Really easy to find.
https://weathersats.com/the-evolution-and-future-of-ai-in-weather-forecasting/#:~:text=AI Models Enhancing Numerical Weather,forecasts more frequently and accurately.
Normalization and Formatting: Data, often in GRIB or NetCDF formats, is cleaned and structured into consistent grids, frequently at 0.25-degree resolution.
Feature Engineering: Raw data is converted into actionable features, such as calculating "feels like" temperature or creating precipitation intensity categories.
Data Assimilation for "Real-Time" Updates
AI models use advanced techniques to ingest the current state of the atmosphere into the model to begin a forecast:
AI-Driven Assimilation: Instead of traditional, slow numerical methods, AI accelerates the integration of new observations by filling in data gaps in regions where measurements are sparse.
Hybrid Approaches: Some models, such as NeuralGCM, blend AI with traditional physical simulations to ensure the input data respects atmospheric dynamics.
Specialized Adapters: To shift from historical training data to live data, models may use specialized neural network architectures (like U-Net) to map real-time, "messy" data into the same format used for training.
Key Technologies Used
Neural Networks (CNNs/RNNs): Convolutional Neural Networks analyze spatial satellite imagery, while Recurrent Neural Networks process time-series data.
Cloud Platforms: Models are often hosted on AWS, Google Cloud, or Microsoft Azure to manage the high volume of data.
Transformers: Used to process long-range dependencies in atmospheric data.