Artificial intelligence for weather and climate: a survey of methods, benchmarks, and scientific machine learning challenges

Weather and climate science is a uniquely demanding testbed for artificial intelligence, combining petabyte-scale spatiotemporal complex data on a rotating sphere, governing partial differential equations coupled to global observational and reanalysis archives, climate non-stationarity that distinguishes it from stationary fluid-dynamics or seismology benchmarks, and operational forecast verification against rare extremes. This survey makes four contributions. First, it organises AI architectures, i.e. convolutional networks, graph neural networks, transformers, neural operators, and generative models, by their physical inductive biases, with geometric deep learning and operator learning as unifying themes. Second, it gives systematic treatment to representation learning and foundation models, uncertainty quantification, and causal discovery as a complement to prediction. Third, it names and analyses evaluation pathologies in scientific machine learning, including the RMSE smoothness bias under which mean-squared-error training systematically rewards blurred forecasts and penalises sharp but slightly displaced predictions, ERA5 training–evaluation circularity, and benchmark overfitting. Fourth, it frames the open problems, i.e. distribution shift under non-stationary climate, physical consistency of learned operators, sample efficiency in data-sparse regimes, and intrinsic interpretability, as scientific machine learning problems generalisable beyond meteorology.