Cross-Provider Differences: RH, Snowfall, Albedo, Snow Depth¶

Tool: src/weather/tests/compare_providers.py Data: 2018, Arctic-edge test cell (70.5°N, 25.0°E — see that script's default), COSMO-REA6 vs ERA5-Land vs MERRA-2. Status: Quantified from real 2018 output; physical explanations below draw on the formulas already implemented in this repo (verified) plus general reanalysis/NWP literature (see References — not independently re-verified for this specific pair of products; ask if you want a literature search for a stronger citation on a specific claim). Pending refresh: COSMO-REA6 2018 was re-run live (adds RELHUM_2M -> RH) after the numbers below were computed — the RH section's "ERA5 vs MERRA-2 only" framing and its table are now stale (COSMO RH is available: whole-Europe June 2018 mean 67.7%, range 0.02-100%, verified via a fresh compare_providers.py run) but the detailed monthly table below has not yet been regenerated against the new data; treat the ERA5/MERRA-2 numbers as still valid, the COSMO-exclusion note as outdated. A fresh xlsx/plots run for June 2018 confirms COSMO RH, MERRA-2 SNODP/PRECSNOLAND, and the new SNOW_DEPTH column all populate correctly end-to-end; re-run compare_providers.py and paste updated numbers here when this file needs a full refresh.

These are real, expected differences between three independently-run reanalysis products — not bugs, and not something to "fix" by unifying the formulas (see CLAUDE.md: "Do NOT unify").

Attribute naming reference: raw source -> canonical output¶

Every provider's transform.py renames or derives its raw downloaded attributes into one shared set of canonical output names (e.g. ERA5-Land's fal -> ALBEDO). Where the physical quantity genuinely differs across providers, the raw formula differs too (see CLAUDE.md: "Do NOT unify") but the name is still shared. Current as of 2026-07-30 (full cross-provider naming-unification pass — see .claude/open.md's ## cross-provider entry for the complete history of what was renamed, when, and why, including two real mislabeling bugs found and fixed along the way: COSMO used to drop its raw wind components, and ERA5-Land's snow_depth entry was mapped to the wrong CDS variable).

Canonical	COSMO-REA6 source	ERA5-Land source	MERRA-2 source
`T`	`T_2M` (K -> degC)	`t2m` (rename)	`T2M` (rename)
`T_DEW`	`T_2M`+`RELHUM_2M` (derived, inverse Magnus-Tetens — no native field exists)	`d2m` (rename, native)	`T2MDEW` (rename, native)
`GHI`	`SWDIFDS_RAD`+`SWDIRS_RAD` (sum, clipped)	`ssrd` (de-accumulated, ÷3600)	`SWGDN` (night-masked, already instantaneous)
`DHI`	`SWDIFDS_RAD` (native, exact)	— (bulk not computed; point-of-use via `dni_pointwise.py`)	— (bulk not computed; point-of-use via `dni_pointwise.py`)
`DNI`	`SWDIRS_RAD`/cos(θz) (native, experimental)	— (point-of-use only)	— (point-of-use only)
`RH`	`RELHUM_2M` (rename, direct measurement)	`t2m`+`d2m` (Magnus formula)	`QV2M`+`PS`+`T2M` (Bolton 1980 formula)
`WS_10M`	`U_10M`+`V_10M` (sqrt(u²+v²))	`u10`+`v10` (sqrt(u²+v²))	`U10M`+`V10M` (sqrt(u²+v²))
`U_10M`/`V_10M`	`U_10M`/`V_10M` (kept as-is, native)	`u10`/`v10` (rename)	`U10M`/`V10M` (rename)
`ALBEDO`	`SWDIFDS_RAD`+`SWDIRS_RAD`+`SOBS_RAD` (derived: `(GHI-SOBS_RAD)/GHI`)	`fal` (rename, native "forecast albedo")	`ALBEDO` (native, unchanged)
`SNOWFALL`	`SNOW_CON`+`SNOW_GSP` (sum)	`sf` (rename)	`PRECSNOLAND` (rename, kg/m²/s -> kg/m²/h)
`SNOW_DEPTH`	`H_SNOW` (rename)	`sde` (rename)	`SNODP` (rename)
`PS`	`PS` (unchanged)	`sp` (rename)	`PS` (unchanged)

Provider-unique fields (no cross-provider equivalent, kept under their own raw name, not part of the canonical set above):

