acceptodds
Withdrawn from ICLR 2027

ClimLoRA: Frequency-Aware Dual-LoRA for Fine-Scale and Physics-Decoupled Climate Downscaling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-resolution meteorological fields are essential for extreme-weather analysis and regional climate studies. In recent years, machine-learning-based super-resolution methods have provided an efficient approach to climate downscaling. However, existing methods typically optimize data reconstruction and physical constraints within a shared parameter space, coupling the two objectives and making it more difficult to achieve both reconstruction accuracy and physical structural consistency. To address this challenge, we propose ClimLoRA, a frequency-aware Dual-LoRA framework for climate downscaling that separates parameter updates for fine-scale reconstruction and physical structure correction and coordinates the two objectives through sequential training. We first formulate climate downscaling as learning the residual between low-resolution inputs and high-resolution targets, and then decouple the learning objectives into two LoRA parameter spaces: one learns fine-scale reconstruction through high-frequency modeling and spherical harmonic spectral constraints, while the other promotes physical consistency through physical structural constraints on spherical grids. We conduct global multivariable joint downscaling experiments under both ERA5-to-ERA5 in-domain and MPI-ESM-to-ERA5 cross-source settings, covering near-surface and upper-air meteorological variables. Experimental results show that ClimLoRA achieves leading overall downscaling performance while maintaining strong reconstruction capability in global downscaling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
climate Downscaling, diffusion models, physics, earth science

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_uzYqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ClimLoRA, a Dual-LoRA framework that separates fine-scale reconstruction and physical consistency optimization for climate downscaling. The method uses frequency-aware constraints for reconstruction and physical diagnostics for structural correction. Experiments on ERA5 and MPI-ESM datasets demonstrate improvements in reconstruction accuracy and physical diagnostic metrics over existing baselines. Ablation studies further support the effectiveness of the proposed components and sequential optimization strategy.

Critical strengths

This paper addresses the potential optimization interference between data reconstruction and physical constraints in global climate downscaling. It proposes a decoupled optimization strategy using separate parameter spaces, which is an interesting and relevant research direction.

Critical weaknesses
  1. Insufficient justification and analysis of the Dual-LoRA design

Although the authors experimentally demonstrate some advantages of decoupled optimization, they do not directly show whether significant gradient conflicts exist between the two learning objectives. Moreover, the current experiments do not clearly distinguish whether the performance improvements come from parameter-space decoupling or sequential training. In particular, the results do not sufficiently explain why two independent LoRA modules are necessary for optimizing different loss functions.

For example, could similar improvements be achieved by using a single LoRA module trained sequentially with LFSRL_{\mathrm{FSR}} and LPhyL_{\mathrm{Phy}}? Alternatively, what would happen if two LoRA modules were trained jointly rather than sequentially?

I suggest including additional ablation studies, such as single-LoRA sequential training and dual-LoRA joint training. An analysis of gradient similarity or gradient conflicts between the two objectives would also provide stronger evidence for the proposed design.

  1. Limited evaluation of physical consistency and extreme weather events

The authors evaluate improvements in several physical diagnostics, including divergence and vorticity. However, these quantities are also directly included in the training objectives of Physics-LoRA. Therefore, the current results mainly demonstrate that Physics-LoRA can better reconstruct the physical quantities used in its training loss, rather than establishing that the predicted fields satisfy independent physical laws. I suggest including at least one physical consistency evaluation that is independent of the training losses to provide stronger evidence for the claimed physical consistency.

In addition, the Introduction highlights extreme-weather analysis as an important motivation, but the experiments mainly report overall RMSE, MB, and LFD, without specific evaluations of extreme weather events. Global average error metrics alone provide limited insight into the model's performance across different regions, climate conditions, and spatial scales, potentially masking substantial regional errors and poor reconstruction of local-scale features and extreme events.

For example, high-frequency reconstruction may smooth extreme temperatures, underestimate strong winds, or introduce errors in the spatial structures of extreme events. I suggest evaluating the model on extreme heat and strong wind events, using metrics such as extreme percentiles (P95/P99), peak errors, and spatial reconstruction accuracy.

  1. Limited comparisons with physics-guided downscaling methods

Physical consistency is highlighted as one of the main contributions of this paper. However, the current baselines mainly focus on general super-resolution and data-driven climate downscaling methods.

I suggest including comparisons with relevant physics-guided downscaling methods, where applicable, or implementing physics-constrained baselines using the same backbone architecture.

Such comparisons would help clarify the advantages of Dual-LoRA over existing physics-guided approaches in terms of reconstruction accuracy, physical consistency, and optimization efficiency, thereby strengthening the experimental evidence for the paper's contributions.

Questions for authors

Please see Weakness

4 comments

Thank you for your detailed comments. We add adapter ablations, independent physical and extreme-weather evaluations, and a physics-guided baseline comparison, and clarify the Dual-LoRA design.

