Coupled climate emulation with AI
HuggingFace Repository | Link to Paper Collaborators: James Duncan, Elynn Wu, Surya Dheeshjith, Adam Subel, Troy Arcomano, Spencer Clark, Brian Henn, Anna, Kwa, Jeremy McGibbon, Andre Perkins, Carlos Fernandez-Granda, Julius Busecke, Oli Watt-Meyer, Bill Hurlin, Alistair Adcroft, Laure Zanna, Chris Bretherton
Coupled climate models have been the cornerstone of Earth system research for decades and have allowed us to answer critical questions about how Earth’s climate responds to changes in greenhouse gas emissions and therefore what the future might look like under different warming scenarios. However, these simulations are expensive. The state-of-the-art GFDL climate model CM4 has a throughput of 16 SYPD (Simulated Year Per Day) on 5535 CPU cores. To accurately understand climate variability and change we need to run long simulations—the deep ocean takes hundreds to thousands of years to reach an equilibrium, or spun-up, state. A 100 year simulation of CM4 takes almost 1 week and requires massive resources, meaning that we often do not run simulations long enough to appropriately spin up the ocean. Recent work in the space of AI and climate has shown that AI models can skillfully simulate individual Earth system components. For example, ACE and Samudra are auto-regressive AI emulators of the full 3D atmosphere and ocean, respectively. When forced with observed boundary conditions in the style of AMIP and OMIP experiments, these emulators can integrate, or rollout, stably for decades, while accurately capturing the mean state and variability of the atmosphere and ocean. In this latest pre-print, led by James Duncan at the Allen Institute for AI (Ai2), we couple the 3D atmosphere and ocean emulators, ACE and Samudra, to produce the first coupled climate emulator, SamudrACE. SamudrACE is a 1-degree resolution coupled model that has been trained on a coarsened 200-year CM4 pre-industrial control simulation. SamudrACE is significantly faster than CM4, achieving 1500 SYPD on 1 H100 GPU. This speed makes it a fast and flexible tool for modelling long-term climate and large-ensemble analysis for studying extreme events. For more information on SamudrACE, check out our M$^2$LInES blog post!
Building a Hybrid (AI+physics) climate model
GitHub Repository Collaborators: Mitch Bushuk, Yongfei Zhang, Alistair Adcroft, Laure Zanna
Part I: Machine learning model error from data assimilation corrections
Climate models contain structural errors as a result of poorly parameterised or missing physics, as well as errors in the discretisation of continuous equations and errors in surface forcing. These structural errors lead to systematic biases in numerical simulations. For example, a climate model which systematically produces too much sea ice relative to a set of observations. Data Assimilation (DA) is a Bayesian framework which can reduce model biases by applying a correction, or increment, to the model state based on the current set of observations. These increments actually contain information about the systematic biases of a model. For example, if a model has a systematic positive bias, then the corrections generated from DA will be systematically negative (the DA is always trying to pull the model down from its positively biased state). We can therefore think of these increments as comprising some nonlinear combination of predictable model error growth associated with model bias, and an unpredictable component associated with short-term dynamics. In collaborative work with Princeton University, the Geophysical Fluid Dynamics Laboratory, and New York University, we investigated whether we could learn the predictable component of these increments using machine learning. To do this we trained convolutional neural networks to predict the increments based on the current state of the model (i.e. based on the current sea ice, ocean, and atmosphere conditions). This therefore gives rise to a state-dependent representation of the systematic component of model error. In our study published in the Journal of Advances in Modeling Earth Systems, we found that we can predict these increments very well in both the Arctic and Antarctic, and across all seasons. The figure on the right for example is a snapshot of the increments produced from DA (i.e. using observations), compared to the increments predicted by machine learning (i.e using only model state variables). The spatial pattern correlation (rho) is given between these two snapshots.
Based on this work, we suggested that this machine learning model could be used to bias correct numerical simulations during forward integration of the model, when we do not have observations—see Part II.