Field	Provider	What it is
`asn`	ERA5-Land only	ECMWF "snow albedo" — reflectivity of just the snow-covered fraction, narrower than `ALBEDO`'s whole-grid-cell blend
`QV2M`	MERRA-2 only	2 m specific humidity (kg/kg) — the raw input `RH`'s Bolton formula is derived from; kept in the output alongside the derived `RH`
`U_2M`/`V_2M`	MERRA-2 only	2 m wind components (renamed from `U2M`/`V2M` for internal consistency, no derived scalar computed from these)
`U_50M`/`V_50M`	MERRA-2 only	50 m (hub-height) wind components (renamed from `U50M`/`V50M`) — added for a confirmed downstream wind-power consumer, see `.claude/merra2/merra2_context.md`

1. Relative humidity: MERRA-2 reads ~6 points higher than ERA5-Land¶

What's different¶

Each provider computes RH from different source fields via a different formula — this is by design, documented in CLAUDE.md:

Provider	RH source	Formula family
COSMO-REA6	`RELHUM_2M` (direct model output)	none — RH is a native model field
ERA5-Land	2 m dew-point (`d2m`) + 2 m temperature	Magnus formula
MERRA-2	2 m specific humidity (`QV2M`) + `PS`, `T2M`	psychrometric (specific-humidity-based)

(COSMO has no RH in the currently-generated 2018 files for an unrelated reason — see .claude/open.md — so this comparison is ERA5-Land vs MERRA-2 only.)

What the data shows¶

Monthly mean RH, MERRA-2 minus ERA5-Land, at the test cell:

Month	ERA5-Land	MERRA-2	Diff (M−E)
1	81.28	91.06	+9.78
2	81.14	90.42	+9.28
3	80.72	83.06	+2.34
4	84.87	88.57	+3.70
5	77.24	86.92	+9.68
6	79.67	85.55	+5.88
7	79.39	85.99	+6.60
8	83.95	89.47	+5.52
9	85.38	91.22	+5.84
10	88.29	92.89	+4.60
11	88.23	92.90	+4.67
12	84.83	91.50	+6.66

MERRA-2 reads higher every single month — annual mean gap +6.2 pts (std 2.3 pts across months). This is a consistent bias, not noise.

The gap also has its own diurnal cycle, not just a flat offset — full-year mean by hour-of-day:

Smallest gap (~4.9–5.3 pts): 03:00–06:00 UTC (near the daily temperature minimum)
Largest gap (~7.1–7.3 pts): 14:00–19:00 UTC (near the daily temperature maximum)
Range across the day: 2.4 pts

Why¶

RH depends strongly on temperature (via saturation vapor pressure), and the two formulas propagate a temperature swing differently:

ERA5-Land (Magnus, dew-point-based): RH = f(T, T_dewpoint). Both T and T_dewpoint respond to the diurnal cycle, but not identically — daytime heating drives T up faster than T_dewpoint typically follows, widening the T − T_dewpoint spread and pulling RH down more at midday.
MERRA-2 (specific-humidity-based): RH is derived from a near-conserved quantity (specific humidity QV2M doesn't track T directly) divided by a saturation vapor pressure that does track T. The result has a different, slightly damped diurnal sensitivity, and a different absolute-value convention baked into MERRA-2's own internal humidity diagnostics.

The two formulas are correct for what they're computing; they're simply not computing bit-identical physical quantities from bit-identical inputs. A ~5–10 point spread between independently-formulated RH diagnostics is consistent with what's reported in reanalysis intercomparison literature (see References).

2. Albedo: ERA5-Land reads higher than MERRA-2, most in the snow-transition months¶

What's different¶

Provider	Field	What it represents
COSMO-REA6	`ALBEDO`	`(GHI - SOBS_RAD) / GHI`, `GHI = SWDIRS_RAD + SWDIFDS_RAD` — built as of the `SOBS_RAD` addition (see `cosmo_rea6/transform.py::compute_albedo`); NaN at night. Not yet incorporated into the `compare_providers.py` table below (needs a rerun against a COSMO archive with `SOBS_RAD`)
ERA5-Land	`ALBEDO` (renamed from cfgrib's `fal`)	"forecast albedo" — total background reflectivity, bare land + variable snow cover combined
MERRA-2	`ALBEDO`	native surface albedo (M2T1NXRAD collection)

What the data shows¶

Monthly mean albedo (0–1 fraction) at the test cell:

Month	ERA5-Land	MERRA-2	Diff (E−M)
1	0.535	0.457	+0.077
2	0.522	0.463	+0.059
3	0.583	0.484	+0.099
4	0.525	0.432	+0.093
5	0.294	0.197	+0.097
6	0.153	0.138	+0.015
7	0.152	0.139	+0.013
8	0.148	0.134	+0.014
9	0.143	0.135	+0.008
10	0.197	0.162	+0.035
11	0.298	0.241	+0.056
12	0.516	NaN	n/a

(December MERRA-2 was NaN at this cell in the current 2018 data — not investigated further here.)

