# TCOscope validation against CAMS and observations, 2026 YTD

Generated on 2026-08-09. The common forecast-init window is 2026-01-21 through
2026-06-20 (151 daily initializations, T+1 through T+5). The checkpoint was trained
through 2024, validated on 2025, and 2026 is the held-out period.

## Protocol

- TCOscope uses the production Earthformer EMA checkpoint and archived GFS forecasts.
- CAMS uses the 00 UTC `gtco3` forecast at T+24, +48, +72, +96, and +120 hours.
- Gridded comparisons conservatively remap CAMS 0.4 degree fields to the TCOscope
  1.2 degree grid. Area-weighted metrics are the preferred global headline.
- Ground validation uses WOUDC Level-1 daily representative total ozone from Brewer
  and Dobson instruments. TCOscope and CAMS are bilinearly sampled on their native
  grids at station coordinates.
- Bias is forecast minus observation. All units are Dobson Units (DU).

## Overall results

| Truth source | System | Bias | MAE | RMSE | Centered RMSE |
|---|---|---:|---:|---:|---:|
| Mature TROPOMI, area-weighted | TCOscope | -0.80 | 6.80 | 10.63 | 10.60 |
| Mature TROPOMI, area-weighted | CAMS | -8.66 | 11.92 | 16.55 | 14.10 |
| OMPS-NPP L3 daily, area-weighted | TCOscope | +5.34 | 12.60 | 18.57 | 17.78 |
| OMPS-NPP L3 daily, area-weighted | CAMS | -2.75 | 13.79 | 21.92 | 21.74 |
| WOUDC Brewer/Dobson | TCOscope | +2.43 | 13.89 | 19.76 | 19.61 |
| WOUDC Brewer/Dobson | CAMS | -6.20 | 17.18 | 24.10 | 23.29 |

TCOscope reduces RMSE versus CAMS by 35.8% against area-weighted TROPOMI, 15.3%
against area-weighted OMPS, and 18.0% against WOUDC. Paired seven-day block-bootstrap
95% confidence intervals are respectively 29.6–36.1%, 15.1–17.7%, and 15.8–19.9%
for the corresponding unweighted/ground protocols used by the bootstrap.

## RMSE by lead

| Truth | System | T+1 | T+2 | T+3 | T+4 | T+5 |
|---|---|---:|---:|---:|---:|---:|
| TROPOMI | TCOscope | 9.65 | 10.74 | 11.66 | 12.84 | 14.71 |
| TROPOMI | CAMS | 17.00 | 17.27 | 17.63 | 18.27 | 19.32 |
| OMPS | TCOscope | 19.97 | 19.88 | 19.99 | 20.22 | 20.87 |
| OMPS | CAMS | 23.99 | 24.03 | 24.09 | 24.28 | 24.60 |
| WOUDC | TCOscope | 18.34 | 18.65 | 19.25 | 20.37 | 21.99 |
| WOUDC | CAMS | 23.29 | 23.37 | 23.63 | 24.28 | 25.85 |

## Brewer and Dobson

| Instrument subset | System | N | Bias | MAE | RMSE | Pearson r |
|---|---|---:|---:|---:|---:|---:|
| Brewer | TCOscope | 9,250 | -0.65 | 13.86 | 19.28 | 0.933 |
| Brewer | CAMS | 9,250 | -10.37 | 18.25 | 25.39 | 0.904 |
| Dobson | TCOscope | 10,310 | +5.02 | 14.37 | 20.96 | 0.924 |
| Dobson | CAMS | 10,310 | -2.45 | 16.79 | 23.54 | 0.898 |
| Strict direct-sun | TCOscope | 16,195 | +2.92 | 12.21 | 17.19 | — |
| Strict direct-sun | CAMS | 16,195 | -6.09 | 15.24 | 21.35 | — |

The complete WOUDC sample has 22,217 matched forecast-observation pairs from 45
stations and 155 valid dates. Equal station weighting gives RMSE 18.91 DU for
TCOscope versus 22.80 DU for CAMS, so the result is not caused only by a few dense
stations. Pearson correlation on all matched pairs is 0.950 for TCOscope and 0.928
for CAMS.

## Moscow-bias interpretation

WOUDC has no matched Moscow station in this window. The nearest station is Kaunas,
about 876 km away. At Kaunas, TCOscope bias is +3.97 DU and CAMS bias is -10.38 DU;
RMSE is 11.35 versus 15.25 DU. Across all Northern midlatitude stations, TCOscope
bias is +1.75 DU and CAMS bias is -8.97 DU, with RMSE 19.26 versus 25.45 DU.

This supports a systematic low CAMS bias in the broader Northern midlatitudes and
explains part of the current Moscow difference. It does not prove that the full
approximately 20 DU Moscow gap belongs to CAMS: no colocated Moscow ground truth is
available, and CAMS 00 UTC instantaneous fields are being compared with a model
trained on daily satellite aggregates.

## Production-source sensitivity

The long YTD benchmark uses mature TROPOMI inputs because a complete historical
direct-CDSE NRTI archive is unavailable. The earlier NRTI-input benchmark through
2026-05-07 remains consistent: against WOUDC, TCOscope RMSE was 19.16 DU versus
24.19 DU for CAMS, with bias -0.86 versus -6.11 DU. The direct-CDSE replay had
RMSE 10.56 DU, MAE 7.44 DU, and bias +2.57 DU against mature TROPOMI on its 15
forecast/lead cases.

## Sources and limitations

- CAMS global forecasts: https://ads.atmosphere.copernicus.eu/datasets/cams-global-atmospheric-composition-forecasts
- WOUDC total ozone: https://woudc.org/en/data/dataset-information/totalozone
- NASA OMPS-NPP NMTO3 L3 daily: https://cmr.earthdata.nasa.gov/search/site/collections/directory/GES_DISC/gov.nasa.eosdis
- TROPOMI is not independent of TCOscope training. OMPS and WOUDC are the stronger
  external checks, although CAMS also assimilates satellite ozone products.
- Six late-June TROPOMI forecast/lead rows have zero observed cells and are excluded
  by the observation mask; OMPS and WOUDC still score those forecasts.
- Ground stations are sparse and point-versus-grid representativeness remains a
  source of error, especially for the 1.2 degree TCOscope grid.
