AWS Builder Center
Building a Virtual Storm Chaser with AWS

Building a Virtual Storm Chaser with AWS

This article covers a side project I worked on this summer while I was a Solutions Architect intern with AWS. I developed an application to track storms in realtime across the continental United States, enabling analysis of tornadic thunderstorms and machine learning models.

Introduction

Ever since I was a kid, I’ve held a deep fascination for severe weather. Over the years, that interest evolved alongside a growing passion for technology, inspiring me to start designing digital tools to explore the atmosphere further.
The privilege I had this summer, with the expanse of AWS cloud services and the latest frontier models at my fingertips, motivated me to put both passions together. I wanted to learn, be curious, and see what I could build.
My initial goal was to see if I could make a pipeline to help give earlier warning for tornadic events. That goal morphed into something I did not plan for, an end-to-end virtual storm chasing application.
Real storm chasers drive toward the danger while everyone else is fleeing. What I built captures that same thrill through a live radar screen instead of a windshield.
Want to see it in action? I have prepared a 40-minute  video demonstration  of the application. Feel free to skip around to the parts that interest you most.

What I made!

In patches of time throughout the summer, I created an end-to-end architecture that identifies, tracks, interprets, and stores data from storms occurring across the continental United States, 24/7.
Every five minutes, the pipeline runs a full chase cycle.
  • It checks NOAA's live severe-storm feed to decide which storms are worth chasing right now.
  • It pulls the newest scan from the nearest of roughly 160 weather radars in the national network.
  • It records everything, including the model score, the storm's properties, the exact radar file, and the storm's track over time.
High level diagram illustrating the workflow of the storm tracking application.

A storm is only flagged when two independent sources agree. NOAA's storm tracker must report tornado potential, and the model must independently give the same storm's radar a high score. One noisy signal on its own is not enough.

How the chase flows through AWS

Here is that same cycle again from the infrastructure's point of view, and why each service earned its spot.
Illustration of the flow of AWS cloud services behind the application.

The timer. Amazon EventBridge starts a chase cycle every five minutes. The cadence is deliberate. The radars themselves complete a new volume scan roughly every four to six minutes, so polling faster would mostly re-analyze the same data.
The scout. The first AWS Lambda function, the dispatcher, reads NOAA's free ProbSevere feed, which already fuses radar, satellite, and lightning data into per-storm hazard estimates. This feed is the cost-control decision of the whole project. Running a deep-learning model continuously against every scan from every radar in the country would be expensive and almost entirely wasted, since the vast majority of scans contain no tornadic activity. Instead, NOAA's operational tracker nominates the candidates, and only those storms get the expensive radar analysis. The dispatcher maps each flagged storm to its nearest radar site, groups the storms by site, and drops one message per site onto a queue.
The waiting line. Amazon SQS holds those per-site messages, and it is what keeps the system calm during an outbreak. Radar sites are independent, so a late or unreadable scan at one site gets retried without holding up storms assigned to every other site, and a message that keeps failing is set aside in a dead-letter queue for inspection instead of silently disappearing. Because there is exactly one message per activated site, the number of workers running at once can never exceed the number of radar sites in the country, no matter how large the weather event. Concurrency stays bounded by design, not by luck.
The heavy lifting. A worker Lambda, packaged as a container image so it can carry the radar-science libraries, picks up each site's message. Downloading and preparing a radar volume is the expensive step, roughly ninety seconds of decoding and signal processing, while scoring one more storm inside an already-prepared scan takes about a second. So each volume is prepared once and reused for every storm near that radar. One scan covers everything within roughly 230 kilometers of the station, while the model examines a single storm at a time. The worker therefore cuts a storm-centered slice out of the prepared volume for each flagged storm, aimed using that storm's tracked position, and sends each slice to the model separately. One download, one preparation, many slices. The radar itself comes from NOAA's NEXRAD archive, hosted as an AWS Open Data bucket. Because the pipeline runs in the same region as that bucket, the downloads travel AWS's internal network with no transfer charges.
The model host. Amazon SageMaker hosts TorNet on a serverless endpoint. The worker sends a prepared, storm-centered slice of radar, and the endpoint returns a tornado-resemblance score in under a second. Serverless matters here because storms are bursty. The endpoint costs nothing while the weather is quiet and wakes up when the dispatcher finds something worth chasing.
The memory. Results land in two places on purpose. Amazon DynamoDB keeps a compact, queryable index of every observation, which is what lets the dashboard answer "show me every storm from the last three hours" instantly. Amazon S3 keeps the full research archive, holding the score, the exact radar file reference, the storm's complete property set, and data-quality details. The index is for finding. The archive is for proving.
Live track of a storm in Wisconsin with associated hazard probabilities and issued alerts alongside.