Pattern: the two products nearly agree during the snow-free summer (Jun–Sep, bare-ground albedo, gap 0.01–0.02) and diverge most during the snow-transition months (Mar–May and Nov–Jan, gap 0.06–0.10) — exactly when the fraction of the grid cell covered by snow vs bare ground is most sensitive to small differences in each product's snow-cover/melt timing and sub-grid representation.

Why¶

fal is explicitly a combined land+snow reflectivity in ERA5-Land's own documentation (see downloaded_attributes.py's description), while MERRA-2's ALBEDO is that product's own land-surface-model diagnostic. Both are physically the same quantity in principle, but each product's snow-cover fraction, snow age/depth, and land-surface scheme differ, so they agree closely only when the surface state is unambiguous (all snow or all bare ground) and diverge when the cell is a partial/transitional mix — precisely the pattern seen above.

3. Snowfall: same monthly totals in the same ballpark, but very different hourly timing¶

What's different¶

Provider	Field	Meaning
COSMO-REA6	`SNOW_GSP` + `SNOW_CON` (combined into one `SNOWFALL` field 2026-07-26, see below)	stratiform + convective snow, hourly-accumulated (kg/m²)
ERA5-Land	`sf` (renamed `SNOWFALL` 2026-07-26)	snowfall rate, hourly-accumulated (kg/m²/h)
MERRA-2	`PRECSNOLAND` (renamed `SNOWFALL` 2026-07-26)	snowfall rate over land, converted kg/m^2/s -> kg/m^2/h (the `M2T1NXLND` collection, added and live-verified this session — see `.claude/open.md`)

All three renamed to the shared canonical SNOWFALL name as of 2026-07-26 (see .claude/open.md's ## cross-provider naming-unification entry); the already-completed 2018 archive this section's numbers come from still has the OLD names above, unaffected by the rename.

What the data shows¶

Monthly accumulated snowfall (kg/m², summed over the month) at the test cell:

Month	COSMO sum	ERA5-Land sum
1	22.68	26.27
2	17.72	21.84
3	54.82	67.42
4	31.32	33.18
5	0.00	2.98
6	0.71	2.77
7	0.00	0.00
8	0.00	0.00
9	1.20	2.81
10	9.28	18.24
11	18.35	37.99
12	36.83	32.63

Totals are in the same rough range (ERA5-Land usually somewhat higher, notably ~2x COSMO in October/November) — but the hourly timing barely correlates:

Hourly correlation across the full year (nearest-timestamp matched): r = 0.39
COSMO reports "snowing" (SF > 0) in 19.7% of hours; ERA5-Land in 46.0% of hours — more than double
Jaccard overlap (hours where both say snowing / hours where either says snowing): 0.44 — the two products disagree about whether it's snowing right now more often than they agree

Why¶

The monthly totals landing in the same range while the hourly pattern disagrees this much points to a temporal distribution difference, not a magnitude difference. COSMO-REA6 (~6 km, convection-permitting) and ERA5-Land (driven by ERA5's own coarser ~31 km atmospheric forcing, no independent convection scheme of its own) run genuinely different dynamics and microphysics, so they place precipitation events on different hours even for the same real month. This general mechanism (coarser/parent-driven atmospheric models spreading precipitation across more hours at lower intensity than a convection-permitting downscaling of the same period) is directly supported by a study that compares a convection-permitting COSMO-CLM downscaling of ERA5 against ERA5 itself over Italy (Adinolfi et al. 2023, in References): the convection-permitting model showed clear gains in reproducing hourly precipitation characteristics in mountainous areas, though with losses in coastal/flat areas — a nuanced result, not a blanket "COSMO always better." Directionally consistent with what's measured above (ERA5-Land snowing nearly 2.4x as often as COSMO), but this is still a different model pair (COSMO-CLM/ERA5 over Italy, not COSMO-REA6/ERA5-Land over the Arctic) — treat as supporting evidence for the general mechanism, not a direct validation of this specific comparison.

