Machine Learning for numerical weather and climate modelling: a review

Abstract

Abstract. Machine learning (ML) is increasing in popularity in the field of weather and climate modelling. Applications range from improved solvers and preconditioners, to parametrisation scheme emulation and replacement, and recently even to full ML-based weather and climate prediction models. While ML has been used in this space for more than 25 years, it is only in the last 10 or so years that progress has accelerated to the point that ML applications are becoming competitive with numerical knowledge-based alternatives. In this review, we provide a roughly chronological summary of the application of ML to aspects of weather and climate modelling from early publications through to the latest progress at the time of writing. We also provide an overview of key ML concepts and terms. Our aim is to provide a primer for researchers and model developers to rapidly familiarize and update themselves with the world of ML in the context of weather and climate models.

A Review of Application of Machine Learning in Storm Surge Problems

“…The ML techniques applied for storm surge forecasting still expose flaws and weaknesses and are not mature enough for operation prediction. There exists a large gap between research findings and the successful integration of innovative new approaches into major model upgrades [148]. In this part, we aimed to summarize the strengths and shortcomings of both ML and numerical methods concerning the efficiency, overall predictive accuracy and generalization ability to extreme events, timeliness of forecast, interpretability and transferability.…”

“…Only a small part of studies extended the lead time to 24 h [44,83], and fewer studies attempted to output the forecast water level time series with a length of over 24 h [38,48,62,78,79]. The ML models now show great capability in short-range forecasting, and a flurry of DL-driven models are being used in nowcasting in the atmospheric sciences [148], while for long-term forecasting, it still has a long way to go to be competitive with numerical models.…”

“…One of the primary drawbacks of data-driven methods lies in that there are no physical constraints [25]. A promising method known as physics-informed ML (PIML) is becoming increasingly popular in the earth system sciences and is conducive to improving model performance and interpretability [148]. The key approaches to incorporate physics and domain knowledge into ML models include direct approaches, like enforcing the conservation of physical properties and indirect approaches, such as designing model architectures under the guidance of domain expertise [150].…”

Machine Learning Methods in Climate Prediction: A Survey

“…However, machine learning algorithms serve as auxiliary tools for parameter tuning, feature engineering, or addressing specific limitations in the physical models, thereby improving their overall predictive accuracy and reliability. In this survey, ML-enhanced was divided into three catagories: bias correction, down-scaling, and emulation [66]. Model.…”

Moving beyond post-hoc XAI: Lessons learned from dynamical climate modeling

“…Machine learning (ML) is becoming increasingly utilized in climate science for tasks ranging from climate model emulation (Beucler et al 2019), to downscaling (McGinnis et al 2021), forecasting (Ham, Kim, and Luo 2019) and analyzing complex and large datasets more generally (for an overview of ML in climate science, see Reichstein et al 2019;Molina et al 2023;de Burgh-Day and Leeuwenburg 2023). Compared with physics-based methods, ML, once trained, has a key advantage: computational efficiency.…”