Machine Learning Weather Models: The Good, the Bad, and the Tropical

The field of weather forecasting is undergoing a transformation. Machine Learning Weather Prediction models (MLWP) are at the forefront, promising to reshape how we predict and prepare for weather phenomena. Yet, regions like the Global South, their efficacy remains largely uncharted territory.

The Brazilian Challenge

A recent study evaluated GraphCast, a machine learning model, against the deterministic ECMWF IFS HRES across four Brazilian climatic sub-regions. The focus was on key tropospheric variables: $T_{850}$, $Q_{850}$, and $Z_{500}$. These variables were tracked over distinct seasonal windows, with the operational IFS analysis providing the ground truth.

In the austral winter, GraphCast struggled with medium-range predictions, particularly for $Z_{500}$ in southern Brazil where baroclinic systems move swiftly. However, it regained composure in the extended range, where its ability to smooth chaotic variability shone. Was this smoothing a triumph or a flaw?

Summer Successes

During the wet austral summer, GraphCast excelled at capturing large-scale moisture transport. Yet, its tendency to dampen high-frequency convective variability challenged traditional NWP models, indicating both a strength and a notable caveat. The key finding: MLWP models can outperform in certain conditions but falter where high-frequency variability is essential.

A Call for Tropicalization

The study's findings pave the way for "tropicalization" efforts. Brazil's diverse climate offers a unique testing ground, essential for optimizing these machine learning models. But will these efforts be enough to make MLWP models a reliable tool in such complex environments?

Crucially, these models establish a baseline for Brazil, highlighting the specific physical boundaries that need addressing. They show promise, yet there's much work ahead to ensure they can handle the dynamic and convective nature of tropical climates.

The ablation study reveals that while these models can bring significant advancements, they must be fine-tuned for each region's unique climate patterns. The question remains: How soon can these models be adapted for reliable use in the Global South?