4. Snow depth: all three providers are physical depth (corrected 2026-07-30)¶

What's different¶

Provider	Field	Meaning
COSMO-REA6	`H_SNOW` (renamed `SNOW_DEPTH` 2026-07-26)	physical snow depth, meters
MERRA-2	`SNODP` (renamed `SNOW_DEPTH` 2026-07-26)	physical snow depth, meters
ERA5-Land	`sde` (renamed `SNOW_DEPTH` 2026-07-30)	physical snow depth, meters

Correction (2026-07-30): an earlier version of this section, and of era5_land/downloaded_attributes.py's snow_depth entry, claimed ERA5-Land's field was meters of water equivalent, genuinely different from the other two, and left it deliberately unrenamed. That claim was wrong. Verified via three independent sources: the real CDS request payload (variable snow_depth, not snow_depth_water_ equivalent), the raw GRIB's own decoded metadata (GRIB_shortName 'sde', long_name 'Snow depth' — not 'sd', which is the OTHER, water-equivalent variable), and the already-processed 2018-03 output (the variable was literally named sde, not sd). All three providers' snow depth is the same physical quantity after all, now renamed to the shared canonical SNOW_DEPTH. Importantly, this mislabeling never affected any actual downloaded/exported VALUE — the water-equivalent -> physical-depth conversion the old entry described was never actually implemented in era5_land/transform.py either, so every number below has always been correct; only the description and the naming decision were wrong.

What the data shows¶

Whole-Europe domain stats, June 2018:

Provider	Field	Mean (m)	Max (m)	NaN fraction
COSMO-REA6	`H_SNOW`	0.00	40.00	0.00
MERRA-2	`SNODP`	0.00	0.53	0.00
ERA5-Land	`sde`	0.02	33.33	0.49 (ocean cells — ERA5-Land is land-only, this NaN fraction matches every other ERA5-Land attribute in the same domain-stats table, not a data problem)

All three are now directly comparable — no conversion needed.

June is close to snow-free at most of the Europe domain except high terrain and the far north (COSMO's max of 40 m physical depth is almost certainly a high-Alpine or Scandinavian-mountain outlier cell, not typical). A winter month (e.g. January) would show a much larger, more informative spread and is the natural next data point if this comparison needs to go further.

Why¶

Not run yet as a quantified difference (unlike sections 1-3 above). Now that all three are confirmed to be the same physical quantity (see the 2026-07-30 correction above), any remaining spread between them would be a genuine modeling difference in the three products' actual snowpack simulation, worth a real comparison once a winter-month snapshot exists.

5. GHI/DHI/DNI: COSMO stores an exact direct/diffuse split; ERA5-Land and MERRA-2 don't¶

What's different¶

Provider	GHI	DHI	DNI
COSMO-REA6	`SWDIFDS_RAD + SWDIRS_RAD`	`SWDIFDS_RAD` (native, exact)	`SWDIRS_RAD / cos(θz)` (native, exact — not a GHI decomposition)
ERA5-Land	de-accumulated `ssrd`	pvlib DIRINT decomposition of GHI (only option — no native split)	pvlib DIRINT decomposition of GHI
MERRA-2	`SWGDN` (already instantaneous)	pvlib DIRINT decomposition of GHI	pvlib DIRINT decomposition of GHI

This is the single biggest methodological difference between COSMO and the other two: COSMO-REA6's regional model separately simulates direct and diffuse shortwave as distinct physical quantities, so its DHI/DNI are as trustworthy as its GHI. ERA5-Land and MERRA-2 only ever store GHI in bulk — DHI/DNI have to be decomposed from GHI alone via pvlib's DIRINT algorithm (era5_land/dni_pointwise.py, merra2/dni_pointwise.py), a fundamentally lossier operation with no independent direct/diffuse information to draw on.

