Climate and planetary models must compute how heat radiates through a stack of atmospheric layers. Doing this exactly means tracking millions of sharp molecular absorption lines at every wavelength. The cheapest shortcut is the grey approximation, which pretends each layer absorbs equally at all wavelengths. In effect, each layer is described by a single number, its optical depth (how opaque it is), and the column's opacity is just the sum of those numbers. Grey models are fast enough to run hundreds of thousands of planetary evolution tracks. They also misplace the edges of habitability. This proposal asks a composition question about this shortcut. Suppose each layer may pass only $r$ numbers to the rest of the column, and those numbers must combine the way optical depths do: by adding across layers and then being read out through exponentials. What is the smallest $r$ that reproduces the exact fluxes to a target accuracy, and how does that $r$ grow as the atmosphere moves into hot, chemically mixed, or perturbed regimes?
We call this smallest $r$ the composition width of an atmospheric regime. The grey model has width 1. We claim width is a computable, physically meaningful complexity measure. It is the radiative analogue of the composition complexity studied for Boolean functions, where one asks how many local pieces an arbitrary combiner needs. It explains when grey models fail. It also gives a design rule for neural emulators: build them with exactly this additive, width-$r$ structure, and choose $r$ from the measured width of the regimes they must cover. The main bet is empirical and can be checked. Regimes where width rises sharply are where grey models break and where off-the-shelf emulators fail out of distribution.
Core Claim
Let $\tau_l(\nu)$ be the optical depth of layer $l$ at wavenumber $\nu$, and let $w(\nu)$ be a spectral weight such as the Planck function. The exact band transmission between levels $a$ and $b$, the quantity from which fluxes and heating rates are built, is
$$T_{ab} = \int w(\nu)\, \exp\Big(-\sum_{l=a}^{b} \tau_l(\nu)\Big)\, d\nu .$$
A width-$r$ composition replaces this with
$$\hat T_{ab} = \sum_{i=1}^{r} c_i \exp\Big(-\sum_{l=a}^{b} s_{l,i}\Big), \qquad s_{l,i} = \phi_i(\text{state of layer } l) \ge 0 .$$
Here $c_i \ge 0$ are learned weights and $\phi_i$ are learned maps from a layer's temperature, pressure and composition to $r$ nonnegative "pseudo-channel" optical depths. The composition width $W_\varepsilon$ of a regime is the smallest $r$ for which $\max_{a,b}|T_{ab}-\hat T_{ab}| \le \varepsilon$ over all columns in that regime. The grey model is $r=1$. If the spectrum keeps the same ordering of strong and weak wavenumbers in every layer, a correlated-$k$ table with $K$ quadrature points gives $r = K$. When that ordering breaks, the required $r$ grows. This happens when overlapping gases change their mixing ratios with height, or when line shapes shift with temperature, as in dissociating steam.
The new element is treating $W_\varepsilon$ as a quantity to measure and bound, rather than a tuning knob. Two uses follow. First, the growth of $W_\varepsilon$ across regimes is a diagnostic of where grey and band models break. Second, the same additive structure, used as a neural architecture, should resist distribution shift better than generic architectures, because its way of combining layers matches the physics.
Where This Sits
The planetary motivation comes from two lines of work. Coupled interior–atmosphere studies using the line-by-line GARLIC solver, which draws on the HITEMP2010 and HITRAN 2016 line databases plus a water-vapor continuum model, show that steam above about 2000 K dissociates into H$_2$ and O$_2$. This lowers outgoing longwave radiation by 1–6% relative to pure steam (Evolution and Spectral Response of a Steam Atmosphere for Early Earth with a Coupled Climate–Interior Model; Coupled atmosphere-interior modeling during Earth's magma ocean phase). Meanwhile, the 280,000-track stagnant-lid habitability survey relied on grey radiation because anything finer was too expensive (Redox state and interior structure control on the long-term habitability of stagnant-lid planets). These papers show the gap: a cost–accuracy tradeoff that nobody has quantified per regime. Our width measurement supplies that quantification.
The complexity side supplies a lens but not a method. The composition complexity of majority counts how many $k$-local functions an arbitrary combiner needs. It shows Parity costs exactly $n/k$ while Majority pays an unavoidable $\log k$ overhead. The grey model is the radiative "Parity" case: optical depths add with no overhead. Our open question is whether realistic non-grey mixtures carry an analogous overhead that grows with the dynamic range of mixing ratios. We do not import the Boolean proofs. The setting is continuous, and the combiner here is restricted to add-then-exponentiate. That restriction is exactly what makes the measure physically interpretable.
