AI in Atmospheric Sciences: An introduction

This document provides an overview of machine learning (ML) applications in weather, climate, and atmospheric sciences, focusing on observations, forecasting, data assimilation, post-processing, and operational meteorology.

1. Observations

Machine learning enhances observational datasets by improving data quality, fusing data sources, and extracting meaningful features.

Applications:

• Quality Control:

  • Example: Identifying anomalies in radar and satellite observations.

  • Reference:

• Data Fusion:

  • Example: Merging data from multiple satellites and sensors for higher accuracy in storm monitoring. NVIDIA’s ML-based data fusion models have been instrumental in creating unified observational datasets.

  • Reference:

• Feature Extraction:

  • Example: Extracting storm patterns from geostationary satellite imagery using DeepMind’s convolutional neural networks.

  • Reference: Lee, C., et al. (2019). “Deep learning based cloud detection for medium and high resolution remote sensing images of different sensors” Remote Sensing.

---

2. Forecast Model

ML enhances traditional physics-based forecast models by accelerating simulations and improving subgrid parameterizations.

Applications:

• Surrogate Modeling:

  • Example: Pierre Gentine’s group used ML as surrogates for convection parameterization in atmospheric models.

  • Reference:

• Hybrid Models:

  • Example: NVIDIA’s hybrid models combine GPUs for physics simulations with ML-based turbulence models.

  • Reference:

• Ensemble Forecast Emulation:

  • Example: ECMWF developed ML techniques to emulate their ensemble systems, accelerating probabilistic forecasting.

  • Reference:

---

3. Data Assimilation

ML enhances the process of integrating observations into forecast models.

Applications:

• Enhanced Observation-Model Integration:

  • Example: Microsoft Research applied deep learning to reduce biases in data assimilation systems.

  • Reference:

• Sparse Data Assimilation:

  • Example: Huawei’s AI Lab leveraged autoencoders to handle sparse observational data for tropical cyclone analysis.

  • Reference:

• Hybrid Data Assimilation:

  • Example: DeepMind explored combining traditional 4D-Var with neural networks for long-range forecasts.

  • Reference:

---

4. Post-Processing

ML improves forecast outputs through bias correction, downscaling, and uncertainty calibration.

Applications:

• Bias Correction:

  • Example: Random forests used to adjust biases in ECMWF’s rainfall forecasts.

  • Reference:

• Downscaling:

  • Example: Huawei’s super-resolution ML models improve climate model resolutions from 50 km to 5 km.

  • Reference:

• Forecast Calibration:

  • Example: DeepMind’s probabilistic calibration of weather forecasts using Gaussian processes.

  • Reference:

---

5. Ocean & Climate

ML enables detailed predictions of ocean dynamics and climate change impacts.

Applications:

• Ocean Current Prediction:

  • Example: Predicting Gulf Stream dynamics using neural networks.

  • Reference:

• Climate Change Impact Assessment:

  • Example: Evaluating fire risk due to climate change in Europe.

  • Reference:

• Climate Downscaling:

  • Example: High-resolution projections using LEAP’s ML-powered Earth system models.

  • Reference:

---

6. Operational Meteorology

ML supports real-time meteorological decision-making and early warnings.

Applications:

• Nowcasting:

  • Example: DeepMind’s precipitation nowcasting in London using deep generative models.

  • Reference:

• Extreme Event Detection:

  • Example: CNN-based detection of tropical cyclones from satellite imagery.

  • Reference:

• Decision Support Systems:

  • Example: Microsoft’s AI-based systems integrating flood forecasts with evacuation planning.

  • Reference: