Sparse-observation reconstruction in coupled atmospheric world models

1. Research question and context

Coarse reanalysis and regional climate simulations provide broad atmospheric context but cannot directly resolve every local process across Jamaica's complex terrain. Surface observations constrain conditions at individual locations; satellite products represent different spatial and temporal supports. The research question is whether locally coupled state models can combine those information sources while preserving a defensible account of exchanges between neighbouring volumes.

Recent numerical–AI research has sharpened the evaluation problem. Sha et al. [1] measured the effects of mass and energy constraints on an AI forecasting model. AIFS 1.1.0 incorporates physical bounds for selected forecast variables [2]. WIND formulates atmospheric reconstruction and downscaling as inverse problems [3], while generative assimilation and prediction combines observations, forecast dynamics and longer-term simulation within one probabilistic framework [4]. These studies establish relevant methods, not their performance in the Caribbean.

Observation-based evaluation also matters. Gupta et al. [5] reported differences between AI-model error measured against reanalysis and error measured against independent observations. Tian et al. [6] documented sensitivity limitations in learned weather models used for data assimilation. These findings motivate separate evaluation of observational fit, physical consistency and free-running prediction.

2. Numerical formulation

Conservative shared-face transport

The controlled experiment represents a scalar state x over an 8 × 8 grid with no external sources or boundary exchange. Neighbouring cells exchange a quantity using a symmetric diffusion coefficient κ and fixed step Δt. The exchange is added to one cell and removed from the other.

x_i(t + Δt) = x_i(t) + Δt κ ∑_j∈N(i) [x_j(t) − x_i(t)]

Because each interface transfer is paired, the scalar total remains constant to floating-point tolerance. The reference uses κ = 0.16; the candidate forecast uses κ = 0.055. Both start from smooth but spatially different states adjusted to have the same total.

Sparse observation update

Eight fixed cells provide reference observations during the analysis interval. A Gaussian distance kernel interpolates observation-minus-prior residuals across the field. The constrained variant subtracts the spatial mean of the proposed correction before applying it, leaving the total unchanged. An unconstrained variant applies the same interpolated update without this correction.

x^a = x^f + δx − mean(δx) · 1

Here x^f is the model prior, x^a the updated state, and δx the interpolation of observed residuals. This operation enforces conservation of one synthetic scalar only. A three-dimensional atmosphere requires explicit dry-air mass, moisture, energy and momentum budgets with boundary fluxes and source terms.

3. Experiment and results

The model runs for 80 time steps. Observations are supplied every two steps until step 40 and withheld thereafter. Eight additional, non-overlapping cells are reserved for evaluation. The open-loop interval, steps 41–80, tests whether the corrected state improves subsequent prediction when no further observations are available.

RMSE is dimensionless, averaged over eight withheld sites. Results are deterministic for this fixed synthetic configuration.

The constrained model's maximum budget drift over the run was approximately 2.13 × 10⁻¹⁴ scalar units. The unconstrained observation update reached a drift of 0.09432. The mismatched-observation control increased analysis error rather than improving reconstruction.

Explore the numerical experiment and compare model variants. The implementation provides the exact numerical update, observation indices, coefficients and scoring code.

4. Application to Caribbean atmospheric reconstruction

Observations and baselines

The proposed Caribbean evaluation begins with Jamaica rainfall and near-surface temperature, using timestamped large-scale forcing and separately verified observational products. Candidate sources include ERA5 [7], NASA IMERG [8], CHIRPS [9], and surface and radiosonde observations from NOAA archives [10, 11]. Differences in spatial support, instrument uncertainty, product version and gauge overlap must be accounted for when selecting independent evaluation targets.

The existing Jamaican rainfall experiment offers an important adverse baseline: an exact-aggregation-constraint neural model recorded 14.23 mm/month RMSE against 9.73 mm/month for seasonal allocation. That result prevents a conservation constraint alone from being treated as evidence of improved local prediction. See the experimental record and evidence classification.

Evaluation of a coupled atmospheric model

A regional study needs time-ordered training and evaluation splits, geographically withheld observation locations, frozen issue times and matched baseline models. The necessary comparisons include conventional interpolation, seasonal climatology, uncorrected numerical or learned priors, observation-corrected states, and coupled versus uncoupled local fields. Reconstruction error, conservation residuals, extreme-event error and forecast stability should be reported separately.

A learned latent-state prior, such as a joint-embedding predictive representation, is one possible extension. Its value would need to be established by independent observational improvement or forecast stability relative to a simpler model with comparable inputs and computational cost. The initial experiment provides a testable numerical foundation for that comparison, not a substitute for atmospheric training.

5. References and reproducibility

1. Sha, Y., Schreck, J. S., Chapman, W., & Gagne II, D. J. (2025). Improving AI Weather Prediction Models Using Global Mass and Energy Conservation Schemes. Journal of Advances in Modeling Earth Systems. doi:10.1029/2025MS005138.
2. Moldovan, G., et al. (2026). AIFS Single 1.1.0: an update to ECMWF's machine-learned weather forecast model AIFS. Geoscientific Model Development, 19. Article.
3. Aich, M., Fürst, A., Sestak, F., Ruiz-Gonzalez, C., Boers, N., & Brandstetter, J. (2026). WIND: Weather Inverse Diffusion for Zero-Shot Atmospheric Modeling. Proceedings of the 43rd International Conference on Machine Learning, 306, 1404–1433. Proceedings.
4. Yang, S., et al. (2026). Generative assimilation and prediction for weather and climate. npj Climate and Atmospheric Science, 9, 218. Article.
5. Gupta, et al. (2026). MAUSAM: An Observations-Focused Assessment of Global AI Weather Prediction Models During the South Asian Monsoon. Journal of Advances in Modeling Earth Systems. doi:10.1029/2025MS005568.
6. Tian, X., Holdaway, D., & Kleist, D. (2026). Evaluating Machine Learning Weather Models for Data Assimilation: Fundamental Limitations in Tangent Linear and Adjoint Properties. Geophysical Research Letters. doi:10.1029/2025GL119402.
7. ECMWF Copernicus: ERA5 single levels.
8. NASA GPM: IMERG.
9. Funk, C., et al. (2015). The climate hazards infrared precipitation with stations. Scientific Data, 2, 150066. doi:10.1038/sdata.2015.66.
10. NOAA: Integrated Surface Database.
11. NOAA: Integrated Global Radiosonde Archive.

Research material. Adrian Dunkley, Climate Studies Group Mona, The University of the West Indies, Mona. Source code, simulation configuration and generated comparison figures are provided with the numerical experiment. Numerical results above come from constructed data; observational validation remains a separate research stage.