MWmetalhead wrote: ↑Tue Feb 02, 2021 10:26 pm
Outstanding insight; thank you for posting.
I recall hearing recently that the GFS's equations recently (within the last year, let's say) received a major update. Is that correct, do you know?
Thank you!
I am not a model developer, nor am I actively participating in model development, so I can't speak to well to the specifics, but I'll try my best.
You are correct, the GFS did a major change in its computational methods. Prior to 2019, the GFS used spectral methods to solve the basic physical equations. Spectral methods typically use Fourier transforms, which fit a series of sinusoidal equations to estimate a solution (to visualize this, think if you were to model the ocean surface, you'd see a series of small waves embedded within larger waves extending to diurnal and monthly lunar tidal patterns). This method has worked reasonably well over the years and requires fewer individual computations to run, but there is a limitation to its accuracy.
In 2019, the GFS now adopts a Finite Volume Cubic-sphere (FV3) grid. That is, it assumes the Earth's atmosphere is of finite volume (a fair assumption, since we're neither gaining nor losing atmosphere to any reasonable approximation; unlike Mars, which is losing atmosphere due to a weaker magnetic field leaving the planet more exposed to the impacts of solar wind). It then partitions the Earth into a cubed sphere, pictured below (image from NASA, but the cubed sphere grid was developed at NOAA's Geophysical Fluid Dynamics Laboratory at Princeton.
It's a strange shape, but dividing a sphere into roughly equal parts at high spatial resolution, yet maintaining lower computational costs and mitigating numerical instabilities is a complex problem. The promise of the FV3 grid is that it is supposed to resolve larger scale features better than spectral methods over longer forecasting periods. I don't have the specifics in how much this has improved the GFS's forecast over time, but the modeling community seems happy.
As a brief interesting aside, the big motivation for American forecasting model improvements was Hurricane Sandy. The ECMWF was the first global model to predict a leftward turn, making landfall over the eastern US, whereas the GFS predicted the storm would eject out to see over the North Atlantic. Both models at the time used spectral methods to solve the basic equations, and the ECMWF still uses spectral methods operationally today in their forecast models. However, post Sandy model analysis revealed that the key difference happened to be the data ingested. NOAA tends to be pretty aggressive with quality controlling observations that go into their models; but the ECMWF group tends to be more inclusive with what data they used. When using the same observational data that the ECMWF used, a rerun of the GFS also showed a leftward turn into the east coast.
Given that, it would seem that the FV3 fix seems questionable, but the goal was to create better modeling techniques than the ECMWF group, rather than merely copy them. Other fixes in the future include increasing vertical resolution (adding more model layers in the atmosphere, since vertical processes are extremely important for resolving weather features), better representation of aerosols (important for clouds and heating/cooling), among others.
I'm here for a good, hearty debate, to agree and disagree respectfully, and commiserate on the current state of terrestrial radio.