Knowledge Base

OpenSnow's Severe Weather Forecast System (StormNet) leverages cutting-edge artificial intelligence to deliver advanced severe weather forecasting with unmatched precision.

By analyzing massive datasets from Super-Res Radar, satellite, and ground-based sensors in real-time, our AI models identify patterns that indicate heightened risks for:

Unlike traditional forecasts, our system is continuously learning and adapting, offering faster, more accurate predictions when every second counts.

StormNet was engineered by Andrew Brady.

Andrew's passion for meteorology started when he was a young child. After school, he would read books about meteorology and, later, would read online articles on the subject. One highlight of his childhood was begging his mother for a megaphone so he could make severe weather announcements to the neighborhood where he grew up.

Years later, in 2020, Andrew founded AtmoSphere Analytics with one familiar goal: to improve prediction and communication around impactful weather and, eventually, save lives. It began with experiments using the open-source WRF model on his low-end PC. This evolved into WRF experiments on second-hand Dell PowerEdge servers. It started with one, which grew to four -- all running in parallel for WRF experiments. Andrew would tinker with configurations and source code, learning about how it all worked along the way. Eventually, he arrived at a version that he liked -- one that, in his experiments, would generally produce the most realistic forecasts when it came to impactful weather events. A combination of open source WRF with customized segments of code in the physics parameterizations. He then built a system to run WRF automatically, daily, to be posted on the AtmoSphere Analytics website: the Microscale-Mesoscale Prediction System (MMFS). 

While MMFS was helpful to look at and would produce interesting forecasts, he wanted to take it a step further: machine learning. He had been interested in machine learning for some time, but had only tinkered with the idea briefly. He brainstormed: "how can I make this even more centered around the goal? Around impactful weather prediction and communication?". His original goal was to apply machine learning to MMFS outputs to generate higher resolution, refined forecasts.

But, then, in 2021, he arrived at a new idea: severe weather forecasts from MMFS outputs. The idea seemed simple: take MMFS outputs, post-process them using machine learning to produce severe weather forecasts. This was the first step towards StormNet. Throughout 2021, he built a ML algorithm that would take MMFS outputs and would produce severe weather forecasts. It was called HazCast. In 2022, he started providing HazCast forecasts on his website alongside the MMFS forecasts. Throughout 2022 and 2023, he worked on tweaking HazCast and MMFS to make them as accurate as possible.

HazCast was interesting. It would produce broad forecasts that were, at times, quite accurate. HazCast had some significant limitations, however. Namely, HazCast would produce very broad/imprecise forecasts. It was also limited by the biases and data availability of MMFS. HazCast had a relatively simple algorithm, primarily due to data limitations from MMFS. The biggest limitation, however, was that Andrew felt that it wasn't quite 'there' when it came to making a difference (such as saving lives). The key to making a true difference with impactful weather prediction/communication is precision and accuracy. Over time, the idea of StormNet was formed: a very complex deep learning model which would take various data sources, as well as current conditions, and would produce hyper-local precise severe weather forecasts. StormNet: Severe storm and TOrnado  Real-Time Monitoring NETwork

Andrew started working on StormNet in 2023. He quickly realized that such a complex model wouldn't be able to run on his Dell poweredge server cluster. It needed graphics processing units (GPUs). AtmoSphere Analytics was making some money from subscriptions at the time, but not enough to purchase massive GPUs. Andrew had to decide between abandoning the project or making the model simple enough to work without GPUs. A third option came to mind -- building a GPU server at home. He decided to embrace this crazy idea, with hopes that StormNet would eventually be success and save lives -- it would be worth it. He researched different motherboards that had the specific adapter needed for high powered GPUs and learned that nearly all of those require a data-center setup.

At this point, he decided to do this completely from scratch. He ordered a motherboard that had the correct adapters, ordered the GPUs second hand, and ordered all of the remaining supplies to build this computing system. Then, he went to the local home improvement store to purchase a piece of plywood to put the system on. He had no idea how this was going to work -- or if it would work.. but it was worth a shot if he could build something special. Over the next several weeks, the parts came in and he built the system. It didn't work at first, but after lots of trial and error and re-builds, the green light on the motherboard finally came on -- it was working. 

 

With this computing system working, he then moved towards actually building the model. He had to collect all of the training data that he would use, then design an architecture. This had never been done before, so he had to figure it out on-the-fly. He tried various different architectures (graph attention networks, convolutional neural networks, multi-layer perceptron, etc) before eventually deciding that he would need to build something custom for this task.

After weeks of trial-and-error, StormNet v0.1 was trained. This was cool -- very cool. He ran the model on various past events, such as the Mayfield, KY 2021 EF4 tornado. StormNet was actually able to predict that a tornado would form and move into Mayfield with 30min+ lead-time. Trial-and-error continued through the end of 2023, and finally in January 2024, he announced StormNet to the world. Around that time, he set up a real-time inference system. At the time, StormNet was producing 1-hour tornado predictions across the eastern CONUS every 5 minutes. It quickly gained popularity in the weather community. Andrew quickly iterated updates and upgrades, allowing the system to learn from its own performance. By March 2024, after a ton of R&D effort, Andrew released StormNet v1.0.

