Abstract Model instability remains a core challenge for data‐driven parameterizations, especially those developed with supervised algorithms, and rigorous methods to address it are lacking. Here, by integrating machine learning (ML) theory with climate physics, we demonstrate the importance of learning spatially non‐local dynamics using a 1D quasi‐biennial oscillation model with parameterized gravity waves (GW) as a testbed. While common offline metrics fail to identify shortcomings in learning non‐local dynamics, we show that the receptive field (RF) can identify instability a‐priori. We find that neural network‐based parameterizations, though predicting GW forcings from wind profiles with 99% accuracy, lead to unstable simulations when RFs are too small to capture non‐local dynamics. Additionally, we demonstrate that learning non‐local dynamics is crucial for the stability of a data‐driven spatiotemporal emulator of the zonal wind field. This work underscores the need to integrate ML theory with physics in designing data‐driven algorithms for climate modeling.

Predicting gray swan weather extremes, which are possible but so rare that they are absent from the training dataset, is a major concern for AI weather models and long-term climate emulators. An important open question is whether AI models can extrapolate from weaker weather events present in the training set to stronger, unseen weather extremes. To test this, we train independent versions of the AI weather model FourCastNet on the 1979–2015 ERA5 dataset with all data, or with Category 3–5 tropical cyclones (TCs) removed, either globally or only over the North Atlantic or Western Pacific basin. We then test these versions of FourCastNet on 2018–2023 Category 5 TCs (gray swans). All versions yield similar accuracy for global weather, but the one trained without Category 3–5 TCs cannot accurately forecast Category 5 TCs, indicating that these models cannot extrapolate from weaker storms. The versions trained without Category 3–5 TCs in one basin show some skill forecasting Category 5 TCs in that basin, suggesting that FourCastNet can generalize across tropical basins. This is encouraging and surprising because regional information is implicitly encoded in inputs. Given that current state-of-the-art AI weather and climate models have similar learning strategies, we expect our findings to apply to other models. Other types of weather extremes need to be similarly investigated. Our work demonstrates that novel learning strategies are needed for AI models to reliably provide early warning or estimated statistics for the rarest, most impactful TCs, and, possibly, other weather extremes.

Climate science has been revolutionized by the combined effects of an exponential growth in computing power, which has enabled more sophisticated and higher-resolution simulations to be made of the climate system, and an exponential increase in observations since the first weather satellite was put in orbit. Big data and associated algorithms, coalesced under the field of machine learning (ML), offer the opportunity to study the physics of the climate system in ways, and with an amount of detail, that were previously infeasible. Additionally, ML can ask causal questions to determine whether one or more variables cause or affect one or more outcomes and improve prediction skills beyond classical limits. Furthermore, when paired with modelling experiments or robust research on model parameterizations, ML can accelerate computations, increasing accuracy and generating very large ensembles with a fraction of the computational cost of traditional systems. In this Review, we outline the accomplishments of ML in climate physics. We discuss how ML has been used to tackle long-standing problems in the reconstruction of observational data, representation of sub-grid-scale phenomena and climate (and weather) prediction. Finally, we consider the benefits and major challenges of exploiting ML in studying complex systems.

Abstract Neural networks (NNs) are increasingly used for data‐driven subgrid‐scale parameterizations in weather and climate models. While NNs are powerful tools for learning complex non‐linear relationships from data, there are several challenges in using them for parameterizations. Three of these challenges are (a) data imbalance related to learning rare, often large‐amplitude, samples; (b) uncertainty quantification (UQ) of the predictions to provide an accuracy indicator; and (c) generalization to other climates, for example, those with different radiative forcings. Here, we examine the performance of methods for addressing these challenges using NN‐based emulators of the Whole Atmosphere Community Climate Model (WACCM) physics‐based gravity wave (GW) parameterizations as a test case. WACCM has complex, state‐of‐the‐art parameterizations for orography‐, convection‐, and front‐driven GWs. Convection‐ and orography‐driven GWs have significant data imbalance due to the absence of convection or orography in most grid points. We address data imbalance using resampling and/or weighted loss functions, enabling the successful emulation of parameterizations for all three sources. We demonstrate that three UQ methods (Bayesian NNs, variational auto‐encoders, and dropouts) provide ensemble spreads that correspond to accuracy during testing, offering criteria for identifying when an NN gives inaccurate predictions. Finally, we show that the accuracy of these NNs decreases for a warmer climate (4 × CO 2 ). However, their performance is significantly improved by applying transfer learning, for example, re‐training only one layer using ∼1% new data from the warmer climate. The findings of this study offer insights for developing reliable and generalizable data‐driven parameterizations for various processes, including (but not limited to) GWs. , Plain Language Summary Scientists increasingly use machine learning methods, especially neural networks (NNs), to improve weather and climate models. However, it can be challenging for an NN to learn rare, large‐amplitude events because they are infrequent in training data. In addition, NNs need to express their confidence (certainty) about a prediction and work effectively across different climates, for example, warmer climates due to increased CO 2 . Traditional NNs often struggle with these challenges. Here, we share insights from emulating known physics (gravity waves) with NNs in a state‐of‐the‐art climate model. We propose specific strategies for effectively learning rare events, quantifying the uncertainty of NN predictions, and making reliable predictions across various climates. For instance, one strategy to address the learning of rare events involves inflating the impact of infrequent events in the training data. We also demonstrate that several methods could be useful in determining the uncertainty of the predictions. Furthermore, we show that NNs trained on simulations of the historical period do not perform as well in warmer climates. We then improve NN performance by employing transfer learning using limited new data from warmer climates. This study provides lessons for developing robust and generalizable approaches for using NNs to improve models in the future. , Key Points Whole Atmosphere Community Climate Model’s orographic, convective, and frontal gravity wave parameterizations are emulated using neural nets to inform future modeling efforts Data imbalance is addressed via resampling and weighted loss; uncertainty quantification via Bayesian, dropout, and variational methods Performance of the neural nets in a warmer climate is improved via transfer learning with ∼1% new data