From Theorems to Thunderstorms: Forecasting West Africa’s Rains with AI
Rainfall forecasting in tropical Africa is notoriously difficult, and even current, state of the art numerical weather prediction (NWP) models often fail to deliver reliable guidance for farmers, water managers, and disaster response teams. When Boston University Professor of Mathematics & Statistics and Computing & Data Sciences Yves Atchadé first heard from a colleague that “rainfall forecasting is really hard under the tropics, particularly in Africa,” he saw a challenge he could not ignore.
“That’s actually what sparked my curiosity,” Atchadé says. “I’ve spent most of my career writing theorems and working with Monte Carlo methods, but I’ve always wanted to put that statistical machinery into something applied. Weather forecasting under the tropics is a perfect place where the math really matters.”
Atchadé and his team focused on Ghana in West Africa as a case study, using rainfall estimates from NASA and the Japanese space agency’s Global Precipitation Measurement (GPM) mission, which has been collecting satellite observations since around 2000. They combined these with the ERA5 reanalysis dataset from the European Centre for Medium‑Range Weather Forecasts (ECMWF), which provides a consistent global picture of the atmosphere over recent decades.
The team built a deep learning model that is specifically tuned to predict 24 hour rainfall there 12 or 30 hours ahead. Atchadé explains: “We wanted to show that a small team, with a small model, could still match or even beat the state of the art NWP system in this region.”
One of the natural worries with deep learning rainfall models is that they might just “memorize” past weather instead of learning real physical patterns. Atchadé’s group tackled this by treating their model as a statistical model first and using regularization and careful validation to guard against overfitting. “After we fit the model, we looked to understand what was learned by the model, and we compared that to what scientists know about rainfall formation in the area. A lot of interesting phenomena were captured by the model.”
At the core, the model learns that certain combinations of humidity, wind patterns, and large-scale wave features are meaningful predictors of rainfall across the region. In their Ghana experiments, the deep learning model outperforms ECMWF’s own forecasts at matching observed rainfall and identifying heavy rain events, even though the model is much smaller.
For Atchadé, rainfall forecasts are not just an academic exercise. “In many African countries, people still rely on predictions from Europe, not from their own institutions,” he notes. “If local researchers can build models that match or surpass those, it’s a way to train local scientists and to give farmers better information for planting and harvesting.”
He also stresses that models are inherently probabilistic. “Statistical models only make probability statements, and there will always be uncertainty,” he says. To manage that risk, his team utilized ensemble modeling, which means running multiple simulations or using multiple models to chart out different possible futures. “That’s exactly how operational weather forecasts are presented: the storm is most likely to go this way, but it could also go that way.”
To explain the model to non-technical decision makers, Atchadé keeps the language simple. “We talk about the model output,” he says. “We talk about possible scenarios and the likelihood of each.” Farmers want to know: is it likely to rain in the next two days, or is there a real chance of a heavy downpour that could flood my fields? The model packages the output so it’s not about neural networks, but rather about risk, and what to do under each scenario.
For Atchadé’s lab, the rainfall work is already evolving. The next step “is a super exciting thing for us,” he adds. His postdoc and PhD student collaborators are now turning from daily rainfall to a bigger, more seasonal question: predicting the West African monsoon months in advance.
“For farmers, it is perhaps the most important prediction we can make,” Atchadé says. If the model can give them a credible forecast in March or April about the timing of the monsoon, that’s far more useful than a single day‑ahead rainfall number.
-- Shriya Jonnalagadda (CDS'28), Data Science Research Communications Intern