Search

General circulation models (GCMs) are the foundation of weather and climate prediction 1,2 . GCMs are physics-based simulators that combine a numerical solver for large-scale dynamics with tuned representations for small-scale processes such as cloud formation. Recently, machine-learning models trained on reanalysis data have achieved comparable or better skill than GCMs for deterministic weather forecasting 3,4 . However, these models have not demonstrated improved ensemble forecasts, or shown sufficient stability for long-term weather and climate simulations. Here we present a GCM that combines a differentiable solver for atmospheric dynamics with machine-learning components and show that it can generate forecasts of deterministic weather, ensemble weather and climate on par with the best machine-learning and physics-based methods. NeuralGCM is competitive with machine-learning models for one- to ten-day forecasts, and with the European Centre for Medium-Range Weather Forecasts ensemble prediction for one- to fifteen-day forecasts. With prescribed sea surface temperature, NeuralGCM can accurately track climate metrics for multiple decades, and climate forecasts with 140-kilometre resolution show emergent phenomena such as realistic frequency and trajectories of tropical cyclones. For both weather and climate, our approach offers orders of magnitude computational savings over conventional GCMs, although our model does not extrapolate to substantially different future climates. Our results show that end-to-end deep learning is compatible with tasks performed by conventional GCMs and can enhance the large-scale physical simulations that are essential for understanding and predicting the Earth system., (© 2024. The Author(s).)

Projecting climate change is a generalization problem: We extrapolate the recent past using physical models across past, present, and future climates. Current climate models require representations of processes that occur at scales smaller than model grid size, which have been the main source of model projection uncertainty. Recent machine learning (ML) algorithms hold promise to improve such process representations but tend to extrapolate poorly to climate regimes that they were not trained on. To get the best of the physical and statistical worlds, we propose a framework, termed "climate-invariant" ML, incorporating knowledge of climate processes into ML algorithms, and show that it can maintain high offline accuracy across a wide range of climate conditions and configurations in three distinct atmospheric models. Our results suggest that explicitly incorporating physical knowledge into data-driven models of Earth system processes can improve their consistency, data efficiency, and generalizability across climate regimes.

By accounting for most of the poleward atmospheric heat and moisture transport in the tropics, the Hadley circulation largely affects the latitudinal patterns of precipitation and temperature at low latitudes. To increase our preparednesses for human-induced climate change, it is thus critical to accurately assess the response of the Hadley circulation to anthropogenic emissions 1-3 . However, at present, there is a large uncertainty in recent Northern Hemisphere Hadley circulation strength changes 4 . Not only do climate models simulate a weakening of the circulation 5 , whereas atmospheric reanalyses mostly show an intensification of the circulation 4-8 , but atmospheric reanalyses were found to have artificial biases in the strength of the circulation 5 , resulting in unknown impacts of human emissions on recent Hadley circulation changes. Here we constrain the recent changes in the Hadley circulation using sea-level pressure measurements and show that, in agreement with the latest suite of climate models, the circulation has considerably weakened over recent decades. We further show that the weakening of the circulation is attributable to anthropogenic emissions, which increases our confidence in human-induced tropical climate change projections. Given the large climate impacts of the circulation at low latitudes, the recent human-induced weakening of the flow suggests wider consequences for the regional tropical-subtropical climate., (© 2023. The Author(s), under exclusive licence to Springer Nature Limited.)

Cardiovascular disease (CVD) prediction models are widely used in modern medicine and are incorporated into prominent guidelines. Coronary artery calcium (CAC) is a marker of coronary atherosclerotic disease and has proven utility for predicting cardiovascular disease. Despite this, current guidelines recommend against including CAC scores in CVD prediction models due to the medical and financial costs of acquiring it, and the insufficient evidence concerning its ability to improve existing models. Modern machine learning models are capable of automatically extracting coronary calcium scores from existing chest computed tomography (CT) scans, negating these costs. To determine whether the inclusion of CAC scores, automatically extracted using a machine learning algorithm from chest CTs performed for any reason, improves the performance of the American Heart Association/American College of Cardiology 2013 pooled cohort equations (PCE). A retrospective cohort of patients with available chest CTs prior to an index date (2012) was used to compare the performance of the PCE model and an augmented-PCE model which utilizes the CT-based CAC scores on top of the existing model. The PCE and the augmented-PCE predictions were calculated as of an index date (2012) using data from the electronic health record and existing chest CTs. The performance of both models was evaluated by comparing their predictions to cardiovascular events that occurred during a 5-year follow-up period (until 2017). A total of 14,135 patients aged 40-79 years were included in the study, of whom 470 (3.3%) had documented CVD events during the follow-up. The augmented-PCE model showed a significant improvement in c-statistic (0.64 ≥ 0.69, Δ = 0.05, 95% CI: 0.03 to 0.06), sensitivity (53% ≥ 57%, Δ = 4.7%, 95% CI: 0-9.0%), specificity (67% ≥ 70%, Δ = 2.8%, 95% CI: 0.9-5.1%), in positive predictive value (5% ≥ 6%, Δ = 0.9%, 95% CI: 0.4 to 1.4%), negative predictive value (97.7% ≥ 97.9%, Δ = 0.3%, 95% CI: 0.1 to 0.5%), and in the categorical net reclassification index (7.4%, 95% CI: 2.4 to 12.1%). Automatically generated CAC scores from existing CTs can aid in CVD risk determination, improving model performance when used on top of existing predictors. Use of existing CTs avoids most pitfalls currently cited against the routine use of CAC in CVD predictions (e.g., additional radiation exposure), and thus affords a net gain in predictive accuracy., (© 2022. The Author(s) under exclusive licence to Society for Imaging Informatics in Medicine.)

Global climate models represent small-scale processes such as convection using subgrid models known as parameterizations, and these parameterizations contribute substantially to uncertainty in climate projections. Machine learning of new parameterizations from high-resolution model output is a promising approach, but such parameterizations have been prone to issues of instability and climate drift, and their performance for different grid spacings has not yet been investigated. Here we use a random forest to learn a parameterization from coarse-grained output of a three-dimensional high-resolution idealized atmospheric model. The parameterization leads to stable simulations at coarse resolution that replicate the climate of the high-resolution simulation. Retraining for different coarse-graining factors shows the parameterization performs best at smaller horizontal grid spacings. Our results yield insights into parameterization performance across length scales, and they also demonstrate the potential for learning parameterizations from global high-resolution simulations that are now emerging.

Cells probe their mechanical environment and can change the organization of their cytoskeletons when the elastic and viscous properties of their environment are modified. We use a model in which the forces exerted by small, contractile acto-myosin filaments (e.g., nascent stress fibers in stem cells) on the extracellular matrix are modeled as local force dipoles. In some cases, the strain field caused by these force dipoles propagates quickly enough so that only static elastic interactions need be considered. On the other hand, in the case of significant energy dissipation, strain propagation is slower and may be eliminated completely by the relaxation of the cellular cytoskeleton (e.g., by cross-link dissociation). Here, we consider several dissipative mechanisms that affect the propagation of the strain field in adhered cells and consider these effects on the interaction between force dipoles and their resulting mutual orientations. This is a first step in understanding the development of orientational (nematic) or layering (smectic) order in the cytoskeleton. We use the theory to estimate the propagation time of the strain fields over a cellular distance for different mechanisms and find that in some cases it can be of the order of seconds, thus competing with the cytoskeletal relaxation time. Furthermore, for a simple system of two force dipoles, we predict that in some cases the orientation of force dipoles might change significantly with time, e.g., for short times the dipoles exhibit parallel alignment while for later times they align perpendicularly.