D8 Numerical Modeling and Prediction

Topic

Numerical weather prediction involves the numerical integration of the atmospheric primitive equations—which describe how air moves and changes—starting from an initial state and projecting it into the future. This initial state is simply the analysis produced by data assimilation; numerical weather prediction models take this analysis, advance it through time by solving the atmospheric primitive equations step-by-step, and thereby generate a forecast.

In the realm of numerical weather prediction, various meteorological centers operate their own global forecasting models. Each features a horizontal resolution of between nine and thirteen kilometers and over a hundred vertical levels, and each is capable of predicting the atmospheric state up to fifteen days in advance—though accuracy diminishes the further into the future the forecast extends. High-resolution regional models, by contrast, sacrifice global coverage for much finer resolution—ranging from one to three kilometers. This resolution is fine enough to explicitly resolve mesoscale phenomena such as deep convection, rather than having to approximate them as global models with coarser resolution must do.

Since no global forecasting model can determine the initial state of the atmosphere with perfect precision, ensemble forecasting quantifies the uncertainty introduced by this lack of exact knowledge. It runs the same model multiple times, each with a slightly different initial state—sometimes using dozens of perturbed initial states, such as fifty-one—and occasionally employing different versions of the model itself. The spread among the various members of this forecast ensemble serves as a measure of uncertainty: the greater the spread between ensemble members, the lower the atmospheric predictability at that time. Ultimately, the quality of any of these predictions—whether derived from a global model, a regional model, or an ensemble—is measured using various statistical scores: some assess the average error between the model's prediction and actual events; others evaluate how closely the predicted anomaly pattern matches the observed one; and still others measure the calibration of probabilities provided by ensemble forecasts—such as the likelihood of rain.

In recent years, artificial intelligence applied to meteorology has offered a different path toward achieving this same predictive capability: various neural models, trained by different laboratories, demonstrate predictive skill comparable to that of conventional global models, yet they require only minutes of computation rather than hours. The fundamental difference lies in how they arrive at the prediction: while numerical weather prediction explicitly solves atmospheric primitive equations step-by-step, these neural models are trained on vast reanalysis datasets and learn the equations governing atmospheric fluid dynamics directly from that data, without explicitly solving them for each individual forecast.

For the shortest timeframe—zero to two hours ahead—neither numerical weather prediction nor neural models can deliver results in time; instead, "nowcasting" addresses this window by combining the extrapolation of the latest radar and satellite imagery with high-resolution regional models, leveraging the fact that, on a scale of minutes and a few kilometers, tracking the movement of current atmospheric activity suffices. Thus, ranging from the slow, explicit integration of atmospheric primitive equations to the direct learning of neural models and the simple extrapolation used in nowcasting, the field of weather modeling encompasses various approaches pursuing the same goal: transforming the atmosphere's current state—captured through analysis—into the best possible prediction of its future.

ID:38