Part II: The first hybrid global sea ice model
In a follow-up Geophysical Research Letters study, we used the convolutional neural network which was trained to predict sea ice concentration DA increments, to do online bias correction in global ice-ocean simulations. A snapshot of the model error in summer is shown in the figures below for both the Arctic and Antarctic, where the observed ice edge is shown by the black contour. The colours then represent sea ice concentration errors relative to observations. We can see that the free-running model generally has too much sea ice in summer in both hemispheres—highlighted by the positive errors equator-ward of the observed ice edge contour. The simulation which assimilates observations (DA simulation) then reduces the ice-edge errors significantly. Impressively, the simulation which applies the machine learning correction (ML simulation) also significantly reduces the model errors, despite never seeing any observations—the corrections being applied during the model simulation are only a function of the model state variables themselves.
The results we showed in this study were for forced simulations, which is where the atmosphere is prescribed from reanalysis data. Therefore the ocean and sea ice are not impacting the atmosphere in any way. The next step in this work is therefore to assess how well the machine learning correction scheme does at improving real sea ice forecasts in a fully-coupled climate model—see Part III.
Part III: Improving global fully-coupled climate model forecasts
In our latest work (accepted at Science Advances), we used the convolutional neural network which was trained to predict sea ice concentration DA increments, to do online bias correction in a set of 1-year initialized sea ice forecasts, using the fully-coupled GFDL SPEAR model. The developments from stage II therefore include implementation in a climate model with interactive atmospheric feedbacks, as well as ML inference embedded into the Fortran source code of the GFDL sea ice model (see GitHub), rather than offline every 5 days.
The gif on the right shows the mean (2018–2024) bias of 1-year March-initialized sea ice forecasts. The sea ice concentration bias of the free-running SPEAR model on the top left and the hybrid SPEAR+ML model on the top right. The bottom panel then tracks the respective sea ice extent in each simulation. Here we can see how the machine learning model is able to significantly improve the seasonal forecast capabilities of SPEAR. In this work we also paid close attention to how different training regimes can impact online performance during fully-coupled simulations. To do this, we performed sea ice bias correction using the ML model which was developed in part II, which had a forced atmosphere. We then compared this to a ML model which was trained in a nudged configuration of the fully-coupled SPEAR model. While the nudged configuration still constrains the sea ice and will ultimately be different from the eventual free-running atmosphere of SPEAR, comparison of forced vs nudged atmosphere allows us to determine the importance of atmosphere-ice-ocean feedbacks in ML training data. We subsequently find them to be very important for online generalization. Training on data from a forced atmosphere and then implementing in a fully-coupled model causes runaway processes such as ice-free Antarctic summers. This happens by the ML model pre-conditioning the ocean and sea ice state in winter by triggering ocean convection, which brings heat to the surface and reduces winter ice growth rates. This pre-conditioning then triggers ice-albedo feedbacks in summer, facilitating rapid ice loss.
GPSat: Scalable Gaussian process interpolation
GitHub Repository | Link to Paper Collaborators: Ronald MacEachern, So Takao, Isobel Lawrence, Carmen Nab, Marc Deisenroth, Michel Tsamados
Satellite altimeters have been revolutionary in our understanding of sea ice, as they allow us to measure sea ice thickness directly from space. This has therefore given us a crucial insight into energy budgets in the Arctic, and the volumetric response of sea ice to global warming. Despite this, altimeters generally have narrow spatial footprints, on the order of 10s to 100s of metres. This means that on a given day, we may only sample a small fraction of a given spatial domain. For the CryoSat-2 radar altimeter, if data are binned to a 25 km grid, then it takes 30 days to uniformly sample the sea ice cover in the Arctic. At the footprint level it will take on the order of 1 year. These gaps in the data record make it difficult to understand how sea ice thickness changes on timescales ranging from days to weeks. Over the past few years I’ve been working with colleagues at University College London and the European Space Agency to develop a python programming library called ‘GPSat’ which attempts to fill these gaps in polar altimetry data. The idea behind GPSat is to use Gaussian Process (GP) models to learn how these observations co-vary in space and time, and to predict their values at unobserved locations. GP models are nice as they follow Bayesian principles; guiding our predictions based on our prior beliefs. The downside of GP models is that they are typically very expensive when dealing with large data sets. In this study in Nature Communications, we showed that our GPSat library can interpolate sea ice observations at relatively high spatial resoluton (5 km), and crucially, in affordable time. This is because we adopt a local approach, whereby we distribute GP models (we call them local experts) across the spatial domain (see image above), and then share information across these models when predicting at unobserved locations. We can also significantly speed up computations by leveraging GPU hardware.
