Atmospheric Model at 220m Resolution: How it Eliminates Phantom Precipitation

Imagine a weather model capable of distinguishing every alleyway in a city: the calculation grid now measures only 220m on each side, equivalent to a basketball court. At this resolution, global climate simulations no longer display the infamous 'popcorn' showers, intense and localized rains that never exist in reality but appear in coarser-resolution models. The result? More reliable predictions of extreme events like violent thunderstorms or sudden floods at the global level.

Model Result: Disappearance of 'Popcorn' Precipitation at Global Scale

By refining the grid to 220m, researchers observed that intense rain episodes previously generated by numerical artifacts disappeared completely. These 'popcorns' were symptomatic of a poor representation of microphysical processes, notably droplet formation and their interaction with atmospheric dynamics. The new model reproduces precipitation coherently with ground station data and satellite observations, confirming that the ultra-fine resolution captures temperature and humidity gradients that were previously smoothed and poorly treated at the previous scale.

This improvement is not just an aesthetic correction: it reduces prediction uncertainty in areas where thunderstorms are frequent, such as Southeast Asia or the US West Coast. According to initial tests, the average difference between predictions from the 220m resolution model and real observations decreases by 30% for precipitation above 20mm/h. This increased precision opens the door to more targeted alerts and better emergency planning in case of sudden flooding.

Method: Ultra-Fine Resolution, Satellite Data, and Machine Learning

The heart of the new system relies on a neural network specifically trained to optimize energy flow calculations at the meter scale. By combining machine learning with atmospheric data from Copernicus and other satellites, the model integrates temperature, humidity, and wind speed information with unprecedented granularity. This deep learning process allows the model to correct historical biases and better represent convection processes.

Meanwhile, scientists have integrated modeling techniques used in projects like GraphCast and Pangu-Weather, which use graphical neural network architectures to manage complex interactions between grid points. The result is a hybrid predictive model where traditional physical calculation coexists with AI modules, offering better uncertainty management. Satellite data, notably height measurements...