Deep-dives into tuning WRF and RASP for soaring - physics, resolution, and validation.
4 km resolves the mesoscale. 2 km starts resolving the terrain that actually triggers thermals. It also costs roughly eight times the compute. Here is the trade, honestly, before any of it has been benchmarked on the live box.
Every previous optimisation post says 'we'll measure this when validation lands'. The first piece of that pipeline shipped this month - a head-to-head diff between two cycle runs at named pilot sites. Here is what it does, what it does not, and why it was the right first step.
The first step of any WRF cycle is downloading the global model fields that drive it. Pulling full GFS files takes a long time. Pulling only the subset we need takes about two minutes. Here is how the live ingest works.
WRF's radiation schemes are expensive, so they run on a longer interval than the dynamics. The shipping interval is 30 minutes. Here is the trade and what we would test before changing it.
Global weather models stop being useful for soaring long before you reach the thermal layer. Here's why Convek runs WRF at 4 km instead of serving GFS like everyone else.
Comparing a full-UK 4 km box against a tighter trimmed domain on the same cycle, the trimmed run finished WRF in about 13 minutes instead of nearly an hour. Here is the trade and why we made it.
We do not yet have an automated verification pipeline. Here is what we currently rely on to know whether the forecast is working, what we plan to build, and how to spot a bad day in the meantime.
We do not have automated validation against radiosonde soundings or XContest flight activity yet. Here is what we plan to build, and why it is non-trivial.
WRF performance tuning is a deep rabbit hole. Here is what we know about our MPI setup, what we have measured directly, and what is still educated guessing from sensible defaults.
The first few hours of any WRF run are contaminated by the model adjusting to its initial conditions. Here is how we think about it for the UK pipeline today, and what we plan to change.
We run the original Noah land surface model in the Convek pipeline, not the more sophisticated Noah-MP. Here is why, what it costs us, and what would push the upgrade onto the roadmap.
Soil moisture in the live pipeline comes from GFS, which can lag reality after fast-moving wet fronts. Switching to ERA5 reanalysis would help. Here is the trade and where it sits in the queue.
Cloudbase in the API today comes from a surface-parcel LCL using 2 m temperature, 2 m mixing ratio, and surface pressure. Here is what that gets right, what it gets wrong, and where it is going.
Today's cloudbase comes from a surface-parcel LCL. The standard upgrade is a mixed-layer LCL averaged over the diagnosed PBL depth. Here is what that would change and what is in the way.
We ship the YSU planetary boundary layer scheme. We picked it the way most operational regional setups do - from the published evidence on what works at this resolution. Here is the reasoning.
We ship YSU. MYNN-2.5 is the most credible alternative for daytime convective forecasting. Here is the comparison we plan to run, and what makes it worth running rather than picking from the published evidence alone.
WSM3 is the cheapest microphysics scheme in WRF. It is also what we ship. Here is why that was the right starting point and what would push us to change it.
We ship WSM3 today. Thompson is the obvious next-step microphysics for soaring forecasting. Here is what we would want to validate before paying the extra runtime cost.
WRF gives you a vertical stack of eta levels to spend however you like. For soaring, almost all of them should live in the lowest few kilometres.
The first decision in any WRF setup is where the model sees air. Here is how the Convek UK domain is laid out and why the boundaries sit where they do.