Every piece scales down when the weather is quiet, which put the design target in only paying for what you need instead of an always-on server bill. Here is one table as a recap.
Need
AWS service
Role
Start a chase cycle every five minutes
Amazon EventBridge
The timer
Decide which storms to chase
AWS Lambda (dispatcher)
The scout
Keep radar sites independent
Amazon SQS
The waiting line
Download and prepare radar
AWS Lambda (worker)
The heavy lifting
Run the machine-learning model
Amazon SageMaker
The model host
Find recent storms quickly
Amazon DynamoDB
The searchable index
Preserve detailed records
Amazon S3
The research archive

Riding along in 3D

A user can view a live dashboard where they can see current and past storm tracks across the country, check which storms have active National Weather Service alerts, and select any storm to inspect it up close.
The above rendering shows a storm approaching the Outer Banks (OBX) of eastern North Carolina, with reflectivity indicating the location of the precipitation.

The centerpiece is a 3D rendered playback. The dashboard rebuilds the storm's radar volume in three dimensions, so you can rotate around it, compare different heights inside it, and switch between properties like precipitation intensity and wind velocity. Playback controls let you pause, step scan by scan, scrub the timeline, change speed, and loop the moments where rotation appears. The view stays centered on the storm as it moves, the way a chaser follows the storm rather than parking next to the radar station.
This is the part a real chase cannot give you. A chaser on the ground sees one side of a storm. Here, you can circle it, look inside it, and replay it. I like to think of it like looking out the window of a plane and seeing the expanse of clouds spread out below you.
The above 3D storm velocity rendering captures a tornado vortex signature (TVS), which indicates an area of significant rotation favorable for a tornado. This was a confirmed tornado in Appleton, Wisconsin on 7/27/26.

Getting the visuals right took real testing. My first version flipped the storm from north to south, and another view reversed the apparent direction of rotation. I caught both by rendering known test patterns through the same code path.

What happened after the storm?

After storm events have occurred, a post-storm research component takes over. Its job is to answer the question a chaser answers by sight. Did a tornado actually happen?
The flow mirrors the live side. When a storm disappears from NOAA's feed, a tracker Lambda closes out its history and places the completed storm on another SQS queue. From there, a scheduled research Lambda checks back on each storm as its evidence matures, using a DynamoDB index to know which storms are due for a look. The schedule follows how the real world publishes evidence. Preliminary storm reports can appear within hours, official damage surveys within days, and the final national storm-events archive can take weeks or months. Each storm's record is upgraded as stronger evidence arrives.
Above is the post-storm analysis component of the dashboard. The storm selected has been automatically marked up with reported tornado sightings and storm damage assessments.

I held this component to a strict standard. Warnings, radar rotation, and model scores never count as confirmation. Only official records and surveyed damage do, and if the final records are not in yet, the storm stays marked unresolved instead of being labeled "no tornado."

Final thoughts

Each piece started as a question. Could I follow storms automatically? Could I see one in three dimensions? Could I find out what actually happened afterward? The cloud turned each question into a running experiment within days. I watched it carry my own use case, one curious question at a time.
If you have a passion outside of tech, point your builds at it. You will learn the services faster because you care about the answers, and you may walk away with things you didn't initially plan on!
Cheers,
Jonah
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article