acceptodds
Under review as a conference paper at ICLR 2027

UniTeD: Resolving Sub-Grid Heterogeneity for Robust Grid Cell Prediction

14 views

Abstract

Machine learning (ML) has seen widespread adoption in the physical sciences for developing data-driven emulators of grid-based simulations. Despite this rapid progress, physical systems characterized by sub-grid heterogeneity remain largely underexplored. In domains ranging from land surface modeling in Earth sciences to multi-scale stress analysis in materials science, a grid cell's state is far from uniform. Instead, the cell's state is highly heterogeneous: distinct sub-grid “tiles” exhibit intrinsic processes that must be resolved before computing the grid cell's aggregate behavior. Directly applying standard ML models, which lack the inductive bias to account for this heterogeneity or resolve these underlying processes, can lead to structural entanglement and poor generalization when sub-grid fractions shift. To address this gap, we introduce UniTeD (Unified Tiled-Expert Demixer), a sub-grid-aware ML architecture that strictly enforces physical aggregation logic. UniTeD isolates latent constitutive functions purely from aggregate supervision, enabling robust generalization across varying sub-grid fraction compositions. Evaluated on synthetic benchmarks and global ERA5 surface energy flux inference, UniTeD achieves zero-shot compositional generalization, reducing prediction errors by up to 28% relative to the strongest baseline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cell heterogeneity, Sub-grid Physics, Disentanglement, Composition, Mixture-of-Experts, Sub-grid parameterization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TVtDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies grid-cell prediction in physical systems with heterogeneous sub-grid structure, where the observed cell-level target is an aggregate of several latent constituent responses. The authors argue that conventional monolithic neural models may entangle the effects of different sub-grid components, leading to poor generalization when constituent fractions change.

To address this issue, the paper proposes UniTeD (Unified Tiled-Expert Demixer), which combines a shared trunk encoder, a FiLM-conditioned shared expert for modeling tile-specific latent responses, and a physics-prescribed aggregation layer that combines these responses using known tile fractions. By restricting the fractions to the final aggregation step, UniTeD explicitly enforces linear superposition and prevents the latent tile responses from directly depending on the mixture configuration. The paper further provides a theoretical identifiability analysis under regularity, structural-exclusion, and local non-degeneracy assumptions.

The method is evaluated using both controlled synthetic benchmarks and ERA5 surface-energy-flux prediction. Synthetic experiments examine latent-process recovery, zero-shot compositional generalization, few-shot transfer, nonlinear interactions, and noisy tile fractions. On ERA5, UniTeD achieves substantially better robustness under mixture-to-pure and pure-to-mixture distribution shifts and reports up to approximately 28% error reduction relative to the strongest baseline, together with improved geographical generalization.

Critical strengths

The paper's strongest aspect is the close alignment between the model architecture and the assumed physical structure. UniTeD does not ask a generic network to infer how constituent fractions should interact; instead, the model learns tile responses independently of the fractions and combines them through the prescribed physical aggregation rule. The use of a shared FiLM-conditioned expert also provides a natural mechanism for transferring information between related, data-rich and data-scarce tiles.

The experimental design is another important strength. The synthetic experiments directly test latent disentanglement, unseen mixture generalization, and few-shot transfer, while the ERA5 experiments demonstrate meaningful gains at much larger scale. The additional experiments involving fraction noise, nonlinear coupling, and geographical distribution shift substantially improve the completeness of the empirical study.

Critical weaknesses

The first concern is that the paper somewhat overstates the connection between the theoretical identifiability result and the ERA5 experiments. The theoretical argument relies on locally full-rank, sufficiently well-conditioned fraction matrices under nearly fixed forcing, whereas the empirical condition-number distribution contains a substantial ill-conditioned tail. Moreover, aggregate predictive stability in these regions does not by itself imply that all latent constituent functions are uniquely and accurately recovered.

A second concern is that the real-data experiments cannot directly verify the central latent-disentanglement interpretation because ERA5 does not provide ground-truth tile-level fluxes. Improved performance under composition shifts is certainly consistent with successful physical disentanglement, but it is not uniquely diagnostic of it. Relatedly, the nonlinear-coupling robustness experiment evaluates recovery of pure, uncoupled tile responses after training on coupled mixtures; this is useful as a stress test, but it does not demonstrate that the current linear aggregation architecture can accurately predict genuinely coupled mixture outputs at deployment time.

Questions for authors

Could the authors clarify more quantitatively how the ERA5 condition-number analysis supports the local non-degeneracy assumption in Theorem 3.1? Given the very large condition numbers in the upper tail, it would be helpful to report the distribution of σmin⁡(AB)\sigma_{\min}(A_B), or the fraction of neighborhoods satisfying a concrete lower-bound criterion required for numerically stable latent recovery.

Could the authors also clarify the intended strength of the latent-function recovery claim on ERA5? Since only aggregate flux targets are available, the current experiments directly validate aggregate prediction and compositional generalization rather than the correctness of each latent tile response. An external tile-level diagnostic, comparison against outputs from a process-based land-surface model, or a counterfactual consistency test would help distinguish true physical disentanglement from a useful but potentially non-identifiable internal representation.