How Compositional is a Model? bounds how well a model generalizes to unseen combinations by its "locus of influence," meaning how much one input token can sway the output. Our architecture caps each layer's influence to an additive contribution in $r$ channels. That paper is motivation for why this should help out of distribution. We do not claim to compute its bound for atmospheres.
ClimART is the empirical anchor for emulators (ClimART). Its benchmark covers more than 10 million CanESM5 columns. There, structured 1D convolutional networks and graph networks generalize to future and volcanic climates, while multilayer perceptrons blow up. ClimART reports that MLP error "jumped from under 1.0 for the present day to over 100 for future conditions." ClimART shows that structure matters but does not say which structure or why. Our proposal gives a specific answer to test.
Two radiative transfer surrogates attack a different axis of the problem. Physics Informed Neural Networks for Simulating Radiative Transfer trains networks to satisfy the transport equation's residual. Uniformly accurate structure-preserving neural surrogates for radiative transfer handles the optically thick limit with an asymptotic-preserving parity decomposition. Both concern angular and spatial transport, mostly in idealized or grey settings. Our concern is the spectral dimension, so the two lines of work are complementary rather than competing.
Research Program
The first component is a width estimator. For a given set of columns, we compute exact layer optical-depth spectra and all level-to-level transmissions with a line-by-line code using HITRAN/HITEMP data, as in the Katyal et al. pipeline. We then fit the width-$r$ model for increasing $r$ by gradient descent on $\phi_i$ and $c_i$, with restarts. The smallest $r$ meeting tolerance gives an empirical upper bound on $W_\varepsilon$. To estimate lower bounds, we also fit a relaxed model whose per-layer summaries feed an unrestricted neural combiner. If even this model needs $r$ channels, the additive restriction is not the bottleneck.
The second component is a regime atlas. We measure $W_\varepsilon$ across several families of columns:
- Earth-like clear-sky profiles from the 2,580 EUREC4A dropsondes (Atmospheric radiative profiles during EUREC 4 A).
- Pure steam atmospheres from 300 K to 4000 K.
- Dissociating steam with H$_2$/O$_2$ fractions as reported by Katyal et al.
- CO$_2$–H$_2$O mixtures spanning the oxidized-to-reduced outgassing range relevant to the stagnant-lid survey.
The theoretical target is a conjecture analogous to Majority's overhead. For two gases whose lines overlap randomly and whose mixing ratio varies by a factor $R$ through the column, we conjecture $W_\varepsilon$ grows at least like $\log R$, whereas for a single gas it stays bounded. Proving even the single-gas upper bound and a two-gas lower bound in a stylized line model would be a clean result. We treat the general case as open.
The third component is the emulator. The same width-$r$ structure becomes a neural network: a small per-layer network outputs $s_{l,i}$, the column sums them along partial paths (computable for all level pairs by cumulative sums), exponentiates, and reads out fluxes. On ClimART, where inputs are layer states and targets are flux profiles, this "pseudo-spectral" emulator is trained with $r$ chosen from the atlas.
Baselines
The naive reference point is the best grey model: our architecture with $r=1$, fitted on the same data. This is not a strawman. It is the optimally tuned version of what planetary evolution codes actually use. If larger $r$ does not clearly beat it in a regime, that regime does not need non-grey treatment, and that is a useful finding in itself.
The strongest physics competitor is a correlated-$k$ model with the same number of quadrature points per band as our $r$. It is built from the same line-by-line spectra and scored on the same partial-path transmissions and heating rates. Correlated-$k$ is the established, physically derived answer to exactly this question. The comparison is fair only at matched width, because width sets the downstream cost. Our model beats it if, at equal $r$, it achieves lower worst-case heating-rate error in the mixed-gas and dissociating-steam regimes. Equivalently, it should reach the same error with fewer channels.
The strongest learned competitors are the ClimART 1D convolutional network, the overall winner on present-day accuracy, and GraphNet, which had the lowest mean bias error. We rerun both with ClimART's released code on identical splits, with matched parameter counts and the same hyperparameter-search budget as our model. Exact RMSE values for these architectures on the out-of-distribution splits: no reported result found on Lacuna for these (they appear only in ClimART's figures). We therefore report our own reruns alongside the MLP anchor quoted above. Beating them means matching present-day error within noise while showing smaller degradation on the 2097–2099 future split and the 1991 Pinatubo split.
