Silurian AI

Hurricane Isaias went from a 30 kt tropical depression to a 105 kt Category 3 in about two and a half days. We’ve been running GFT-C, our AI hurricane model, on it every six hours since it formed (including at AOML servers). This chart shows all ten runs against what the storm actually did. A few things stood out to us: - The track held up well. At 24 hours out, the forecasts were off by about 24 miles on average, and every run had the same turn north toward the Pensacola area. - Intensity. Every run made Isaias a major hurricane. The first two runs peaked around 80-90 kt, short of what eventually happened. The Wednesday 1pm run did better: Isaias was a 50 kt tropical storm then, and that run forecast 102 kt for this morning. It came in at 105. Rapid intensification is still one of the hardest problems in hurricane forecasting, and we’re keeping score storm by storm. If you’re on the Gulf coast, follow the National Hurricane Center and your local officials for warnings.

Hurricane Isaias went from a 30 kt tropical depression to a 105 kt Category 3 in about two and a half days. We’ve been running GFT-C, our AI hurricane model, on it every six hours since it formed (including at AOML servers). This chart shows all ten runs against what the storm actually did. A few things stood out to us: - The track held up well. At 24 hours out, the forecasts were off by about 24 miles on average, and every run had the same turn north toward the Pensacola area. - Intensity. Every run made Isaias a major hurricane. The first two runs peaked around 80-90 kt, short of what eventually happened. The Wednesday 1pm run did better: Isaias was a 50 kt tropical storm then, and that run forecast 102 kt for this morning. It came in at 105. Rapid intensification is still one of the hardest problems in hurricane forecasting, and we’re keeping score storm by storm. If you’re on the Gulf coast, follow the National Hurricane Center and your local officials for warnings.

Last week we released radrs, our open-source NEXRAD reader. Today we're writing up on Raystack, the layout radrs uses to get radar data ready for machine learning. NEXRAD scans the atmosphere in volume coverage patterns. The number of sweeps, their elevations, and their timing change with the pattern, and some patterns revisit the lowest sweep partway through a volume. CfRadial stores this as a tree of volumes and sweeps. That works for storage, but it's awkward to batch for training, and it bakes in assumptions about scan timing that a model could learn itself. A Raystack instead flattens the tree. Each radar variable becomes one time-ordered stream of rays, and sweep and volume boundaries are kept as separate metadata streams. Rays are folded into fixed-width returns, so every variable is a plain 2D array with time and geometry on each row. A training sample is then just a range of rows. It can cover a single volume or run across several to follow a storm. We built this during our NOAA collaboration on compact radar representations. The full write-up, with format and batching guides: https://lnkd.in/gYWJbBpk

Silurian AI is taking the stage at The Nest Campus next week during Climate Week NYC. We're one of five semifinalists in the Climate Intelligence & Technology cohort, presenting Wednesday morning at the Javits Center alongside Accelerate Wind, RAMO Earth, Tirion Fleets and Vital Ocean®. At Silurian, we're building Earth-system AI that helps operators understand what weather will do to their specific assets, early enough to act. I'll be at Nest Campus and around Climate Week all week. If you're around and want to catch up, send me a DM. Thanks to The Nest Campus and KITE Scouting for bringing this group together.

We’re open sourcing radrs: a Rust-powered Python library that turns NEXRAD weather radar files directly from cloud storage into xarray-ready data for analysis. Point it at the existing archive and start working with observations. No need for first downloading, converting and maintaining a separate copy of the dataset. During our collaboration with NOAA on compact radar representations, we needed a faster path from radar files to model inputs. Reading and decoding volume after volume adds up across a training dataset. radrs builds on the nexrad Rust project for decoding, with concurrent fetching that overlaps network reads with parsing. It returns familiar xarray structures for sweep-based analysis, or a flat Raystack representation that assembles observations across sweeps and volumes for training while retaining time and spatial metadata. On one KTLX radar volume, radrs loaded all 20 sweeps approximately 17x faster than xradar on an Apple M2 Max. We’re sharing this tooling so other teams can work directly with the public radar archive and spend more time on analysis and modeling. Thank you to the nexrad Rust authors and contributors for the foundation. Available under Apache 2.0 for Python 3.11+. Install with `uv add radrs`. Read the post: https://lnkd.in/gKwiPE_a

Announcing a new #NOAA Partnership! Atlantic Oceanographic and Meteorological Laboratory (AOML) and Silurian AI have signed a #CRADA to develop AI-driven tropical cyclone forecasting. The #CRADA partnership will improve AI-based tropical cyclone tracking and intensity prediction, with the goal of mitigating damage from extreme weather events. Read more at https://lnkd.in/ggwEsnWA. #partnerships #AI #hurricanes

NOAA's Atlantic Oceanographic and Meteorological Laboratory (AOML) and Silurian AI partner on AI-driven tropical cyclone forecasting. Through this collaboration, Atlantic Oceanographic and Meteorological Laboratory (AOML) and Silurian AI will develop a dedicated ML-ready database for tropical cyclone data and refine Silurian’s Generative Forecast Transformer (GFT) to improve tropical cyclone track and intensity prediction. We’re excited to work with AOML on one of the most challenging and consequential forecasting problems in Earth science. Better tropical cyclone forecasts translates into better decisions for communities and operators exposed to storm risk. Read NOAA’s announcement here: https://lnkd.in/ggwEsnWA

The Silurian AI Team will present at #AMS2026 how the Weather Foundation Models are revolutionizing operational forecasting across scales and industry verticals. See you in Houston! #WeatherForecast #FoundationModels #GFT #Silurian #Resilience #PowerGrid #TropicalCyclones