4 comments

We thank the reviewer for their constructive feedback. We have addressed the reviewer’s questions regarding the quantitative numerical stability of our theoretical assumptions and the intended scope of our latent recovery claims on the ERA5 dataset in our point-by-point responses below.

1. Could the authors clarify more quantitatively how the ERA5 condition-number analysis supports the local non-degeneracy assumption in Theorem 3.1? Given the very large condition numbers in the upper tail, it would be helpful to report the distribution of σmin(AB)\sigma_{min}(A_B), or the fraction of neighborhoods satisfying a concrete lower-bound criterion required for numerically stable latent recovery.

The minimum singular values of the local fraction matrices have the following distribution:

Percentileσmin⁡(AB)\sigma_{\min}(A_B)
5th2.57×10−42.57 \times 10^{-4}
25th3.43×10−23.43 \times 10^{-2}
50th0.3970.397
75th1.201.20
95th3.943.94

We found these results consistent with the long-tailed condition-number distribution reported in Appendix D.6. As discussed there, we observed that extreme condition numbers tend to occur in vast, homogeneous regions, where some tile types are absent or have negligible coverage and the corresponding fraction-matrix columns are zero or nearly zero.

In the limiting case of a neighborhood composed entirely of one tile aa, the aggregation equation simplifies to:

yi=1⋅Φa(xi)+∑k≠a0⋅Φk(xi)=Φa(xi)y_i = 1\cdot\Phi_a(x_i) + \sum_{k\neq a}0\cdot\Phi_k(x_i) = \Phi_a(x_i)

Thus, although the full fraction matrix is rank-deficient, the active tile response is directly supervised by the aggregate target, without requiring unmixing. The absent tile responses remain locally unconstrained, but they do not contribute to these cells’ aggregate outputs. Therefore, a large condition number in such a region does not by itself imply difficulty in aggregate prediction.

For neighborhoods whose fraction matrices have full column rank, Equation 11 in Appendix A.3 provides the latent-recovery error bound through the pseudoinverse, with sensitivity governed by 1/σmin⁡(AB)1/\sigma_{\min}(A_B). Larger minimum singular values yield tighter bounds when the other terms are held fixed. For rank-deficient neighborhoods, the pseudoinverse constrains only observable combinations of tile responses, rather than identifying every tile response.

Homogeneous neighborhoods can directly supervise their active tile responses, whereas recovering all latent responses requires the local non-degeneracy assumption. These diagnostics do not establish uniformly stable recovery of all latent tile responses on ERA5; rather, they characterize the strength and limitations of the local constraints. We will clarify this distinction in the revised manuscript.

2. Could the authors also clarify the intended strength of the latent-function recovery claim on ERA5? Since only aggregate flux targets are available, the current experiments directly validate aggregate prediction and compositional generalization rather than the correctness of each latent tile response. An external tile-level diagnostic, comparison against outputs from a process-based land-surface model, or a counterfactual consistency test would help distinguish true physical disentanglement from a useful but potentially non-identifiable internal representation.

We agree with the reviewer’s assessment. Since the public ERA5 archive only provides grid-level aggregate fluxes, we cannot directly verify the strict correctness of each individual latent tile response. In the revised manuscript, we will explicitly clarify the intended strength of our claims on ERA5: the latent tile functions learned on ERA5 should be interpreted as physically guided approximations rather than strictly verified ground-truth responses.

Regarding the suggestion of an external tile-level diagnostic, we note that regenerating these global datasets using a coupled, process-based land-surface model (like HTESSEL, the underlying land-surface model for ERA5) to extract internal tile-level states requires significant High-Performance Computing (HPC) infrastructure and massive data storage typically only accessible to major operational climate/research centers. For the broader research community, relying on ERA5 provides a standardized, globally recognized, and operationally validated dataset, which ensures our evaluation is reproducible and comparable to state-of-the-art benchmarks. This structural barrier to accessing real-world tile-level ground truth was the exact motivation for our two-tier experimental design.

Our controlled synthetic benchmark serves as a "counterfactual consistency test" for latent-process recovery. By mimicking the aggregation processes of a land-surface model in a controlled environment, we empirically show that UniTeD can recover good approximations of the tile functions when trained only on mixtures (Figures 2 & 3) and generalizes to novel compositions (Figures 4 & 5). Following this proof of concept, our ERA5 experiments were designed to test whether a machine learning model, structurally constrained to be faithful to the underlying physics of these simulators, could scale to challenging data at global scale.

Even interpreted as internal representations or approximations of tile functions, this capability is vital for practical applications. Under anthropogenic and natural shifts, such as global urbanization (which is transforming global landscapes), models must frequently generalize to entirely unseen landscapes (Chaney et al., 2018; Torres-Rojas et al., 2022; Yin et al., 2023). A highly relevant example of this is counterfactual urban planning: when decision-makers assess how building a new park, an artificial lake, or a massive building complex might impact a city's microclimate, the modeling tool must predict the physical response of a completely novel tile composition. An architecture like UniTeD, which can operate with unknown tile compositions, is more useful for these downstream scientific and planning applications.

Thanks for addressing my concerns. I will maintain my score.

We appreciate the reviewer’s response. Are there any remaining concerns critical to your assessment that we could address?