The key ablation keeps everything about our emulator except the additive coupling. Per-layer $r$-dimensional summaries are fed to an unrestricted combiner network. If this variant generalizes out of distribution as well as the additive one, then the gain comes from the bottleneck width, not from physical structure, and the main architectural claim fails.
Why not just…?
Why not just use correlated-$k$, which already is a width-$r$ composition? This is the real threat, and we say so directly. Correlated-$k$ is exact when spectral ordering is preserved across layers. Our model contains it as a special case, so in single-gas, well-correlated regimes we expect a tie. The open question is whether the tie survives overlapping gases and temperature-dependent line shapes. There, correlated-$k$ must either add channels or make overlap assumptions. The matched-width comparison in the mixed CO$_2$–H$_2$O and dissociating-steam regimes decides this. If correlated-$k$ wins there, the contribution shrinks to the width atlas itself, which is still new as a regime-by-regime measurement.
Why not just train a larger CNN or GraphNet on ClimART? ClimART already shows that convolutional and graph networks generalize far better than MLPs, so capacity is not the issue. What is unclear is whether they fail on shifts they were never structured for. Neither architecture forces a layer's effect to enter additively in optical depth, so a large shift in a gas concentration can be extrapolated arbitrarily. The out-of-distribution comparison at matched parameters and tuning budget tests whether the additive structure buys anything beyond what generic locality already gives. A null result here would be informative and publishable.
Why not just use a physics-informed or asymptotic-preserving solver? These methods address angular transport and optical thickness, not spectral structure. The asymptotic-preserving surrogate of Bai et al. works in grey or idealized slabs, and the PINN work of Mishra and Molinaro reports 0.24% error in 1D monochromatic slabs. Neither tells you how many spectral channels a hot, mixed atmosphere needs. They could in principle be stacked on top of a width-$r$ spectral decomposition, so we do not compete with them on their benchmarks.
Why the complexity framing at all, rather than "spectral quadrature fitting"? Fitting is the method. The framing adds a question fitting alone does not ask: whether width must grow with column properties, the way Majority's cost must grow with locality. If the conjectured $\log R$ growth fails even in the stylized model, we will drop the analogy and keep the atlas.
Proposed Experiments
The first decisive check is whether width predicts grey failure. Across the regime atlas, we plot the optimal grey model's outgoing-radiation and heating-rate errors against $W_\varepsilon$ at fixed $\varepsilon$. We expect width near a small constant for Earth clear-sky EUREC4A columns. We expect a sharp rise in dissociating steam above 2000 K, which should be largest where Katyal et al. find the 1–6% outgoing-radiation shift. We also expect rising width along the oxidized-to-reduced CO$_2$–H$_2$O axis. A flat width curve across these regimes would falsify the diagnostic claim.
The second check uses ClimART. The width-$r$ emulator, the CNN, GraphNet, the free-combiner ablation, and the $r=1$ grey model are trained on 1979–2006. All are evaluated on 2007–2014, 1850–1852, 2097–2099, and 1991 using RMSE and mean bias error on fluxes and heating rates. We sweep $r$ to show where accuracy saturates.
The third check, which we consider the most decisive, is predictive. Before training any emulator, we compute line-by-line widths on representative columns from each ClimART split. We then ask whether the increase in width from training to test regime predicts the ranking of out-of-distribution error across architectures and splits. If it does, width becomes an a priori criterion for when an emulator trained in one climate can be trusted in another. If it does not, the measure is descriptive only, and we will report that.
References
- Evolution and Spectral Response of a Steam Atmosphere for Early Earth with a Coupled Climate–Interior Model
- Coupled atmosphere-interior modeling during Earth's magma ocean phase
- Redox state and interior structure control on the long-term habitability of stagnant-lid planets
- The Composition Complexity of Majority
- How Compositional is a Model?
- ClimART: A Benchmark Dataset for Emulating Atmospheric Radiative Transfer in Weather and Climate Models
- Physics Informed Neural Networks for Simulating Radiative Transfer
- Uniformly accurate structure-preserving neural surrogates for radiative transfer
- Atmospheric radiative profiles during EUREC 4 A