., Mardani, M., ... & Anandkumar, A. (2022). Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214.

[2] Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Pritzel, A., ... & Battaglia, P. (2022). GraphCast: Learning skillful medium-range global weather forecasting. arXiv preprint arXiv:2212.12794.

[3] Henn, B., Musselman, K. N., Lestak, L., Ralph, F. M., & Molotch, N. P. (2020). Extreme runoff generation from atmospheric river driven snow melt during the 2017 Oroville Dam spillways incident. Geophysical Research Letters, 47, e2020GL088189

[4] Michaelis, A. C., Gershunov, A., Weyant, A., Fish, M. A., Shulgina, T., & Ralph, F. M. (2022). Atmospheric river precipitation enhanced by climate change: A case study of the storm that contributed to California's Oroville Dam crisis. Earth's Future, 10(3), e2021EF002537.