Reviewer_P5HzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UniTeD, a framework for predicting grid-level physical quantities from heterogeneous sub-grid components. It combines a FiLM-conditioned shared expert to model tile-specific responses with a physics-prescribed linear aggregation mechanism based on known tile fractions. The authors also provide a theoretical analysis of latent response identifiability. Experiments on synthetic data and ERA5 demonstrate improved generalization under changes in sub-grid composition, particularly when transferring between pure and mixed grid cells.

Critical strengths
  1. Well-motivated modeling framework. The paper addresses a meaningful problem in heterogeneous grid-cell prediction. Incorporating known physical aggregation rules provides a reasonable inductive bias for improving compositional generalization.
  2. Encouraging out-of-distribution results. The experiments on synthetic data and ERA5 show notable improvements over the selected baselines under shifts in tile composition. These results suggest that explicitly modeling sub-grid structure can be beneficial.
Critical weaknesses
  1. Insufficient architectural ablation. The paper lacks comparisons with simpler shared-expert architectures, making it difficult to isolate the specific contribution of FiLM beyond parameter sharing and physics-prescribed aggregation.
  2. Limited evaluation of generality and trade-offs. The experiments mainly focus on compositional distribution shifts, while the method's performance under standard training settings and the potential limitations of its linear aggregation assumption remain insufficiently explored.
Questions for authors
  1. Could the authors compare UniTeD with a shared MLP using tile identity embeddings and the same physics-prescribed aggregation mechanism to better isolate the contribution of FiLM?
  2. Could the authors provide a standard ERA5 evaluation where both training and testing include the full range of pure and mixed tile compositions?
  3. Since linear area-weighted aggregation does not necessarily imply independent tile responses, could the authors further evaluate prediction accuracy on mixed grid cells with nonlinear inter-tile interactions, particularly as the interaction strength increases?
  4. Could the authors provide more diverse visualizations beyond the current response curves, such as spatial error maps on ERA5 or error distributions across different tile compositions? These would help better illustrate where UniTeD performs well and where its limitations arise.
4 comments

We thank the reviewer for their constructive feedback. We have addressed the reviewer’s questions regarding architectural ablations, standard evaluations, physical coupling, and spatial error distributions in our point-by-point responses below.

1. Could the authors compare UniTeD with a shared MLP using tile identity embeddings and the same physics-prescribed aggregation mechanism to better isolate the contribution of FiLM?

To explicitly isolate the contribution of the FiLM conditioning mechanism, we conducted the requested ablation. We replaced FiLM with tile identity embeddings and the same physics-prescribed aggregation mechanism.

As shown in the tables below, the 'UniTeDNoFiLM' ablation is consistently less performant across all evaluated purity and mixture thresholds compared to the full UniTeD architecture. This confirms that FiLM’s multiplicative and additive modulation provides an architectural advantage compared to identity embeddings.

Performance across Purity Thresholds

ModelTest (Mixture).7.8.91.0 (Pure)
UniTeDNoFiLM63.0 ± 0.4120.0 ± 3.3125.7 ± 3.8133.2 ± 6.9122.8 ± 20.4
UniTeD (Ours)62.4 ± 0.8113.4 ± 2.3118.2 ± 2.6122.7 ± 4.092.1 ± 11.4

Zero-Shot Compositional Generalization (Pure → Mixture)

ModelTest (Pure).01.05.08
UniTeDNoFiLM36.0 ± 0.379.6 ± 2.285.4 ± 4.386.8 ± 2.1
UniTeD (Ours)34.5 ± 0.377.1 ± 0.778.2 ± 0.370.3 ± 0.9

2. Could the authors provide a standard ERA5 evaluation where both training and testing include the full range of pure and mixed tile compositions?

We performed a standard baseline evaluation by training and testing all models on the full distribution of ERA5 data (including the complete range of pure and mixed tile compositions in both the training and testing sets). The RMSE results are as follows:

ModelRMSE
DNN56.54 ± 0.001
FTTransformer55.08 ± 0.025
SoftMoE55.72 ± 0.070
FracMoE55.30 ± 0.118
UniTeDNoFiLM55.24 ± 0.198
UniTeD (Ours)55.00 ± 0.044

As expected, under this standard evaluation, all methods perform relatively well and achieve comparable performance. UniTeD continues to deliver the best overall performance (55.00), though the performance gap between the architectures is narrower. These results mirror the in-distribution trends shown in the first columns of Tables 1a and 1b ("Test (Mix)" and "Test (Pure)").

As noted in the manuscript, highly parameterized machine learning methods generally excel at interpolating in-distribution data. However, the ability to generalize to unknown or unseen compositions is critical in practical applications, as standard evaluation (with identical training and testing distributions) fails to capture the realities of modern climate monitoring and Earth system modeling. Under anthropogenic and natural shifts, such as urbanization which is transforming global landscapes, models must frequently generalize to entirely unseen landscapes (Chaney et al., 2018; Torres-Rojas et al., 2022; Yin et al., 2023).

A relevant practical example of generalizing to unknown tile compositions can be seen in urban planning. For instance, when decision-makers assess how building a new park, an artificial lake, or a massive building complex might impact a city's micro-climate, a useful modeling tool must predict the physical response of a completely novel tile composition. An architecture like UniTeD, which can operate with unknown tile compositions, is more useful for these downstream scientific and planning applications.

