Probabilistic recovery of high-impact urban convective rainfall from radar-assimilation experiments using conditional diffusion and distributional attribution

Predicting where rain will fall is one of the hardest tasks in weather science, especially when that rain comes in short, violent bursts over a city. Unlike steady, widespread showers that cover a whole region, these convective storms can form and move with such speed that a tiny error in guessing their starting position leads to a massive mistake in the forecast. If a computer model places a storm just a few miles off, the city might get a dry report while the next town over floods. To solve this, meteorologists often turn to weather radar, which acts like a high-speed camera, capturing the wind and rain inside storms as they happen. They feed this live data into complex computer simulations to correct the forecast. However, even with this live data, the models often still get the location wrong. When the raw forecast fails, scientists are increasingly turning to a different kind of tool: artificial intelligence that learns from the patterns in the failed forecasts to find the correct answer hidden within the data.

A researcher at Morgan State University recently tested a new way to handle this problem using a specific, severe storm that hit Baltimore, Maryland, on July 28, 2026. This event caused flash flooding, submerged vehicles, and widespread power outages. The scientist wanted to see if they could fix the forecast by combining three distinct steps. First, they ran a standard computer simulation of the atmosphere. Second, they tried to improve that simulation by feeding it real radar data from the storm, a process known as data assimilation. Third, they applied a sophisticated machine-learning system, called a conditional diffusion model, to the output of those simulations. This machine-learning tool does not just guess a single number; it learns the shape of uncertainty, allowing it to generate a range of possible outcomes that are statistically more likely to be correct than the raw computer forecast.

The researcher began by running three different versions of their computer simulation. One version ignored the radar data entirely, serving as a control. The other two versions tried to incorporate the radar data: one used only the wind speed measured by the radar, while the other used both the wind speed and the intensity of the rain echoes. The goal was to see if adding the radar data would move the predicted storm closer to where it actually happened. The results were clear and somewhat disappointing for the traditional approach. While the radar data did change the internal state of the computer model, it failed to fix the most critical error: the location of the rain. In all three versions, the model placed the center of the rainfall about 40 kilometers away from where it actually occurred. The model also failed to predict the heavy rain that caused the flooding, missing the event entirely in the raw output. The radar data had been successfully absorbed into the system, but it was not enough to overcome the model's tendency to place the storm in the wrong spot.

Next, the researcher applied their machine-learning post-processing tool to the results of these three simulations. This tool acted like a translator, taking the flawed, displaced forecasts and using the surrounding atmospheric clues to reconstruct where the rain should have been. The improvement was dramatic. By analyzing patterns in the wind, pressure, and moisture fields that the computer model had correctly simulated, the machine learning system shifted the predicted rain back toward the city. The error in the location of the rain center shrank from 40 kilometers down to less than 4 kilometers in the best cases. The accuracy of the rainfall amounts also improved significantly, with the error in the total rain volume dropping by about a quarter. The system did not just get the numbers right; it also provided a much better sense of where the rain was likely to fall, turning a near-zero chance of a correct match into a strong correlation with reality.

The study also looked closely at why the machine learning worked so well. By analyzing which pieces of information the system relied on most heavily, the researcher found that the tool did not simply look at the raw rain forecast to make its correction. Instead, it paid the most attention to the wind speed near the ground, the amount of water in the clouds, the air pressure, and the broader pattern of rain over a wider area. This suggests that the machine learning model was able to "read" the atmospheric conditions that the computer simulation got right and use those clues to deduce the correct location of the storm, effectively bypassing the initial mistake in the rain forecast. The system was particularly good at predicting moderate to heavy rain, though it still struggled with the absolute most extreme amounts of water, a challenge that remains difficult for all forecasting methods.

This work demonstrates that while feeding radar data directly into a weather model is a necessary step, it is not always enough to fix a forecast that has already gone off course. The real breakthrough in this study was showing that a machine-learning system could recover the lost accuracy by finding the correct answer hidden inside the flawed data. The researcher did not claim to have solved the problem of predicting extreme weather forever, noting that the most intense rainfall remained difficult to capture. However, they successfully proved that a hybrid approach, combining physics-based simulations with intelligent post-processing, can significantly improve the prediction of dangerous, localized storms. For cities like Baltimore, where a few inches of rain in the wrong place can cause disaster, this ability to correct the location and intensity of a storm forecast offers a powerful new tool for protecting lives and property.