In this work, we found that GPSat is over 500x faster than a naive implementation of GP models (see e.g., our earlier work). We therefore suggested that GPSat could overcome the computational bottlenecks faced in many altimetry-based interpolation routines, and hence advance critical understanding of ocean and sea ice variability over short spatio-temporal scales.
Connecting the polar climate system with complex networks
Part I: Ice-atmosphere teleconnections
GitHub Repository | Link to Paper Collaborators: Julienne Stroeve, Michel Tsamados
The representation of the climate system as a complex network is a useful framework with which to visualise and understand intrinsic climate variability, as well as coupled climate interactions, or teleconnections. During my PhD at University College London, I developed a workflow based on unsupervised cluster analysis for representing geospatial data sets as complex networks. The first step in this approach is to perform a clustering of grid points based on their temporal correlation structure. This can be seen in the figure on the right, in which grid points have been clustered together into spatially-contiguous geographic regions. For sea ice, these represent regions which have behaved ‘homogenously’ over the length of the time-series record—by homogeneous, we mean a high anomaly correlation coefficient of local sea ice area. In the context of a complex or climate network, each of these geographic areas corresponds to a node of the network, and links between nodes are given by the covariance between regions. Node strength is the sum of the absolute value of a node’s links. Therefore the node with the highest strength represents the hub of the network and can be considered as the leading mode of variability in the climate data set (analagous to Principal Component Analysis). In a publication in The Cryosphere with colleagues at University College London, we used this approach to investigate how CMIP6 models reflect the spatio-temporal variability of the winter Arctic Oscillation (AO; the leading mode of sea-level pressure variability), and how this drives summer sea ice. We found that, while models capture the patterns of AO variability relatively well, the effects of the AO on sea ice may be under-represented in models due to mean state biases in model’s sea ice thickness.
Part II: Data-driven sea ice forecasting
GitHub Repository | Link to Paper Collaborators: Michel Tsamados, Julienne Stroeve, Peter Sollich
The ability of complex networks to extract leading modes of climate variability and derive spatio-temporal connectivity structure, provides a good foundation for prediction. During my PhD at University College London, I used a combined approach of complex networks and Gaussian process models to do data-driven seasonal sea ice prediction of September Arctic sea ice extent. In our publication in Weather and Forecasting, we create sea ice area networks in the 1-3 months leading up to September, and use the connected nature of the network as the ‘prior’ in our Gaussian process model. In the figure below we can see the skill of this method in predicting monthly-mean pan-Arctic September sea ice extent anomalies at different lead times. Naturally, making predictions of mean September ice extent on September 1 produces the highest skill, and the skill then decreases as we increase the lead time. Notice for example how we get good skill in extreme years, such as 2012, as far back as July 1. Note that the metrics given in the figure are the Anomaly Correlation Coefficient, and the Mean-Squared Error Skill Score (Skill).
Every year since 2019 I have submitted real-time forecasts using this method to the Sea Ice Outlook (SIO), under the name ‘CPOM UCL (Gregory et al)’. The SIO solicits calls for predictions of the September Arctic sea ice cover each year, starting on June 1 through to September 1. In a recent co-authored study in the Bulletin of the American Meteorological Society, we performed a large inter-comparison of statistical and dynamical forecast models submitted to the SIO. This study emphasised the skill of many of these systems on seasonal timescales, particularly in extreme years. Our complex networks and Gaussian process approach was generally in the top 1-3 most skillful statistical models, in terms of detrended ACC of pan-Arctic sea ice extent.