3. Since linear area-weighted aggregation does not necessarily imply independent tile responses, could the authors further evaluate prediction accuracy on mixed grid cells with nonlinear inter-tile interactions, particularly as the interaction strength increases?

We would like to direct the reviewer to Appendix C.5 and Figure 7, where we have already conducted this evaluation. By introducing non-linear inter-tile interactions into the data generation process, we stress-tested the models as the interaction strength increased. The results demonstrate that UniTeD remains significantly more robust than the baselines, maintaining stable predictive accuracy even when moderate nonlinear cross-tile interactions are present, and degrading gracefully as the non-linear interaction increases.

Regarding the linear aggregation in our framework, we would like to note that UniTeD is designed to emulate Land Surface Models (LSMs), such as HTESSEL (the underlying LSM used in ERA5). An LSM computes vertical energy fluxes for individual sub-grid tiles and aggregates them linearly at the grid cell level [1]. Lateral interactions (e.g., advection and turbulent heat transport) are usually not resolved at the sub-grid tile level; rather, they are resolved globally at the macroscopic grid-cell level by the coupled atmospheric model. The atmospheric module controls the wind that transports heat and particles across grid cells. After computing these macroscopic interactions, the atmospheric model feeds the updated meteorological variables back to the LSM, which uses them to aggregate new energy fluxes for the next time step. The two modules operate in a continuous coupled feedback loop: the LSM (which our model UniTeD emulates) resolves sub-grid heterogeneity, while the atmospheric model resolves inter-cell interactions.

An intuitive, real-world example of this linear aggregation can be observed in our daily lives. Consider a paved surface and a lawn located next to each other on a sunny day: the pavement will become significantly hotter than the adjacent lawn. This is because the vertical emission of heat is vastly stronger than the lateral heat exchange directly between the two surfaces, which justifies the LSM treating their energy fluxes independently prior to aggregation.

Furthermore, in LSMs, there is usually no spatial topology information about the tiles within a grid cell (e.g., the LSM does not know whether vegetation is physically adjacent to pavement or water; it only tracks their respective percentage coverages). This lack of intra-cell topology makes it structurally impossible to compute lateral interactions directly between specific tiles. Therefore, the assumption of independent tile-level responses directly reflects the established, physically justified assumptions that govern the LSM systems UniTeD emulates.

[1] Balsamo et al., 2009. A revised hydrology for the ECMWF model: Verification from field site to terrestrial water storage and impact in the integrated forecast system, Section 2. TESSEL Hydrology, p. 2.

4. Could the authors provide more diverse visualizations beyond the current response curves, such as spatial error maps on ERA5 or error distributions across different tile compositions? These would help better illustrate where UniTeD performs well and where its limitations arise.

We have added a global spatial error map of UniTeD’s mean absolute error (summed across all predicted outputs) on the ERA5 dataset to Appendix D.8 (Figure 11).

As shown in the new spatial map, the model maintains a low error profile globally (bounded below 20 W/m^2). The geographic distribution of the error aligns with expected physical and structural realities:

  • Homogeneous regions: The model achieves its lowest errors (represented by the darkest regions on the map) in vast, uniform landscapes with low sub-grid heterogeneity, most notably the Sahara Desert, the Arabian Peninsula, and central Australia.
  • Heterogeneous regions: The areas exhibiting relatively higher errors correspond to highly heterogeneous, biologically complex, or topographically varied regions, such as the tropics (e.g., the Amazon and Congo basins) and major mountain ranges (e.g., the Andes). In these regions, high sub-grid land-cover mixing increases the difficulty of prediction.

However, the global error remains tightly bounded even in these complex regions, demonstrating that UniTeD can scale across entirely distinct climatic and geographic landscapes without severe failure.

I thank the authors for their detailed responses and additional experiments. I appreciate the new architectural ablation and standard ERA5 evaluation, which help address some of my concerns.

However, two concerns remain:

  1. Nonlinear inter-tile interactions: The experiments in Appendix C.5 mainly evaluate pure-tile recovery under nonlinear coupling, rather than prediction accuracy on mixed grid cells where nonlinear interactions are present. Therefore, my original concern about the robustness of the linear aggregation assumption remains insufficiently addressed.

  2. Spatial error analysis: While the added spatial error map is helpful, it only shows UniTeD's absolute errors. Comparisons with baselines or error distributions across different tile compositions would better illustrate the method's advantages and limitations.

Overall, I appreciate the authors' efforts and acknowledge the improvements, but I believe these concerns remain important. Therefore, I will maintain my original score.

We thank the reviewer for the follow-up and for the opportunity to address the remaining concerns. Below are our responses:

  • Nonlinear inter-tile interactions: Figure 8 of Appendix C.5 shows the nonlinear coupling stress test on mixed grid cells. The unconstrained baselines (Soft-MoE and FT-Transformer) exhibit instability and error spikes across the coupling spectrum, whereas UniTeD maintains lower error from mild to strong coupling and degrades smoothly as the coupling strength (γ\gamma) increases. This demonstrates that UniTeD provides a stable, well-behaved approximation without the high variance seen in baselines when inter-tile coupling is introduced.

