SOLAR VS HAIL: PIVOTING AWAY FROM DANGER
Vaisala Xweather Perspective
Expert Analysis from Vaisala Xweather
Author: Scott Mackaro, Ph.D., Director of Innovation & Forecasting, Vaisala Xweather
Hail is one of the most damaging forms of severe weather, yet forecasting it accurately remains a formidable scientific challenge. At its core, hail is a product of cloud microphysics—the small-scale processes that govern how particles form, grow, and evolve in the atmosphere. While scientists have a good theoretical understanding of how hail forms and under what conditions it thrives, the reality is that microphysics is one of the hardest areas of atmospheric science to model. The processes are short-lived, nonlinear, and notoriously difficult to observe directly.
That said, progress has been substantial. Technology advancements in radar, lightning detection, and satellite sensing have expanded both our understanding and forecasting capabilities. But, even today, observing hail remains a significant obstacle. In regions with radar coverage, forecasters can often identify hail aloft within storms. In regions without radar, lightning data serves as a useful proxy—after all, lightning cannot occur without the presence of ice in the cloud. Satellites offer additional promise however their spatial resolution and data latency are often mismatched with the short lifespan of a hail-producing core.
One common but flawed assumption is that hail falls straight down. In reality, it rarely does. The motion of the storm and surface-level winds can shift where hail impacts the ground. Accurately tracking the hail core and considering wind direction is key to pinpointing surface impacts—especially when assets are at risk.
Addressing Bias
Historically, ground truth has come from storm reports—human observations often submitted hours after the fact. These are inherently biased toward populated areas and suffer from limited spatial and temporal coverage. In the US, the NOAA NSSL/SPC team has developed the MRMS dataset, a powerful radar-based data fusion system. Two key products from this system—maximum estimated size of hail (MESH) and probability of severe hail (POSH)—have become widely used for validation and model training by commercial weather technology providers. But MRMS has some limitations. Its resolution is coarse when considering the size of hail impacts, and the dataset relies on several assumptions and approximations, much like the forecasts used to verify. Comparing MRMS to highly localized datasets, like those from solar farm operators or the National Renewable Energy Lab (NREL), reveals many missed events and false alarms—especially due to spatial mismatches.
Many hail forecasts today are built directly on MRMS—either by projecting it forward with optical flow techniques, validating against it, or using it to train AI models. These approaches have utility but inherit the dataset’s shortcomings. To overcome this, some groups are turning to alternative observational methods, including kinetic energy sensors, impact-based detectors, and post-damage assessments, such as when solar panels are shattered.
We operate both the National Lightning Detection Network (NLDN) and the Global Lightning Dataset (GLD360) at Vaisala. Recognizing that lightning is an indicator of ice aloft, we collaborated with the University of Oklahoma’s AI2ES program to explore whether lightning data could improve hail prediction alongside radar. The results were compelling—not only did lightning correlate with hail, but in many cases it rivalled radar reflectivity in predictive power.
This led to our Phase 1 approach: enhancing MRMS products with lightning data. Early case studies showed notable improvements in resolution and accuracy. Vaisala launched this capability in our Xweather Protect platform, which now delivers alerts for hail, lightning, and damaging winds via email, SMS, and API. These alerts will soon be ready to be integrated into control systems—such as SCADA—for automated asset protection actions, including commands to stow or resume operations.
Phase 2 is underway, focused on robust validation and tuning. We are analyzing two years of hail data across key regions like Colorado and Texas, using contingency metrics to refine performance. The work draws on trusted sources like solar farms and NREL to ground truth our models more accurately than MRMS alone.
In Phase 3, we’re integrating everything into a deep learning framework. Early versions are operational but limited in scope. This next step is critical because AI can learn regional nuances that traditional models miss. For example, in Colorado, the melting layer is close to the ground, so hail has little time to melt before hitting the surface. In eastern Oklahoma, the melting layer is higher, but storms tend to be stronger. These differences matter—and AI can adapt to them.
US lightning density as detected by the Vaisala Xweather National Lightning Detection Network (NLDN).
Global lightning density as detected by the Xweather Global Lightning Dataset (GLD360).
While we started in the US, we’re now extending this hail prediction system globally. By combining satellite data and our global lightning network, we can identify large storms anywhere on earth. With proper thresholds—calibrated using ground-truth validation—we are poised to take this capability worldwide. Once operational, we’ll reprocess historical data to build a high-resolution global hail risk inventory.
