Description: Earth system models (ESMs) are our main tools for quantifying the physical state of the Earth and predicting how it might change in the future under ongoing anthropogenic forcing. In recent years, however, artificial intelligence (AI) methods have been increasingly used to augment or even replace classical ESM tasks, raising hopes that AI could solve some of the grand challenges of climate science. In this Perspective we survey the recent achievements and limitations of both process-based models and AI in Earth system and climate research, and propose a methodological transformation in which deep neural networks and ESMs are dismantled as individual approaches and reassembled as learning, self-validating and interpretable ESM–network hybrids. Following this path, we coin the term neural Earth system modelling. We examine the concurrent potential and pitfalls of neural Earth system modelling and discuss the open question of whether AI can bolster ESMs or even ultimately render them obsolete.

Description: Quantifying and monitoring terrestrial water storage (TWS) is an essential task for understanding the Earth's hydrosphere cycle, its susceptibility to climate change, and concurrent impacts for ecosystems, agriculture, and water management. Changes in TWS manifest as anomalies in the Earth's gravity field, which are routinely observed from space. However, the complex underlying distribution of water masses in rivers, lakes, or groundwater basins remains elusive. We combine machine learning, numerical modeling, and satellite altimetry to build a downscaling neural network that recovers simulated TWS from synthetic space‐borne gravity observations. A novel constrained training is introduced, allowing the neural network to validate its training progress with independent satellite altimetry records. We show that the neural network can accurately derive the TWS in 2019 after being trained over the years 2003 to 2018. Further, we demonstrate that the constrained neural network can outperform the numerical model in validated regions.

Description: In Irrgang et al. (2020), we have trained a convolutional neural network to perform a so-called downscaling task. This downscaling aims to recover the fine-structure continental water storage distribution on the South American continent from coarse-resolution space-borne gravimetry observations. Here, we share data sets that were used for training the neural network, namely (1) monthly pairs of gridded terrestrial water storage anomalies (TWSA) of the South American continent and (2) surface water storage anomalies (SWSA) in the Amazonas region for the time period 2003-2019. TWSAs were used as target (output) values of the neural network and were derived from the Land Surface Discharge Model (LSDM, Dill, 2008). The corresponding input values were calculated by spatially smoothing the TWSA fields with a 600 km Gaussian filter. After training the neural network over the time period of 2003 to 2018, its performance was tested and compared to LSDM for the subsequent year 2019.

Description: The characterization of uncertainties in geophysical quantities is an important task with widespread applications for time series prediction, numerical modeling, and data assimilation. In this context, machine learning is a powerful tool for estimating complex patterns and their evolution through time. Here, we utilize a supervised machine learning approach to dynamically predict the spatiotemporal uncertainty of near‐surface wind velocities over the ocean. A recurrent neural network (RNN) is trained with reanalyzed 10 m wind velocities and corresponding precalculated uncertainty estimates during the 2012–2016 time period. Afterward, the neural network's performance is examined by analyzing its prediction for the subsequent year 2017. Our experiments show that a recurrent neural network can capture the globally prevalent wind regimes without prior knowledge about underlying physics and learn to derive wind velocity uncertainty estimates that are only based on wind velocity trajectories. At single training locations, the RNN‐based wind uncertainties closely match with the true reference values, and the corresponding intra‐annual variations are reproduced with high accuracy. Moreover, the neural network can predict global lateral distribution of uncertainties with small mismatch values after being trained only at a few isolated locations in different dynamic regimes. The presented approach can be combined with numerical models for a cost‐efficient generation of ensemble simulations or with ensemble‐based data assimilation to sample and predict dynamically consistent error covariance information of atmospheric boundary forcings.

Description: We suggest to apply data assimilation in glacial isostatic adjustment (GIA) to constrain the mantle viscosity structure based on sea level observations. We apply the Parallel Data Assimilation Framework (PDAF) to assimilate sea level data into the time-domain spectral-finite element code VILMA in order to obtain better estimates of the mantle viscosity structure. In a first step, we reduce to a spherically symmetric earth structure and prescribe the glaciation history. A particle filter is used to propagate an ensemble of models in time. At epochs when observations are available, each particle's performance is estimated and the particles are resampled based on their performance to form a new ensemble that better resembles the true viscosity distribution. Using this algorithm, we show the ability to recover mantle viscosities from a set of synthetic relative sea level observations. Those synthetic observations are obtained from a reference run with a given viscosity structure that defines the target viscosity values in our experiments. The viscosity estimation is applied to a three-layer model with an elastic lithosphere and two mantle layers, and to a multi-layer model with a smoother viscosity profile. We use various subsets of realistic observation locations (e.g. only observations from Fennoscandia) and show that it is possible to obtain the target viscosity values in those cases. We also vary the time from which observations are available to evolve the test cases towards a realistic scenario for the availability of relative sea level observations. The most relevant cases start at 26.5ka BP and at 10ka BP as they mark the beginning of the maximum glaciation and the end of deglaciation with a larger amount of observations following, respectively, and end at present day.