  • Error distribution: In Figure 13 of Appendix D.8, we provide the error distributions across tile compositions for all models under the standard ERA5 evaluation. Grid cells are bucketed by their maximum tile fraction, ranging from the most heterogeneous mixtures (0.00–0.200.00\text{–}0.20) to the most homogeneous cells (0.80–1.000.80\text{–}1.00). Across all models, errors are highest in the most heterogeneous bucket and generally decrease as cells become more homogeneous, illustrating both the challenge of modeling highly mixed cells and the narrower performance gap in less heterogeneous regimes (≥0.4\ge 0.4). UniTeD achieves a lower median error and tighter spread across all buckets, with its largest advantage over baselines in the most heterogeneous tile compositions (0.00–0.200.00\text{–}0.20).

Reviewer_hTggOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UniTeD, a physics-guided neural network for modeling sub-grid heterogeneity in grid-cell prediction. The framework combines a shared encoder, FiLM-conditioned tile-level experts, and a prescribed fraction-weighted aggregation layer to recover latent sub-grid responses from aggregate supervision. The authors provide a theoretical identifiability analysis and evaluate the approach on synthetic benchmarks and global ERA5 surface energy flux prediction. The results demonstrate promising compositional generalization, with reported error reductions of up to 28% compared with the strongest baseline.

Critical strengths

S1. Meaningful and well-motivated research problem.

The paper addresses an important challenge in scientific machine learning by explicitly representing sub-grid heterogeneity, which is often simplified or overlooked in conventional data-driven emulators.

S2. Simple and physically motivated architecture.

The combination of shared representations, FiLM-conditioned experts, and prescribed aggregation offers an intuitive and parameter-efficient approach to incorporating physical structure into neural networks.

S3. Encouraging compositional generalization results.

The synthetic experiments demonstrate the potential benefits of structural disentanglement, while the ERA5 experiments show improvements under shifts in sub-grid composition.

S4. Multiple complementary evaluations.

The paper includes synthetic disentanglement, few-shot transfer, robustness to nonlinear coupling and noisy fractions, and geographical generalization experiments, providing useful evidence of the framework's behavior under different conditions.

Critical weaknesses

W1. The assumption of independent tile-level responses may limit physical validity in coupled land surface systems.

UniTeD assumes that each tile-level response depends only on shared atmospheric forcing and tile-specific attributes, without explicitly accounting for interactions between neighboring surface types. However, real heterogeneous land surfaces exhibit interactions through heat and moisture transport, turbulent mixing, and land–atmosphere feedbacks. Consequently, individual tile responses may depend on the surrounding surface composition, even when aggregate fluxes are computed using area-weighted averaging.

This also raises a concern regarding the interpretation of structural entanglement in Section 3.2. Nonlinear dependence on tile fractions may represent genuine physical interactions rather than spurious statistical correlations. Although Appendix C.5 considers synthetic nonlinear coupling, the prescribed coupling mechanisms may not adequately represent realistic land surface processes. The applicability of the proposed structural assumptions to physically coupled systems therefore remains insufficiently established.

W2. The lack of independent observational validation limits the practical significance of the results.

The empirical evaluation relies primarily on synthetic data and ERA5 surface fluxes, which are generated by a numerical modeling system incorporating tile-based land surface aggregation. Therefore, the reported improvements demonstrate the ability to emulate a numerical system whose aggregation structure closely matches the proposed inductive bias, but do not necessarily establish accurate prediction of observed land–atmosphere exchanges.

Independent validation against site-level measurements, such as FLUXNET or suitable urban flux observations, would substantially strengthen the physical credibility of the framework. Moreover, its practical advantages relative to established process-based land surface models, such as CLMU or SUEWS, remain unclear. Although UniTeD may provide computational benefits as a surrogate model, its value should be discussed in terms of observational accuracy, computational efficiency, and robustness across heterogeneous environments.

W3. The spatiotemporal data coverage and evaluation protocol require further clarification.

The ERA5 experiments use data from only one year, with training samples drawn from the 7th day of six even-numbered months in 2024 and test samples from the corresponding 22nd days. Although the temporal separation helps reduce short-term autocorrelation, the limited sampling of days may not adequately capture the diversity of weather conditions, seasonal transitions, extreme events, or interannual variability.

Furthermore, the paper does not clearly specify the spatial resolution of the processed ERA5 dataset, whether regridding or land masking was applied, or the total number of grid-cell–time samples used for training and evaluation. The sample counts and tile compositions associated with different purity and mixture thresholds are also unclear. These omissions make it difficult to assess the representativeness of the dataset and interpret the non-monotonic error trends in Table 1.

Additional experiments involving multiple years, alternative training-day selections, or temporally independent evaluation periods would provide stronger evidence of temporal robustness and generalization.

W4. The theoretical identifiability analysis and architectural novelty require stronger justification.

Theorem 3.1 establishes latent response identifiability under restrictive assumptions, including local non-degeneracy and sufficiently small variations in forcing and prediction error. However, Appendix D.6 reports highly ill-conditioned local fraction matrices, raising concerns about the stability of latent-process recovery under realistic data conditions.

Furthermore, UniTeD combines established components, including FiLM conditioning, shared experts, and linear aggregation. While their integration is reasonable, the experiments do not fully isolate their respective contributions. Additional comparisons against simpler aggregation-constrained shared networks, together with more complete structural ablations, would help establish the necessity of the proposed architectural design.