Q1. Effectiveness of Sequential Dual-LoRA

We compare sequential single-LoRA training (Single), jointly trained Dual-LoRA (Joint), and sequential Dual-LoRA (ClimLoRA) under matched parameter and training budgets. ClimLoRA achieves the lowest RMSE and LFD across all variables. For V10, T2M, and MSL, ClimLoRA reduces RMSE by 1.22%, 1.73%, and 1.62% relative to Single, and by 1.03%, 1.17%, and 3.34% relative to Joint, respectively.

Single updates previously learned reconstruction parameters during the second stage, while Joint may introduce interference between objectives. These results suggest that sequential training with parameter freezing helps preserve reconstruction and reduce interference. Our focus is potential optimization interference during shared-parameter updates; we do not assume that the two objectives necessarily have strongly conflicting gradients.

MethodU10 RMSE ↓U10 LFD ↓V10 RMSE ↓V10 LFD ↓T2M RMSE ↓T2M LFD ↓MSL RMSE ↓MSL LFD ↓
Single0.269211.19880.263111.15800.386611.927416.22019.3581
Joint0.268111.19730.262611.15390.384411.915116.50919.3857
ClimLoRA0.267711.19550.259911.13900.379911.891515.95719.3229

Q2. Physical Consistency and Extreme Weather

Independent Physical Evaluation

We additionally evaluate the horizontal temperature advection error, A_T, which is excluded from the training loss and measures wind–temperature coupling. ClimLoRA reduces A_T by 11.93% relative to Base and 7.02% relative to FSR-LoRA, and outperforms Single, Joint, and GeoFAR. These results suggest that the physics branch improves physical structures beyond the training constraints.

MetricBaseFSR-LoRAClimLoRASingleJointGeoFAR
A_T (K/day) ↓2.05681.94831.81151.84161.83931.9254

Extreme Weather Evaluation

We evaluate extreme heat (T2M) and strong winds (W10 = √(U10² + V10²)) across all 1,464 six-hourly timestamps from 2020, using shared grid-point-wise P95/P99 thresholds from ground truth. Metrics include RMSE over exceedance regions, peak MAE, CSI, and FSS (5×5 neighborhood).

Compared with FSR-LoRA, ClimLoRA reduces P99 RMSE by 5.59% and 5.63% and peak MAE by 9.72% and 21.01% for T2M and W10, respectively, while achieving higher CSI and FSS. These results indicate more accurate reconstruction of the intensity, peaks, and spatial distributions of extreme heat and strong winds. Details have been added to Appendix A.5.5.

T2M

MethodP95 RMSE ↓P99 RMSE ↓Peak MAE ↓P99 CSI ↑P99 FSS₅ ↑
Base0.59090.64740.72200.67290.9416
FSR-LoRA0.43170.47800.52140.71150.9576
GeoFAR0.45360.50730.51490.69730.9492
ClimLoRA0.41700.45130.47070.72280.9679

W10

MethodP95 RMSE ↓P99 RMSE ↓Peak MAE ↓P99 CSI ↑P99 FSS₅ ↑
Base0.55590.67220.62750.67340.9596
FSR-LoRA0.37370.44380.50510.76420.9736
GeoFAR0.38190.45220.44060.75800.9717
ClimLoRA0.35660.41880.39900.76720.9745

Q3. Comparison with Physics-Guided Methods

We add PhyDL [1] as a physics-guided downscaling baseline. PhyDL combines physical equations with latent-force parameterization and guides prediction through physics-constrained losses. ClimLoRA achieves lower RMSE and LFD across all variables, reducing RMSE by approximately 13.0%–25.8% relative to PhyDL. This comparison supports ClimLoRA’s advantages in reconstruction accuracy and spectral recovery.

MethodU10 RMSE ↓U10 MBU10 LFD ↓V10 RMSE ↓V10 MBV10 LFD ↓T2M RMSE ↓T2M MBT2M LFD ↓MSL RMSE ↓MSL MBMSL LFD ↓
PhyDL0.30960.013011.48060.2989−0.001711.40900.4521−0.003412.236521.4988−3.641119.8781
ClimLoRA0.26770.001811.19550.25990.001111.13900.37990.001211.891515.95680.111319.3229

References

[1] Physics-Guided Learning of Meteorological Dynamics for Weather Downscaling and Forecasting, KDD, 2025.

I appreciate the authors' efforts in addressing my concerns and providing additional experimental results. The new ablation studies, physical diagnostics, and extreme-weather evaluations strengthen the empirical evidence for ClimLoRA's effectiveness.

However, several concerns remain:

  1. While the additional ablations demonstrate empirical improvements, they do not provide sufficient insights into why the Dual-LoRA design is necessary or how it actually mitigates optimization interference. The main contribution remains largely empirical rather than mechanistic.

  2. The performance improvements over Single and Joint are relatively marginal for several variables. Without results from multiple random seeds or confidence intervals, it is difficult to determine whether these improvements are statistically robust.