Description: This is a synthetic dataset. It was created from the outputs of the glacial isostatic adjustment model VILMA (Klemann et al. 2008). It consists of realtive sea level (RSL) data on a global regular grid. The resolution is 256 x 512 points (Lat x Lon). The tomporal range is from 123 ka BP until present day. Time steps vary between 2.5 kyrs at the beginning and 0.5 kyrs towards the end. The data were created for a specific configuration of the GIA model: lithosphere thickness = 60 km, lithosphere viscosity = 1.0E31 Pa s, upper mantle thickness = 610 km, upper mantle viscosity = 1.0E20 Pa s, lower mantle thickness = 3,221 km, lower mantle viscosity = 1.0E21 Pa s. The RSL data are accompanied by a observation locations mask. This mask was used to identify those locations in the global RSL dataset where real observations are available. The dataset consists of realtive sea level (RSL) data on a global regular grid. The resolution is 256 x 512 points (Lat x Lon). The temporal range is from 123 ka BP until present day. Time steps vary between 2.5 kyrs at the beginning and 0.5 kyrs towards the end. The data were created for a specific configuration of the GIA model: lithosphere thickness = 60 km, lithosphere viscosity = 1.0E31 Pa s, upper mantle thickness = 610 km, upper mantle viscosity = 1.0E20 Pa s, lower mantle thickness = 3,221 km, lower mantle viscosity = 1.0E21 Pa s. The RSL data are accompanied by observation locations masks. These masks were used to mark those locations in the global RSL dataset where real-life observations are available in order to restrict usage of the synthetic data to those locations.

Description: Ocean sediment drifts contain important information about past bottom currents but a direct link between the study of sedimentary archives and ocean dynamics is not always possible. To close this gap for the North Atlantic, we set up a new coupled Ice-Ocean-Sediment Model of the N. Atlantic - Arctic region. In order to evaluate the potential dynamics of the model, we conducted decadal sensitivity experiments. In our model contouritic sedimentation shows a significant sensitivity towards climate variability for most of the contourite drift locations in the model domain. We observe a general decrease of sedimentation rates during warm conditions with decreasing atmospheric and oceanic gradients and an extensive increase of sedimentation rates during cold conditions with respective increased gradients. We can relate these results to changes in the dominant bottom circulation supplying deep water masses to the contourite sites under different climate conditions. A better understanding of northern deep water pathways in the Atlantic Meridional Overturning Circulation (AMOC) is crucial for evaluating possible consequences of climate change in the ocean.

Description: Global ocean circulation generates large-scale secondary electromagnetic signals, which provide a potentially interesting source of information about the oceanographic system. Radial magnetic fields caused by the movement of conductive seawater through Earth's magnetic field are, in principle, observable in geomagnetic satellite observations such as magnetometer data from the current Swarm mission. However, contrary to magnetic signals induced by tides, the signals resulting from ocean circulation have not been identified yet. Recently, the detectability of these ocean circulation-induced signals was investigated within an observing system simulation experiment using a Kalman filter-based approach. In this approach, the crucial separation from other magnetic contributions is achieved by predefining the temporal behavior of the oceanic component using presumed estimates. We applied this approach to real Swarm magnetometer observations and determined a scale factor for various a priori assumed ocean circulation-induced magnetic fields through the assimilation. Furthermore, we evaluated the outcomes in terms of identifying magnetic signals caused by ocean circulation for the first time in geomagnetic satellite observations. We present and discuss the evaluation of these results and report on the detectability of magnetic fields induced by ocean circulation in Swarm satellite observations.

Description: The tides are a major driver of global oceanic mixing. While global tidal elevations are very well observed by satellite altimetry, the global tidal transports are much less well known. Since twenty years, magnetic signals induced by the ocean tide are detectable in satellite magnetometer observations as Swarm and CHAMP. Here, we successfully use said satellite magnetometer observations to directly derive global ocean tidal transports.