Glossary

Term	Definition
Backtesting	The process of evaluating a forecast model against historical data (using past initial conditions) to assess its performance and accuracy over past events.
Bias Correction	A post-processing technique applied to model outputs to reduce systematic errors (biases) relative to observations. Often needed when model training targets (like reanalysis) don't perfectly match surface measurements.
BigQuery	A serverless, highly scalable data warehouse on Google Cloud. WeatherNext datasets are available here for large-scale SQL-based analysis.
Blurring	A phenomenon where ML weather forecasts become progressively smoother at longer lead times, washing out fine-scale detail. This happens because deterministic models learn to predict the average of possible outcomes as uncertainty grows. Ensemble models are less affected because individual members preserve sharper structures.
Brier Score	A metric for evaluating the accuracy of probabilistic forecasts of binary events (e.g., "will precipitation exceed 1 mm?"). It ranges from 0 (perfect) to 1 (worst). Lower is better.
CRPS (Continuous Ranked Probability Score)	A primary metric for evaluating the accuracy of a probabilistic (ensemble) forecast. It measures both the accuracy and the "spread" (confidence) of the forecast. Lower is better.
Deterministic Forecast	A single "best guess" forecast that predicts one specific future weather state. This contrasts with a probabilistic forecast.
Diffusion Model	A type of generative AI model used in WeatherNext Gen (GenCast) that works by adding "noise" to data and then learning how to reverse the process iteratively to generate new, realistic outputs.
Earth Engine	A Google Cloud platform for petabyte-scale scientific analysis and visualization of geospatial datasets. WeatherNext data is available here for raster-based analysis.
Ensemble / Ensemble Forecast	A probabilistic forecasting method that runs a model multiple times (or runs multiple models) to generate a collection of possible future weather scenarios, rather than just one.
ECMWF ENS (Ensemble Prediction System)	The primary ensemble operational forecast model from the ECMWF. WeatherNext Gen and WeatherNext 2 are often benchmarked against it.
ECMWF ERA5	A widely used global "reanalysis" dataset from the ECMWF. It provides a comprehensive, high-quality historical record of the Earth's climate and is often used to train AI weather models.
ECMWF HRES (High Resolution Forecast)	The primary deterministic operational forecast model from the ECMWF, a leading traditional NWP model. WeatherNext Graph is often benchmarked against it.
FGN (Functional Generative Network)	The neural network architecture used by WeatherNext 2 and WeatherNext 3. FGN uses a graph transformer with noisy weights to generate a full probabilistic ensemble in a single forward pass, rather than requiring iterative sampling like diffusion models.
Geostationary Satellite Mosaic	A composite image stitched from multiple geostationary weather satellites positioned around the equator, providing near-continuous global coverage. WeatherNext 3 ingests this as a direct model input, enabling hourly forecast initialization.
Geopotential (Height)	A measure of atmospheric pressure defined as the altitude at which a specific pressure value (e.g., the 500 hPa level) is reached in the atmosphere. It is fundamental to tracking large-scale weather patterns.
GFS (Global Forecast System)	The primary global weather forecast model run by NOAA in the United States.
GNN (Graph Neural Network)	A type of AI architecture used in WeatherNext Graph (GraphCast) that represents the Earth as a graph and models the connections between different points.
Google Cloud Storage (GCS)	Google Cloud's object storage service. WeatherNext data is stored here in the Zarr format for direct file access.
hPa (hectopascal)	A standard unit of atmospheric pressure. Used to define "pressure levels" in the atmosphere, which are standard altitudes for weather analysis. A pressure of "one atmosphere" is equivalent to 1013.25 hPa.
HRES-fc0	Refers to the initial step used to generate the HRES forecasts, which are provided by ECMWF. It is a form of operational analysis, but differs slightly from the HRES Analysis, which is the main operational analysis product from ECMWF.
HRES-fc1to5	Refers to the first 5 hourly steps produced by the ECMWF HRES model.
IMERG (Integrated Multi-satellite Retrievals for GPM)	A NASA satellite-based global precipitation dataset that combines data from the Global Precipitation Measurement (GPM) mission constellation. WeatherNext 3 uses IMERG as one of three precipitation training targets.