What the data shows¶

Quantified directly against COSMO's own native DNI/DHI (the one provider where a ground truth exists in this dataset) at the test cell, comparing COSMO's native values to a DIRINT decomposition of COSMO's own GHI (i.e. simulating what ERA5-Land/MERRA-2 are stuck doing) vs. the exact closure formula (GHI-DHI)/cos(θz):

Estimate	bias	MAE	RMSE	r
`DNI_pvlib_closure` (exact, known DHI)	+0.15 W/m^2	0.90 W/m^2	7.19 W/m^2	0.9992
`DNI_pvlib_dirint` (GHI-only decomp.)	+11.53 W/m^2	24.82 W/m^2	43.81 W/m^2	0.9717
`DHI_pvlib_dirint` (GHI-only decomp.)	-2.15 W/m^2	8.69 W/m^2	16.87 W/m^2	0.9835

The exact closure formula (used when DHI is already known, as with COSMO) is essentially noise-level (RMSE 7 W/m^2). The GHI-only DIRINT decomposition — the only option ERA5-Land/MERRA-2 have — is ~6x worse by RMSE. This is a real decomposition-model error inherent to not having a direct/diffuse split, not a solar-position or implementation bug (full derivation, the SWDIRS/SWDIFDS-direct/diffuse assignment verification, and the swap test that confirmed it in docs/dni_methodology.md).

Why, and where to read more¶

This is covered in full depth (methodology derivation, the SWDIRS_RAD/ SWDIFDS_RAD direct/diffuse assignment verified against source documentation, night-masking behavior near the Arctic in summer, and a pvlib-closure-vs-DIRINT validation) in docs/dni_methodology.md — this section is a summary/pointer, not a duplicate. See especially its §4 (why COSMO's split is exact), §11 (the closure-vs-DIRINT validation behind the numbers above), and §9 (COSMO's wind vectors being in the rotated-pole frame, related to §6 below).

6. Wind speed: ERA5-Land reads notably lower than COSMO/MERRA-2 in this domain¶

What's different¶

Provider	Fields	Frame
COSMO-REA6	`U_10M`, `V_10M`	rotated-pole grid, not true north (see below)
ERA5-Land	`u10`, `v10`	true north (WGS84)
MERRA-2	`U10M`, `V10M`	true north (WGS84)

All three compute scalar WS_10M = sqrt(U^2 + V^2), which is rotation- invariant — comparable across all three regardless of frame. Wind direction (atan2(V, U)) is not currently computed by any provider's pipeline; if it ever is, COSMO's U_10M/V_10M would need rotating to true north first (docs/dni_methodology.md sec 9) — ERA5-Land/MERRA-2 need no such correction.

What the data shows¶

Whole-Europe domain stats, June 2018 (single-month snapshot, same run as section 4):

Provider	WS_10M mean (m/s)	WS_10M max (m/s)
COSMO-REA6	4.62	29.13
MERRA-2	4.49	22.25
ERA5-Land	2.58	20.86

COSMO and MERRA-2 agree closely (4.62 vs 4.49); ERA5-Land reads ~45% lower.

Why¶

Plausibly a real, largely domain-composition effect rather than a bug: ERA5-Land is a land-only product (this domain's ~49% NaN fraction in section 4's table is ERA5-Land's ocean mask), so its wind stats exclude the open-water/coastal cells where boundary-layer roughness is lowest and near-surface wind speeds climatologically highest — cells that COSMO-REA6 and MERRA-2 (both covering ocean too) still include. Higher land-surface roughness length also directly suppresses ERA5- Land's own near-surface wind on the land cells it does report. Not independently verified against a land-only-masked recomputation of COSMO/MERRA-2 for this specific domain/month — the mechanism is standard boundary-layer physics, but the magnitude of this particular gap hasn't been decomposed into "domain composition" vs "roughness representation" shares.

7. Temperature and pressure: minor, resolution/orography-driven spread¶

What's different¶

All three read 2 m air temperature and surface pressure essentially directly (T_2M/T, PS for COSMO; t2m, sp for ERA5-Land; T2M, PS for MERRA-2) — no formula-family difference like RH's. Any spread here is attributable to model resolution/orography representation and each reanalysis's own physics, not to a difference in what's being computed.

What the data shows¶

Whole-Europe domain stats, June 2018 (single-month snapshot):

Provider	T mean (degC)	T range (degC)	PS mean (Pa)	PS range (Pa)
COSMO-REA6	17.81	-11.82 to 49.52	99027.50	62621.06 to 107428.22
ERA5-Land	16.77	-7.14 to 44.86	97223.18	70831.94 to 103723.44
MERRA-2	15.37	-5.72 to 46.65	99230.93	76953.53 to 103646.67

Spread is ~2.4 degC across the three (COSMO warmest, MERRA-2 coolest) and ~2000 Pa in pressure (ERA5-Land lowest). COSMO's much wider PS range (down to 62621 Pa, i.e. ~630 hPa) is consistent with its 6 km grid resolving individual Alpine peaks that the coarser 0.1 deg/0.5x0.625 deg grids smooth out — a real, expected orography-representation effect, not a data error.

