All Activity

The NC weenie run

AI GFS bumped south at 12z...still brings precip north, but no where near as much as before. Confluence way stronger than 6z

GFS and ECMWF Core vs AI details: (AI) just for reference ECMWF’s AI system (AIFS) is a fully separate, AI-native forecast model, not the traditional ECMWF model with machine learning layered on top. Unlike the main Integrated Forecasting System (IFS), AIFS does not solve physical equations; it uses deep learning trained on decades of reanalysis and operational forecast data to predict the next atmospheric state directly. It ingests similar inputs (pressure, temperature, wind, humidity fields) but produces forecasts via neural networks rather than numerical integration. The operational IFS remains the authoritative, physics-based backbone for warnings, ensembles, and high-impact decision support. ECMWF runs AIFS in parallel, comparing skill, speed, and bias against IFS at multiple lead times. In medium-range forecasts (≈3–10 days), AIFS has shown skill comparable to—and in some metrics better than—the full physics model, at a fraction of the computational cost. Key takeaways Separate model: AIFS is independent from IFS, not a post-processing or hybrid add-on Massively faster: Enables rapid global forecasts and large ensembles with minimal compute Complementary role: IFS provides physical realism and extremes; AIFS adds speed and pattern skill AIFS (AI Forecasting System): Research & development: 2022–2023 First real-time parallel runs: mid-2023 Publicly released experimental real-time forecasts: 2024 Status today: Operational parallel model (not replacing IFS) ECMWF was the first major global center to run an AI-native global model continuously in real time. -------------------------------------------------------------------------------------------------------------------------------------- NOAA’s GFS AI efforts are not a single AI-native replacement model like ECMWF’s AIFS, but a combination of experimental AI models and AI-enhanced components run alongside the main GFS. The operational GFS remains a fully physics-based numerical weather prediction system using the FV3 dynamical core. NOAA’s AI models are trained on decades of global reanalysis and past GFS outputs to forecast future atmospheric states directly, without explicitly solving physical equations. These AI forecasts are run in parallel to GFS for evaluation, research, and medium-range pattern guidance rather than operational warnings. In addition, machine learning is increasingly used inside the GFS workflow to improve data quality control, parameterizations, and bias correction. The strategy prioritizes reliability and explainability, with AI enhancing—but not replacing—the core physics model. Key takeaways Hybrid approach: Physics-based GFS plus separate experimental AI models No AI-only operational GFS: AI runs inform and augment forecasts, not replace them Risk-aware path: Emphasis on safety, interpretability, and gradual adoption NOAA (GFS + AI) GFS (physics model): Operational since the 1980s (FV3 core since 2019) NOAA AI global models: Research pilots (ML weather prototypes): 2021–2022 Sustained real-time parallel AI runs: 2023 Expanded evaluation vs GFS: 2024–2025 Status today: Experimental / evaluation only (not operational) NOAA deliberately moved slower and more conservatively than ECMWF. ------------------------------------------------------------------------------------------------------------------------------------ At 96h (Day 4), ECMWF’s physics-based IFS/HRES and the AI model (AIFS) are often very close on large-scale “pattern” skill (e.g., 500-hPa height), with AIFS frequently matching IFS despite being cheaper to run. At 144h (Day 6), verification shown by ECMWF commonly has AIFS slightly ahead on upper-air headline scores (ACC/RMSE for fields like 500-hPa geopotential height), which is exactly where ML models tend to shine. At 168h (Day 7), AIFS’ advantage can persist for broad circulation patterns, but that does not mean it’s “better at everything.” IFS still tends to win on fine-scale/high-impact details that depend on explicit physics and higher resolution (e.g., local precipitation structure, boundary-layer processes), while AIFS can be more “smooth” because it’s learned and typically coarser. Importantly, AIFS is its own model (AI-native), but it uses the same initial conditions as IFS (from ECMWF’s assimilation/analysis pipeline), so the comparison is about the forecast engine, not the starting point. Bottom line: by Day 4/6/7, AIFS often matches or edges IFS on upper-air pattern skill, while IFS remains stronger for physically constrained, local extremes and “weather sensible” detail. 96h: usually “tie-ish” on big patterns; IFS often better on local detail 144h: AIFS commonly slightly better on headline upper-air skill (pattern metrics) 168h: AIFS can stay ahead on large scales, but IFS remains the safer bet for extremes/structure ---------------------------------------------------------------------------------------------------------------------------- NOAA’s physics-based GFS still outperforms its experimental AI models at shorter lead times, especially where mesoscale detail and physical constraints matter. At 96 hours (Day 4), the operational GFS generally verifies better against observations for precipitation placement, fronts, and boundary-layer–driven features, while AI runs are competitive mainly on broad upper-air patterns. By 144 hours (Day 6), NOAA’s parallel AI models often match GFS on large-scale circulation metrics (e.g., 500-mb height anomalies) but begin to diverge more in sensible weather details. At 168 hours (Day 7), AI models can sometimes show similar or slightly better pattern skill than GFS, reflecting strengths in learned climatology and flow regimes. However, GFS remains more reliable for physically rare or extreme events (strong cyclogenesis, sharp gradients, convective outbreaks). Because NOAA’s AI systems are still experimental and not ensemble-anchored, GFS remains the authoritative model for verification, warnings, and downstream products. Key takeaways 96h: GFS usually verifies better overall; AI competitive mainly on upper-air patterns 144h: AI ≈ GFS on large scales; GFS better on precipitation and structure 168h: AI can edge pattern scores, but GFS remains stronger for extremes and decision use Highest overall pattern skill (global): GraphCast ECMWF AIFS FourCastNet / Pangu-Weather Best operational reliability: ECMWF IFS NOAA GFS Best short-range precipitation AI: MetNet-3

