Using machine learning to represent sub-grid processes in weather and climate models holds promise, but also faces challenges. Incorporating physical knowledge can help.
Journal of Advances in Modeling Earth Systems (JAMES)
Modeling Entrainment with Machine Learning
Researchers present a new approach to modeling the stochastic mixing process of convection using a machine learning technique.
More Accurately Modeling Rain Formation
Rain and cloud droplets are treated as distinct categories in most models yet lie on a continuous droplet size spectrum in nature. Representing them as part of a continuous spectrum improves models.
Learning from Climate Simulations for Global Seasonal Forecast
A probabilistic deep learning methodology that learns from climate simulation big data offers advantageous seasonal forecasting skill and crucial climate model diagnosis information at a global scale.
Corrective Machine Learning for Improving Climate Models
A machine-learned correction enables an efficient coarse-grid global atmosphere model to better track the weather and time-mean precipitation of an expensive fine-grid ‘digital twin’ reference model.
The AI Forecaster: Machine Learning Takes On Weather Prediction
A novel approach to weather forecasting uses convolutional neural networks to generate exceptionally fast global forecasts based on past weather data.
A New Way to Represent Microphysical Uncertainty
A new way of representing microphysical uncertainty in convective-scale data assimilation reduces biases in model states and improves the accuracy of short-term precipitation forecasts.
Importance of High Clouds and Moisture in Rainstorm Aggregation
A study of the impacts of radiative interactions with different cloud types on aggregation of rainstorms finds that interactions with high-clouds and water vapor are key.
AeroCom Models Improved with Aerosol and Albedo Constraints
Satellite data has been used to correct the aerosol loading and land surface albedo in several AeroCom models, which has improved shortwave flux biases between models and observations.
Uncovering Hidden Errors in Simulated Precipitation
New metrics used to quantify errors in precipitation show that convection permitting simulations outperform coarser resolution simulations.