| 20 Feb 2026
A statistical mechanics model for cloud cover: Case of low clouds over the Gulf of Guinea
Abstract. This study uses satellite (MODIS) and reanalysis (ERA5) data to calibrate a stochastic model for low-cloud cover over the Gulf of Guinea. The model, based on the Ising model of ferro-magnetic materials, tracks cloudy and non-cloudy sites on a rectangular lattice. A Hamiltonian that includes external and internal interaction potentials is adopted. MODIS cloud cover data constrain the first three moments of the cloud area fraction (CAF) frequency distributions while the ERA5 boundary layer temperature (Tb) drives the Ising model. A hybrid optimization method is used to learn the internal and external potentials (J0 and h0). MODIS CAF frequency distributions, binned by Tb values, are bimodal, indicating three distinct cloud regimes: low-, intermediate-, and high-CAF. The calibrated Ising model reproduces these regimes through its metastable behaviour, characterized by triple equilibria. MODIS data and model results showed near-identical power-law fits of mean CAF to Tb. As Tb decreased, CAF frequency distribution changed from low to moderate to high CAF regimes in both data sets. Both distributions had similar meandering behaviour, with inflection points marking regime changes and deviations from the power law. An optimal J0 and h0 parameterization was derived in terms of Tb. The learning algorithm prefers high J0 and h0, leading to stable equilibria and exaggerated modal peaks, reminiscent to the Ising model’s behaviour at low temperature. This shortcoming is perhaps due to the crude learning metric. Information theoretic methods, though costly, could be beneficial.
In the article "A statistical mechanics model for cloud cover: Case of low clouds over the Gulf of Guinea" the authors propose that clouds can be described using the 2D Ising model. As evidence, they fit two parameters of 2D Ising model simulations, to replicate the distribution of cloud cover from satellite-observations.
Although I find the idea compelling, I cannot recommend the article for publication. I find that the connection of clouds to the Ising model requires more justification. The motivation offered is that clouds are self-similar, and the Ising Model also gives rise to self-similar structures. There is no discussion of what physical mechanism in the clouds could play the role of the Ising model's neighbor interaction J.
The only quantity compared between satellite observations and the Ising model is the cloud area fraction. After all the discussion of self-similar structures and distributions of chord lengths, I would like to see some deeper analysis of the clouds and the model, something that measures the spatial structures.
In general, I find the picture too simplistic. The cloud scenes considered are small pieces (31x31 pixels) of MODIS images. This misses cloud processes and patterns at both finer and larger scales. And then only the cloud area fraction of these small pieces are analyzed, ignoring all other spatial structure. The cloud scenes are analyzed by binning only according to a mean atmosphere temperature, which misses a lot of rich cloud behavior controlled by other large-scale quantities. This limited framing would be acceptable if the link to the Ising model could be made very strong and clear, which is not the case here in my opinion.
Smaller remarks:
I'm hesitant on the remapping of the traditional sigma = -1 or 1, to sigma = 0 or 1 here. The model behavior remains the same, but the meanings of the parameters J and h change. With the traditional values, J gives an interaction strength, an alignment energy. With the 0 or 1 values, J also favors one value above the other, which traditionally is the role of h. This becomes important when trying to interpret the fitted values of J and h.
line 68. The term "order parameter" doesn't make sense here in the model definition. For a specific phase transition one could then identify an order parameter, but that seems not to be discussed in this paper.
line 200. I don't see the point of using (4) for h, if h_0 is anyway fitted independently for each temperature bin.
line 208. The 20x20 lattice seems arbitrary and fairly small. The justification "to roughly match the MODIS images of 31x31" is not convincing. If one were to study something else than the cloud area fraction, for example chord lengths, this small size would be limiting.
line 209. How is it determined that the simulation has run long enough?
Code and data availability
In general, making the code and model data publicly available, permanently archived and citable by a DOI, is preferable over "available on request".