Questions for authors

Q1. Physical validity and inter-tile coupling.

How is the assumption of independent tile-level responses justified in heterogeneous land surfaces where lateral heat and moisture transport, turbulent mixing, and land–atmosphere feedbacks introduce interactions between neighboring surface types? Could the authors evaluate UniTeD using physically coupled land surface simulations to distinguish spurious statistical entanglement from genuine physical interactions?

Q2. Real-world validation and comparison with process-based models.

The current evaluation primarily demonstrates UniTeD's ability to approximate ERA5 surface fluxes generated by an existing numerical modeling system. However, it remains unclear whether these improvements translate into more accurate predictions of real-world land–atmosphere exchanges. Could the authors provide independent validation against site-level flux observations, such as FLUXNET or suitable urban flux measurement sites? Furthermore, could the authors clarify the practical advantages of UniTeD over conventional process-based modeling approaches, particularly in terms of computational efficiency, predictive accuracy, and generalization? A comparison with the underlying IFS/HTESSEL land surface scheme, or with suitable models such as CLMU or SUEWS on compatible evaluation tasks, would help establish its practical value.

Q3. Spatial resolution, temporal coverage, and evaluation reliability.

Could the authors specify the spatial and temporal resolutions of the processed ERA5 dataset, any regridding or land masking applied, and the total numbers of training and test samples? Given that training uses only six days in 2024, how sensitive are the results to the choice of training days, and do the conclusions hold across multiple years or independent evaluation periods? Please also report the sample counts and tile compositions for the different thresholds in Table 1 and explain the observed non-monotonic error trends.

Q4. Theoretical assumptions and architectural contributions.

Given the highly ill-conditioned local fraction matrices reported in Appendix D.6, how do the theoretical identifiability guarantees translate into reliable latent-process recovery in practice? Could the authors additionally compare UniTeD with a parameter-matched shared expert using direct feature conditioning and the same fraction-weighted aggregation to isolate the contribution of FiLM conditioning?

4 comments

We thank the reviewer for their constructive feedback and the opportunity to clarify the physical assumptions and operational context underlying UniTeD. Detailed, point-by-point responses to your questions are provided below.

1. Physical validity and inter-tile coupling.

To properly address this concern, it is helpful to clarify the distinction between a Land Surface Model (LSM) and a fully coupled land-atmosphere model, as well as the distinction between sub-grid tiles and grid cells.

Difference between LSM and land-atmosphere model. UniTeD is designed to emulate LSMs (such as HTESSEL, which generates the surface fluxes in ERA5) rather than the entire coupled Earth system, which would require multiple different models for different Earth components. In these operational frameworks, the LSM's explicit role is to compute vertical energy fluxes for individual sub-grid tiles and aggregate them at the grid-cell level [1]. These bulk aggregates are then passed to the atmospheric model. It is the atmospheric model, not the LSM, that accounts for macroscopic dynamics, computing the lateral advection, turbulent mixing, and heat transport between adjacent grid cells, and returning updated meteorological forcings to the LSM to complete the feedback loop, thus working in a coupled manner to ensure both sub-grid resolution inside the grid cell (by the LSM) and interactions throughout the entire domain between grid cells (by the atmospheric model).

Absence of sub-grid topology and assumption of independence. LSMs usually lack intra-grid spatial topology; they track only the fractional coverage of tiles within a cell, not their physical proximity to one another. Because spatial adjacency is unknown, resolving explicit lateral interactions between specific sub-grid tiles is structurally impossible in these models. Physically, this assumption holds because vertical energy fluxes dominate micro-scale lateral exchanges. For instance, on a sunny day, a paved road and an adjacent lawn respond entirely differently to the same solar forcing: the pavement becomes hot to the touch, while the lawn remains relatively cooler. The lateral heat exchange directly between these two surfaces is negligible compared to the massive vertical energy fluxes they independently emit into the atmosphere. Therefore, the assumption of independent tile-level responses is not a limitation of the UniTeD architecture but rather a direct reflection of physical assumptions governing the established systems (such as ERA’s HTESSEL or CESM’s CLM) it is designed to emulate.

[1] Balsamo et al., 2009. A revised hydrology for the ECMWF model: Verification from field site to terrestrial water storage and impact in the integrated forecast system, Section 2. TESSEL Hydrology, p. 2.

2. Real-world validation and comparison with process-based models

We wish to clarify that UniTeD is an emulator for grid-based numerical land surface models, rather than a direct-observation model intended to discover new physics.

Structural mismatch with site-level observations (FLUXNET): Direct evaluation of site-level networks like FLUXNET differs structurally from grid-based simulations. FLUXNET provides isolated, point-based measurements. It therefore lacks the grid-cell aggregates paired with explicit sub-grid fractional coverages (e.g., 40% vegetation, 30% water) required to evaluate a sub-grid-aware method like UniTeD.