Why¶

Not investigated further here — this is a much smaller, less operationally significant spread than the RH/albedo/snowfall differences quantified above, so it hasn't received the same dedicated analysis. Flagged for completeness.

8. Timestamp convention: MERRA-2 labels mid-interval, COSMO/ERA5-Land label on the hour¶

MERRA-2's hourly-mean collections (rad, slv) label each timestamp at the midpoint of the averaging interval (HH:30, e.g. 2018-06-01T00:30), since that's what the mean actually represents. COSMO-REA6 and ERA5-Land both label on the hour (HH:00). This is a genuine provider difference, not a bug, and is not corrected/shifted anywhere in this codebase — any code that merges or directly compares MERRA-2 against the other two on a shared time index must handle the 30-minute offset explicitly (resample/interpolate one onto the other's index) rather than assume the raw indices line up. Full detail: docs/MERRA2_PIPELINE_GUIDE.md's "Timestamp convention" section.

9. Grid and resolution: no cross-provider regridding¶

Provider	Native grid	Cell size (approx.)
COSMO-REA6	rotated-pole, 824x848	~6 km
ERA5-Land	regular lat/lon	~9-11 km (0.1 deg)
MERRA-2	regular lat/lon	~35-55 km (0.5x0.625 deg)

tests/compare_providers.py snaps a requested lat/lon to each provider's own nearest native cell independently — it does not regrid/interpolate any provider onto another's grid, so "the same cell" across providers is only approximate (within roughly half a native grid cell of each product, worse for COSMO since its cell must also be located via an analytically-reconstructed rotated-pole projection rather than stored lat/lon — see that script's module docstring). Whole-Europe domain stats (sections 4, 6, 7 above) are likewise computed independently per provider on its own native grid, so a domain mean is not a like-for-like average over identical physical area between providers. Cross-provider regridding, if ever needed, is a separate, not-yet-built future task.

References¶

Confirmed via live search on 2026-07-24 (titles, authors, journal, volume/pages, and DOI checked against the publisher/ADS record for each; not just recalled from training knowledge):

Bollmeyer, C., Keller, J. D., Ohlwein, C., Wahl, S., Crewell, S., Friederichs, P., Hense, A., Keune, J., Kneifel, S., Pscheidt, I., Redl, S., & Steinke, S. (2015). Towards a high-resolution regional reanalysis for the European CORDEX domain. Q. J. R. Meteorol. Soc., 141(686), 1–15. https://doi.org/10.1002/qj.2486 (COSMO-REA6)
Muñoz-Sabater, J., et al. (2021). ERA5-Land: a state-of-the-art global reanalysis dataset for land applications. Earth Syst. Sci. Data, 13(9), 4349–4383. https://doi.org/10.5194/essd-13-4349-2021 — confirms ERA5(-Land) 2 m RH is derived from temperature + dew point via a Magnus-type formula (ECMWF's own internal implementation uses coefficients a=17.502, b=240.97 — slightly different from this pipeline's a=17.625, b=243.04, a different well-established Magnus-type variant (Alduchov & Eskridge 1996); both are valid, this pipeline's RH is not bit-identical to ECMWF's own archived humidity fields if any exist).
Gelaro, R., et al. (2017). The Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2). J. Climate, 30(14), 5419–5454. https://doi.org/10.1175/JCLI-D-16-0758.1
Adinolfi, M., Raffa, M., Reder, A., & Mercogliano, P. (2023). Investigation on potential and limitations of ERA5 Reanalysis downscaled on Italy by a convection-permitting model. Climate Dynamics, 61, 4319–4342. https://doi.org/10.1007/s00382-023-06803-w — directly relevant: COSMO-CLM (same model family as COSMO-REA6) at convection-permitting scale vs. its parent ERA5, precipitation frequency/intensity specifically evaluated.
Sun, Y., Solomon, S., Dai, A., & Portmann, R. W. (2006). How often does it rain? J. Climate, 19(6), 916–934. https://doi.org/10.1175/JCLI3672.1 — general (not COSMO/ERA5- specific) "drizzle bias" finding: coarser models overestimate light- precipitation frequency while broadly matching its intensity pattern.

To regenerate or extend this analysis (different cell/month, or add weekly/diurnal breakdowns for other attributes), see tests/compare_providers.py's RobustnessChecker and the monthly_summary()/weekly_summary()/diurnal_summary() methods it already computes.