Which storm ingredients tell thunder days apart?
Finished · 14 November 2026
The question. Of the convective fields GEFS gives (CAPE, most-unstable CAPE, lifted indices, CIN, bulk shear, storm-relative helicity, rain), which separate the days airports report thunder from the days they do not?
Data
Summer 2025 (1 June – 31 August), all 31 GEFS members, 00z runs, leads 0–24 h ("day 0") and 72–96 h ("day 3"), on the 0.5° grid. Observations: METAR/SPECI reports from 101 airports in the UK and Ireland (Iowa Environmental Mesonet); a station-day counts as thunder if any report has TS or VCTS in its present weather. 8,905 station-days for day 0, 142 of them with thunder (base rate 1.6 %).
For each field we took, per member, the most thunder-favourable 6-hourly value of the day at the airport's grid cell (largest CAPE, smallest lifted index, and so on), and then the ensemble mean; the score is the ROC area (AUC) of that one number at ranking thunder days above the others. 0.5 is chance; below 0.5 means the field ranks thunder days lower.
Results
| Field (daily extreme, ensemble mean) | AUC day 0 | AUC day 3 |
|---|---|---|
| Best (four-layer) lifted index, 4LFTX | 0.860 | 0.852 |
| Surface lifted index | 0.857 | 0.850 |
| Most-unstable CAPE (255–0 hPa) | 0.850 | 0.831 |
| Largest 6-hour rain | 0.838 | 0.766 |
| Surface CAPE, when wet | 0.788 | 0.770 |
| Surface CAPE | 0.756 | 0.753 |
| Storm-relative helicity | 0.451 | 0.477 |
| Bulk shear (500 hPa − 10 m wind) | 0.414 | 0.442 |
| CIN (least inhibition) | 0.363 | 0.349 |
The same order held over the 3×3 neighbourhood of cells (best LI 0.856, MU-CAPE 0.858, shear 0.413, helicity 0.477, CIN 0.369).
What worked
- Lifted index and most-unstable CAPE. Both rank thunder days well, at day 3 almost as well as at day 0. The best lifted index (the lowest over four layers) is marginally the best single field.
- Rain. Model rain is a good second ingredient: showers that grow tall enough for thunder rain.
What did not
- Shear and helicity scored below 0.5. They are the classic ingredients of organised and severe storms, but most thunder in a British summer comes in slack, weakly sheared air (heat-driven showers under a weak low or a cold pool aloft), while strongly sheared days are often windy, stable and dry. Used as a gate for "any thunder", they would point the wrong way.
- CIN. The design we started from blocked storms where inhibition was strong. CIN's AUC of 0.36 says thunder days in this data usually had some inhibition in the model, so the block would have removed real storms. It was dropped.
- Surface CAPE on its own was clearly weaker than the lifted index or most-unstable CAPE: elevated storms and the model's shallow surface layer both blur it.
What we changed
The first thunder step became "rain in the 6 hours and (CAPE ≥ 200 J/kg or best lifted index ≤ 0)"; shear, helicity and CIN were left out of it (see the storm rule). The calibration that turns this into a chance uses the ensemble-mean lowest best lifted index and the largest 6-hour rain (see thunder nearby).
Caveats
Airport reports cannot tell an isolated storm from an organised or severe one, so this says nothing against shear and helicity as severity ingredients; it only says they do not help to find thunder at all. Testing severity needs another source (lightning density, hail and damage reports).
Scripts: scripts/thunder/extract.py, scripts/thunder/evaluate.py.