Emulation of process-based models, not direct competition (HTESSEL/CLMU): UniTeD is an emulator of process-based models (such as ECMWF's IFS/HTESSEL and NCAR's CESM CLMU). It is not designed to compete with these conventional schemes in terms of discovering new physical phenomena, as these are the state-of-the-art ground-truth grid-based systems used in operational frameworks worldwide. We evaluated UniTeD on ERA5, which uses HTESSEL as its underlying land surface model, because it is the most comprehensive, high-resolution global dataset publicly available at this massive scale. Generating custom global gridded datasets by running full coupled land-surface models (such as HTESSEL or CLMU coupled with the rest of the Earth system) is computationally intensive, requiring significant High-Performance Computing (HPC) resources and massive data storage. Relying on ERA5 provides a standardized, globally recognized, and operationally validated dataset, which ensures our evaluation is reproducible and comparable to state-of-the-art benchmarks. However, because CLMU and HTESSEL share the same fundamental sub-grid tiling paradigms, UniTeD is agnostic to the underlying process model. Just as we trained UniTeD on ERA5, it can be trained directly on CLMU outputs to serve as a fast, differentiable surrogate.

Practical advantages: A few practical advantages of a machine learning model such as UniTeD over conventional process-based models include:

  • (i) Computational efficiency: Machine learning provides a highly efficient, lightweight forward pass compared to the iterative solvers and complex parameterizations required by full process-based models.
  • (ii) Differentiability and parameter calibration: Process-based models heavily rely on heuristic or manual calibration of physical properties and coefficients. A fully differentiable neural architecture enables efficient, gradient-based parameter calibration and downstream integration.

3. Spatial resolution, temporal coverage, and evaluation reliability.

To address these concerns, we have provided the exact dataset specifications, detailed sample distributions, and a new multi-year evaluation to confirm temporal generalization.

ERA5 resolution and land masking. We used the ERA5 dataset at its native 0.25° spatial resolution without applying any regridding. We applied the land mask to exclude all ocean grid cells to focus exclusively on terrestrial sub-grid heterogeneity.

Multi-year evaluation. We used models trained on 2024 and evaluated them on 2023 and 2025 data to verify sensitivity to other years. As shown in the tables below, the performance trends remain consistent across different years, and UniTeD still maintains a lead over the baselines.

2023 Evaluation

ModelTest (mixture).70.80.91
DNN63.56128.11129.24141.51150.35
FT-Transformer70.94129.83131.05140.69145.26
Soft-MoE72.14126.05130.89141.78123.53
Frac-MoE71.73142.99151.48168.41174.10
UniTeD (Ours)69.55120.46125.01121.77114.88
ModelTest(Pure)0.010.050.08
DNN30.17114.70117.33124.91
FT-Transformer36.2088.7295.0699.76
Soft-MoE36.9196.2993.8598.48
Frac-MoE33.4099.82103.24101.28
UniTeD (Ours)34.0980.8381.1277.87

2025 Evaluation

ModelTest (mixture).70.80.91
DNN65.38124.92130.05140.06151.55
FT-Transformer69.27124.08129.46142.03149.52
Soft-MoE69.71124.13129.09140.75143.01
Frac-MoE69.29142.03150.93163.74168.07
UniTeD (Ours)68.18118.98115.42120.9598.93
ModelTest(Pure)0.010.050.08
DNN34.35104.11105.56108.70
FT-Transformer35.5389.9593.1797.71
Soft-MoE37.2097.7496.11113.02
Frac-MoE33.71100.1996.92108.12
UniTeD (Ours)34.1082.5776.3979.05

Sample counts, tile compositions, and non-monotonic error trends. The non-monotonic trend behavior is a statistical artifact of tile composition scarcity and compositional skewing at the distribution tails of the dataset. As we tighten the purity/mixture thresholds, the available evaluation sample size collapses drastically.

Here are the sample counts:

Task 1 (Mixture → Pure) Test Counts: Total count [Tile 1 count, ..., Tile 6 count]

  • Mixture (Test): 418,383 [418383, 418383, 418383, 418383, 418383, 418383]
  • Threshold 0.7: 2,702,483 [121197, 573677, 119398, 716137, 115854, 1056220]
  • Threshold 0.8: 2,243,396 [91780, 385155, 98326, 547230, 91086, 1029819]
  • Threshold 0.9: 1,296,723 [61246, 17972, 76229, 233559, 13330, 894387]
  • Threshold 1.0: 353,277 [15041, 0, 40659, 0, 0, 297577]

As the threshold tightens toward 1.0, not only does the total sample count drop significantly, but the composition becomes highly imbalanced. For instance, some tiles have 0 samples with 100% coverage, skewing the aggregate metric toward a few dominant tile types.

Task 2 (Pure → Mixture) Test Counts:

  • Pure (Test): 407,576 [15041, 22816, 40659, 17901, 13582, 297577] (Note: we used closest to 1 for zero-valued tiles to ensure evaluation coverage)

Note that for the rest of the thresholds below, the mixture condition requires all 6 tiles to satisfy the threshold simultaneously; therefore, the same count applies to every tile.

  • Threshold 0.01: 39,335 [39335, 39335, 39335, 39335, 39335, 39335]
  • Threshold 0.05: 3,815 [3815, 3815, 3815, 3815, 3815, 3815]
  • Threshold 0.08: 329 [329, 329, 329, 329, 329, 329]

UniTeD remains the best performing architecture across all evaluated slices.