GFS is a big hit

It's cool dude, I honestly don't mind...and it's kind of a relief. I'll move to Euro pbp since I get it super early. The only reason I did GFS pbp was because I got it early...but now we all get it at the same time, so no point in that. And I can relax and sit back for that one. Now let's reel this mfer home yall

@138 GFS slp about 200-250 WNW of 6z. I'm ok w/ that move. Thats 2 +'s so far.

Man, that is a LONG fetch of frozen precip on the 12z GFS. At 129, snow stretches from southwest New Mexico to Wilmington, NC. Ice sits directly south of that line. Textbook over-running setup.

I received at least 3 inches. Melville.

12z GFS is a reasonable "fail" scenario where the shortwave doesn't eject cleanly -AND- the confluence to the north is stronger. Even with that in mind, it looks like a solid plowable to warning level snowfall for the entire subforum with no temp issue - and it's on a weekend.

Looks like precip coming in around 12-1AM Saturday AM.

Decent... Sent from my SM-G998U using Tapatalk

Pretty big run to run change on the GFS at 500, fair to say it isn’t really locked in yet.

GFS is a slight tick NW but the big thing I saw was that precipitation totals were almost doubled due to the Baja low moving slightly more east.

I’ve been reading and interpreting model runs for 15 years, this isn’t textbook for us in many ways. That said, can still score a decent event with a singular stream wave and some tapping of gulf moisture.

High pressure is stronger this run.

Definitely an improvement compared to 06z.. in my opinion

at least there is no Great Lakes Low

Will check in with you

12z GFS has this starting Friday night now...

There’s the option that things come out, and a phase doesn’t happen. Could still give us love, but wouldn’t be the bomb. to be honest, a setup like this…I think we have to be a bit greedy to make up for the misses we’ve had since 2016 in some areas.

Man that GFS was close... Maybe by 0z...

And where the GFS tries to send a slug of moisture up west of the Apps at the last minute....I don't play that game! That feature almost always trends east. But again, probably a good thing that this system doesn't boomerang trend into the mid-state.

Unless we are looking at different runs, that hp is stronger and pushing farther S Can't make this stuff up.

Lmao you're doing great I'm just here ranting about maps. Speaking of which we flirt for so long with a phase before it starts to retrograde back west. Though... I think it might setup for a different storm later with this look.

Still slides just s and e with the good stuff.