Diffusion and flow matching for precipitation downscaling over India: a benchmark against statistical and deterministic methods

To understand the weather that shapes our daily lives, we must look at the maps scientists use to predict it. Global climate models are powerful tools that simulate the Earth's atmosphere, but they see the world in broad strokes. Their grid cells are often hundreds of kilometers wide, large enough to average over mountains, coastlines, and valleys. For a scientist trying to predict a flood in a specific river valley or manage water for a local farm, this broad view is insufficient. The most dangerous weather events—intense downpours and flash floods—happen on scales these large models simply cannot resolve. To bridge this gap, researchers use a technique called downscaling. This process takes the coarse, low-resolution data from global models and infers what the weather looks like at a much finer, local level. The challenge is that rain is not a smooth, predictable blanket; it is erratic, heavy-tailed, and prone to extreme bursts. Traditional methods often fail here because they are designed to find the average, smoothing out the very extremes that cause the most damage.

A team of researchers led by Shailesh Kumar Jha at the Indian Institute of Technology Mandi has developed a new way to tackle this problem, specifically for the complex monsoon climate of India. They tested a new generation of artificial intelligence models designed not just to guess the average rainfall, but to recreate the full range of possibilities, including the rare, violent storms. By comparing these new methods against older statistical techniques and standard deep-learning networks, they found that only the new generative models could accurately reproduce the extreme rainfall events that drive flood risks. Their work suggests that to predict the weather that matters most for safety and infrastructure, we must stop trying to predict the average and start simulating the full spectrum of what nature might do.

The researchers focused on the Indian subcontinent, a region chosen not just for its importance, but because it presents one of the most difficult challenges for weather modeling. The Indian summer monsoon concentrates the year's rainfall into a few intense events, driven by steep mountain ranges like the Western Ghats and the Himalayas. These landscapes create sharp gradients where rainfall can change drastically over very short distances. To test their methods, the team used a "perfect-model" framework. Instead of trying to correct errors in a global climate model, they took high-resolution, real-world rainfall data, deliberately blurred it to look like a coarse global model, and then asked their algorithms to sharpen it back up. This allowed them to measure the pure skill of the downscaling methods without the noise of other climate model errors. They compared five different approaches: two older statistical methods that rely on finding past weather patterns, a standard deep-learning network that predicts the most likely outcome, and two new generative models based on advanced probability theory.

The results were stark. The older methods and the standard deep-learning network consistently failed to capture the upper limits of rainfall. Because these models are trained to minimize error by finding the middle ground, they naturally smooth out the data. When faced with a single low-resolution input that could correspond to many different high-resolution realities, they chose the average. For a right-skewed variable like rainfall, where most days are dry but a few are extremely wet, the average is a poor representation of reality. These methods underestimated the intensity of extreme rainfall by as much as 67 percent. In practical terms, a method that predicts a 10-year flood level as 40 millimeters when the reality is 126 millimeters is dangerously misleading for anyone building a dam or planning a city. The standard deep-learning network, while better at spatial patterns, still smoothed the peaks, effectively turning rare, catastrophic storms into moderate, manageable showers.

In contrast, the two new generative models, which the researchers call "probability-path" models, succeeded where the others failed. These models do not try to guess a single average outcome. Instead, they learn the entire distribution of possible rainfall patterns for a given set of conditions. When asked to downscale a weather map, they sample from this full range of possibilities, generating a realistic high-resolution field that includes the rare, extreme events. One of these models, based on a technique called Flow Matching, and the other, based on denoising diffusion, both recovered the observed statistical properties of extreme rainfall with remarkable precision. They reproduced the shape of the extreme-value distribution to within a fraction of a percent and estimated the 10-year return level—the amount of rain expected once every decade—within 1 percent of the observed value. This means they could accurately predict the intensity of the most dangerous storms, preserving the heavy tail of the distribution that the other methods had smoothed away.

Beyond accuracy, the study also uncovered a significant difference in efficiency between the two new generative models. While both produced high-quality results, the Flow Matching model achieved this with far less computational effort. Generating a single high-resolution weather map with the diffusion model required many more calculations than the Flow Matching model. The Flow Matching approach could produce results of equal or better quality using only a fraction of the computing steps. This efficiency is crucial because probabilistic forecasting often requires running a model dozens or hundreds of times to create an ensemble of possible futures. If each run is slow, creating a reliable forecast becomes prohibitively expensive. The researchers found that the Flow Matching model could match the performance of the more computationally heavy diffusion model while running significantly faster, making it a more practical tool for operational forecasting and large-scale risk assessment.

The study also examined how well these models captured the texture and persistence of rainfall over time. Floods and droughts depend not just on how hard it rains, but on how long it rains. The older methods struggled here as well. The statistical analogue methods, which stitch together pieces of past weather, tended to break up long dry spells or create unrealistically long wet periods. The standard deep-learning network smoothed the landscape, losing the fine-scale texture of rain bands. The generative models, however, reproduced the natural variability of wet and dry spells with high fidelity. They maintained the correct spatial roughness, ensuring that rain was not just a smooth gradient but a patchwork of intense cells and dry gaps, just as it appears in reality. They also correctly captured the rhythm of the monsoon, distinguishing between active phases of heavy rain and break phases of lighter weather, a nuance that the averaging methods often blurred together.

Despite these successes, the researchers noted that the work is not without limitations. The models still showed a slight tendency to underestimate rainfall in the very wettest, most mountainous areas, a known challenge for all current downscaling techniques. Furthermore, while the models could generate multiple possible outcomes, the spread of these outcomes was sometimes too narrow, meaning they were slightly less variable than the real world. This suggests that while the models are excellent at capturing the extremes, they may still need further refinement to be perfectly calibrated for probabilistic risk assessment. The study was also confined to a "perfect-model" setup, meaning it tested the downscaling logic in isolation from the errors inherent in global climate models. The next step for the researchers is to integrate these tools with real-world climate projections to see how they perform in predicting future climate scenarios.

Ultimately, this research highlights a fundamental shift in how we approach weather prediction. For decades, the goal was to find the single best estimate of the future, the most likely path the weather would take. This study demonstrates that for variables like rainfall, where the extremes drive the consequences, the "most likely" path is often the wrong one. By embracing the full range of possibilities and using generative models to sample from the entire distribution of outcomes, scientists can finally recover the extreme events that define flood risk and water management. The ability to do this efficiently, as shown by the Flow Matching model, opens the door to high-resolution, probabilistic forecasts that can better inform decisions in a changing climate. The findings confirm that to understand the weather that matters, we must stop looking for the average and start simulating the full spectrum of nature's volatility.