4. Theoretical assumptions and architectural contributions.

Regarding the theoretical identifiability guarantees, we first clarify the physical interpretation behind the condition numbers. We then provide the requested ablation to isolate the architectural contribution of FiLM.

Condition numbers. We would like to clarify that the analysis in Appendix D.6 reports a long-tail distribution for the condition numbers κ(A)\kappa(A) of local fraction matrices, with a median of 81.7 and a 95th percentile of 8.31×1048.31 \times 10^4. It is only at the extreme upper tail of the distribution (99th percentile) that the matrices exhibit extreme condition numbers. However, this is physically expected in Earth's geographical domains. Extreme condition numbers occur in regions covered by a single, perfectly homogeneous land-surface type (e.g., a vast desert, the open ocean, or a dense, continuous forest). The local fraction matrix for these regions is largely composed of zeros. Rather than impairing the learning process, this ill-conditioned extreme is actually the most trivial to learn. The target grid cell is no longer a mixture; the network does not need to perform numerical unmixing. Without loss of generality, consider a local batch of grid cells composed entirely of a single tile aa (and 0% of tile bb), with local meteorological inputs xix_i. For any sample ii in this batch, the aggregation equation simplifies to:

1⋅Φa(xi)+0⋅Φb(xi)=yi1 \cdot \Phi_a(x_i) + 0 \cdot \Phi_b(x_i) = y_i

In this extreme case, the aggregate target yiy_i equals the latent response Φa(xi)\Phi_a(x_i). The model directly learns the mapping from xix_i to yiy_i via direct assignment, and the gradient for Φb\Phi_b is safely zeroed out. Thus, the ill-conditioned matrices in the dataset simply represent pure regions where the inverse problem safely collapses.

Furthermore, we note that these condition numbers apply to individual batches randomly sampled during training. Unlike a direct numerical solver that must explicitly invert a static matrix in a single step, a neural network processes these batches iteratively and does not invert the matrices. Therefore, a momentarily ill-conditioned batch does not cause the optimization to fail; rather, it simply contributes to the global loss landscape that the network traverses over multiple iterations and epochs to converge on a robust solution.

FiLM ablation. Per the reviewer’s suggestion, we isolated the contribution of the FiLM conditioning mechanism by replacing it with a parameter-matched shared expert that uses direct feature conditioning. As shown in the tables below, 'UniTeDNoFiLM' with simple direct feature conditioning is consistently less performant than UniTeD with the FiLM generator.

Performance across Purity Thresholds

ModelTest (Mixture).7.8.91.0 (Pure)
UniTeDNoFiLM63.0 ± 0.4120.0 ± 3.3125.7 ± 3.8133.2 ± 6.9122.8 ± 20.4
UniTeD (Ours)62.4 ± 0.8113.4 ± 2.3118.2 ± 2.6122.7 ± 4.092.1 ± 11.4

Zero-Shot Compositional Generalization (Pure → Mixture)

ModelTest (Pure).01.05.08
UniTeDNoFiLM36.0 ± 0.379.6 ± 2.285.4 ± 4.386.8 ± 2.1
UniTeD (Ours)34.5 ± 0.377.1 ± 0.778.2 ± 0.370.3 ± 0.9

Comments on the paper

1 comment

We thank the Area Chair and all reviewers for their constructive feedback and engagement throughout the discussion period. Below, we briefly summarize our responses to the concerns raised by the reviewers:

  • Architectural ablation (Reviewers P5Hz, hTgg): Added the parameter-matched UniTeDNoFiLM ablation (direct feature conditioning / tile identity embeddings with identical physics-prescribed aggregation), confirming that FiLM modulation consistently improves compositional generalization across all purity and mixture thresholds.

  • Standard ERA5 evaluation & compositional error distributions (Reviewer P5Hz): Added a standard in-distribution ERA5 evaluation across the full range of pure and mixed cells (where UniTeD achieves the lowest RMSE at 55.00), alongside a global spatial error map and comparative error distributions across tile-composition buckets (Appendix D.8, Figures 12 & 13) showing UniTeD's lower median error and tighter spread, particularly in the most heterogeneous regimes.

  • Mixed-cell nonlinear inter-tile coupling (Reviewer P5Hz): Extended Appendix C.5 (Figure 8) to evaluate prediction accuracy directly on mixed grid cells under increasing nonlinear coupling (γ∈[0,1]\gamma \in [0, 1]), demonstrating that UniTeD degrades smoothly without the severe error spikes exhibited by baselines.

  • Multi-year & non-adjacent temporal generalization (Reviewer hTgg): Added cross-year evaluations spanning 4- to 6-year gaps (2024→2019,2023,20252024 \rightarrow 2019, 2023, 2025, as well as training from scratch on 2019→2023,20252019 \rightarrow 2023, 2025), confirming that UniTeD's generalization advantage holds across non-adjacent years and is independent of the training year.

  • Dataset specifications & theoretical scope (Reviewers hTgg, TVtD): Added native 0.25∘0.25^{\circ} ERA5 resolution and land-masking details, full per-threshold sample counts explaining tail variance (Appendix D.7), the σmin⁡(AB)\sigma_{\min}(A_{\mathcal{B}}) distribution for local fraction matrices, and a clearer scoping of latent-function recovery claims on ERA5.

We appreciate the reviewers' time and feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.