YSU vs MYNN-2.5: the PBL scheme bake-off
The planetary boundary layer scheme is the single biggest lever on soaring forecast quality. We tested YSU against MYNN-2.5 across a UK convective season. Here is what won and why.
The PBL scheme decides how the boundary layer grows through the day. It is the piece of the model most directly responsible for everything a pilot cares about: boundary layer height, surface heat flux, low-level winds, thermal triggering, moisture mixing up into the cumulus level. If you change nothing else, changing the PBL scheme will move your forecast more than changing any other single piece of the physics.
There are roughly a dozen PBL schemes in WRF. For soaring forecasting, two are serious candidates. YSU (Yonsei University) is a non-local first-order scheme: it represents large convective eddies as a bulk transport term that reaches through most of the PBL in one step. It is the workhorse of a lot of operational regional forecasting. MYNN-2.5 is a local TKE-closure scheme: it evolves a prognostic turbulent kinetic energy field and mixes properties based on the local TKE. More expensive, nominally more physical.
The consensus in the academic literature is that YSU produces deeper, more mixed boundary layers and that MYNN-2.5 produces shallower, more structured ones. For soaring, deeper is better up to a point (more height to work with) but over-deepening gives you overconfident forecasts that under-deliver in reality. We wanted data, not vibes.
The test set was 38 UK convective days across summer 2025, picked from the days when Larkhill and Herstmonceux launched daytime radiosondes (12Z) and when XContest had enough flying activity to give us a verification signal. Two model configurations identical apart from the PBL scheme. Everything else held constant: Thompson microphysics, Noah-MP, RRTMG, same domain, same vertical levels, same GFS initialisation.
On boundary layer height against the 12Z Larkhill sounding, YSU had a mean bias of +120 m (slightly too deep) with RMS error of 290 m. MYNN-2.5 had mean bias of -180 m (too shallow) with RMS error of 340 m. YSU wins on both absolute bias and scatter.
On timing of convective onset, matched against pilot-reported first thermals, YSU was early by an average of 25 minutes, MYNN was late by an average of 45 minutes. YSU's earliness is a known feature of non-local schemes: the bulk transport term fires as soon as surface flux goes positive, whereas MYNN has to build TKE first. For pilots, 25 min early is more forgivable than 45 min late (you can wait on the hill, you cannot un-miss the start of the day).
On cloudbase, surprisingly little difference. Both schemes were within 300 ft of pilot reports on average, which suggests the Thompson microphysics is carrying most of the weight there and the PBL scheme is upstream of the error we care about.
On cost, YSU is roughly 8 to 12 percent faster than MYNN-2.5 at our configuration. Not a huge factor, but meaningful when you are running four cycles a day.
YSU wins. It is now the scheme in the shipping namelist. There are days where MYNN-2.5 is visibly better, usually light-wind days with weak convection where the local scheme captures the structured small-thermal regime more faithfully, but the average across a season favours YSU. Once we are running multiple regional domains, revisiting this on continental-Europe days (stronger convection, different terrain) is on the list.
The raw bake-off data will go into a research note later in the year. For now, if you are running a forecast for the UK and wondering which scheme to pick, the answer is YSU. The WRF model page has the full namelist choices.