Inference	The process of using a trained AI model to make a prediction given some input data (e.g., running the WeatherNext model to generate a forecast).
Initial Conditions	The set of atmospheric data (like temperature, pressure, and wind) at a specific starting time, used to initialize a weather model run.
Lead Time	The duration a forecast predicts into the future measured from the time of the initial condition (e.g., "10-day lead time").
MetNet	A specialized Google AI model designed for nowcasting, focusing on high-resolution precipitation forecasts for the immediate future (e.g., 0-12 hours).
Multi-resolution Forecast	A forecast that produces output at different spatial resolutions for different variables from a single model run. For example, WeatherNext 3 produces station-calibrated temperature and dew point at 5 km, gridded surface fields (including 2m temperature and wind) at 10 km, and atmospheric variables at 25 km.
NeuralGCM	A Google AI model that combines a General Circulation Model (GCM) with AI models. It learns the fundamental physics of the atmosphere to run long-range global weather simulations very quickly.
Nowcasting	The practice of forecasting weather for the very short term, typically 0-6 hours ahead, often with a focus on precipitation.
NWP (Numerical Weather Prediction)	The physics-based mathematical method of weather forecasting that uses complex hydrodynamic models of the atmosphere and ocean. Typically NWP models run on supercomputers (e.g., HRES, GFS).
Operational Analysis	A scientific method that blends real-time observations with an operational NWP to create a comprehensive and consistent "best guess" of the current state of the atmosphere. (Related to Reanalysis).
Pressure Level	A standard surface in the atmosphere defined by its pressure (e.g., 500 hPa), used for analyzing weather patterns at different altitudes.
Probabilistic Forecasting	A type of forecast that provides a range of possible outcomes and the likelihood (probability) of each one occurring, often by means of an ensemble of forecasts.
Reanalysis	A scientific method that blends historical observations with a modern forecast model (like ERA5) to create a comprehensive and consistent "best guess" dataset of past weather. (Related to Operational Analysis).
RMSE (Root Mean Squared Error)	A common metric for evaluating the accuracy of a deterministic forecast. It measures the average magnitude of the errors. Lower is better.
Spatial Resolution	The size of the grid cells in a forecast (e.g., 0.25 degrees, which corresponds to a cell of about 27 km x 27 km at the equator). A smaller grid size (higher resolution) provides more geographic detail.
Temporal Resolution	The frequency of the forecast's time steps (e.g., 1-hour, 6-hour). A 1-hour resolution provides a more detailed view of how weather evolves.
TPU (Tensor Processing Unit)	Google's custom-designed accelerator chip, optimized for machine learning workloads.
Vertex AI Model Garden	A Google Cloud service where developers can discover, deploy, and manage AI models, including WeatherNext, for custom, on-demand inference.
WeatherNext	A family of advanced, AI-powered weather forecasting models developed by Google. The latest model is WeatherNext 3.
WeatherNext 3	Google's most advanced global weather forecasting model. Uses the FGN architecture with live geostationary satellite input, multi-resolution output (0.05° stations, 0.1° gridded surface, 0.25° pressure levels), and station-trained surface variables. Produces 64-member ensembles with hourly timesteps (360h for 6-hourly cycles, 48h for interim hourly runs).
WeatherNext 2 (FGN)	An ensemble forecasting model based on the Functional Generative Network architecture. Generates 64-member ensembles in a single forward pass with 0.25° resolution and 6-hourly timesteps. Superseded by WeatherNext 3.
WeatherNext Gen (GenCast)	A legacy AI model that uses a diffusion-based approach to generate an ensemble forecast. Instead of one single prediction, it produces many possible weather futures to capture uncertainty.
WeatherNext Graph (GraphCast)	A legacy AI model that uses Graph Neural Networks (GNNs) to produce a single, high-resolution deterministic forecast. Known for its high accuracy, outperforming conventional numerical weather prediction models in many metrics.
Zarr	A data format for storing "chunked," N-dimensional arrays. It is highly efficient for cloud-based scientific datasets like weather forecasts, allowing users to read small pieces without downloading the entire file.