Through Spring 2024, StormNet's popularity continued to skyrocket; largely due to the incredible accuracy and usefulness. New versions of StormNet were able to predict more than just tornadoes -- hail and damaging wind outputs were also added. By summer 2024, Andrew met with meteorologists from Fox Weather, The Weather Channel, RadarOmega, and others to gather valuable feedback on StormNet.

In fall 2024, OpenSnow acquired AtmoSphere Analytics with a vision: to make this groundbreaking technology available to many more people. From late 2024 into 2025, Andrew and the OpenSnow team have worked hard to improve StormNet and bring it to where it is today.

StormNet is a deep-learning AI model by OpenSnow.

It uses a proprietary combination of various data sources to analyze the current and predicted state of the atmosphere. Through extensive training, this AI engine can analyze and find very complex patterns that lead to hail, damaging winds, tornadoes, and lightning.

We tested StormNet on over 250 severe weather events from 2024. Data from 2024 was excluded from training for validation purposes. Here are some highlights from our testing:

Brayden Barton with the University of Oklahoma evaluated a very early version of StormNet and found that StormNet is a very impressive deep learning tool, with its claims largely backed by the results of this study'. Barton continues: 'At short lead times prior to tornadogenesis, this study found that StormNet outperforms the National Weather Service in tornado detection percentage at lower tornado probability thresholds, while even possessing the ability to detect potential for tornadogenesis up to an hour before tornadogenesis.' (6)

 

StormNet probabilities update every 2 minutes. Longer range guidance may change little or not at all with every 2-minute update since that data evolves at a slower pace than rapidly changing near-term conditions. 

No, StormNet is not a replacement for National Weather Service warnings, watches, or other guidance. StormNet is to be used as a supplement to any official guidance.

Short-Range Example:

StormNet outputs are plotted with Super-Res Radar.

The example above is displaying damaging wind probabilities within a line of storms. In the example, we are looping Super-Res Radar data with the damaging wind probabilities. The damaging wind probabiltiies are 30 minute windows, valid after the radar frame. The last 4 frames are current forecasts, for now to 30 minutes, 30 minutes to 1 hour, 1 hour to 2 hours and 2 hours to 3 hours. The grey to blue contours ahead of the storm indicate elevated damaging wind probabilities. 

In this example, StormNet is considering several elements:

The final product is contours of probability. In this example, damaging wind probabilities are around 40% to 50% ahead of the main core of the storm. 

Longer-Range Example:

In this example, radar is turned off and we are looking at hail probabilities for the next day.

Specifically, hail probabilities for 5:00 pm to 6:00 pm local time. We can see that StormNet is evaluating the state of the atmosphere, considering how the atmosphere may evolve over the next several hours, and how that may impact hail probabilities specifically at 5:00 pm to 6:00 pm the next day.

The blues indicate lower probabilities whereas the yellows are localized areas where there is increased confidence in hail occurring.

Radar reflectivity measures how much energy a radar signal bounces back from objects in the atmosphere, like precipitation. It indicates the size and concentration of particles (like raindrops or snowflakes) and is used to estimate precipitation intensity and type.

Higher reflectivity values generally mean larger and more numerous particles, often associated with heavier precipitation.

Radial velocity in measures the speed at which precipitation (like raindrops or snowflakes) is moving toward or away from the radar site.

Using the Doppler effect, the radar detects changes in the frequency of the returned signal caused by motion. If precipitation is moving toward the radar, the frequency increases (a positive velocity), and if it’s moving away, the frequency decreases (a negative velocity).

This information helps meteorologists understand wind patterns inside storms, which can reveal important details like rotation in a thunderstorm or wind shear that might indicate severe weather.

Spectrum width measures the variability or spread in the velocities of precipitation particles within a radar beam. Instead of showing the average speed like radial velocity, it tells how diverse the speeds are—like if some raindrops in a radar sample are moving faster or slower than others.

A small spectrum width means the particles are all moving at nearly the same speed, while a large spectrum width suggests turbulence, wind shear, or other chaotic motion. This helps meteorologists identify areas of atmospheric instability, like strong gust fronts or tornadoes.

With Radar + Risk, you can visualize the current and recent radar with StormNet hazard probabilities overlaid. The first group of timesteps are recent radar and StormNet frames, while the last 4 frames are the StormNet hazard risk probabilities for the next 30 minutes, 30 minutes to 1 hour, 1 hour to 2 hours, and 2 hours to 3 hours. 

In the above example, we visualize the radar and lightning probabilities of a line of thunderstorms in Nebraska. Watch as the StormNet lightning probabilities remain elevated along and ahead of the storms before pushing out ahead of the storms in the last 4 forecast frames.

Send an email to hello@opensnow.com and a real human will respond within 24 hours.