  3. Although horizontal temperature advection is excluded from the training loss, it is directly derived from the predicted wind and temperature fields. Therefore, improvements in this metric may simply result from lower prediction errors in these variables, rather than demonstrating improved physical consistency or adherence to physical laws.

Overall, I acknowledge that the rebuttal has strengthened the experimental validation of the proposed approach. However, my primary concern regarding the limited methodological novelty and mechanistic understanding remains unresolved. Therefore, I will maintain my original score.

Thank you for your comments.

1. Performance and Seed Robustness

We have added results from three independent random seeds. For MSL, the mean RMSE reduction relative to Joint is 0.552, approximately 15.35 times the larger standard deviation of the two methods. Even for U10, which shows the smallest improvement, the mean reduction of 0.0004 is approximately 3.25 times the corresponding larger standard deviation. These mean improvements exceed the reported seed-level standard deviations, supporting their reproducibility across seeds.

The gains vary across variables, with particularly clear improvements in MSL: RMSE decreases by 1.62% and 3.34% relative to Single and Joint, respectively. T2M RMSE also decreases by 1.73% relative to Single. Meanwhile, our method further reduces physical diagnostic errors while maintaining strong reconstruction performance, which is another important benefit of the design.

Furthermore, this experiment uses 4× downscaling from 1° to 0.25°. The input grid is already relatively fine, and the reconstruction baselines are strong, leaving relatively limited room for further reductions in overall RMSE. Coarser input grids would provide greater room for improvement.

MethodU10 RMSE ↓V10 RMSE ↓T2M RMSE ↓MSL RMSE ↓
Single0.2692 ± 0.0000630.2631 ± 0.0001010.3866 ± 0.00002516.220 ± 0.011255
Joint0.2681 ± 0.0001230.2626 ± 0.0001580.3844 ± 0.00020316.509 ± 0.035960
ClimLoRA0.2677 ± 0.0000900.2599 ± 0.0002610.3799 ± 0.00001915.957 ± 0.035416

2. Temperature advection

Temperature advection depends on wind velocity, temperature gradients, and their spatial correspondence. Lower errors in individual variables therefore do not necessarily guarantee lower advection errors. Compared with FSR-LoRA, ClimLoRA with physical constraints improves both prediction accuracy and the advection diagnostic excluded from training, supporting its role in reconstructing wind–temperature coupling. Even if part of the improvement arises from more accurate wind and temperature predictions, it still represents a benefit of the physical constraints.

Thank you for your detail suggestions. Our previous presentation did not sufficiently clarify physical consistency. We clarify these points below.

We introduce physical structural constraints based on kinetic energy, divergence, and vorticity to improve downscaling accuracy and physical structural consistency.

  1. Our definition follows Saccardi et al. [1]: diagnostic quantities such as kinetic energy, divergence, and vorticity are used to assess how accurately predictions recover the physical structures of the reference fields.
  2. Regarding independent evaluation metrics, Physics-Informed Diffusion Models [2] likewise evaluates generated samples using the physical residuals constrained during training.
  3. Using the same physical diagnostics for training and evaluation does not, by itself, imply insufficient validation. We evaluate physical consistency on held-out test samples to assess generalization beyond the training data.

References:

[1] Saccardi et al. Assessing the Geographic Generalization and Physical Consistency of Generative Models for Climate Downscaling. NeurIPS 2025 AI for Science Workshop.

[2] Bastek, Sun, and Kochmann. Physics-Informed Diffusion Models. ICLR, 2025.

Reviewer_oc4uOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces a dual low-rank adaptation mechanism to train downscaling models of climate and weather signals defined over a regular pixel grid. Original data are downscaled so that they can be super-resolved back to the original resolution (i.e., full supervision is available). The base training focuses on low-frequency signal components, the first LoRA stage on high frequencies, and the second one on physical consistency. During inference, physical consistency and high frequencies can be combined with different weightings.

Critical strengths

The design choices are sensible and substantiated by experimentation. I particularly like the idea of incorporating physical constraints. The model is compared with multiple baselines, and the ablations justify the addition of all components.

Critical weaknesses

I think it would be nice to provide more details regarding what each physical constraint does and why it is intuitively + physically a good choice. It may be obvious to a domain expert, but it is not obvious to me, and I suppose to many others.

The paper surely has merit; hence I lean towards acceptance. Nevertheless, technical contributions are a bit limited and are a combination of existing techniques.

Questions for authors
  1. There are no details regarding how the low- and high-pass filters are implemented. Are they applied in Fourier space? Is the Euclidean or Spherical transform used? Do HP and LP have any partial overlap? How is their cutoff frequency selected?

  2. Just to confirm, is the LP explicitly used? At line 185 it is introduced as a way to observe the low-frequency components of the target residual (but it does not appear to be the actual target); at lines 213- 214, it appears that the model recovers the large-scale structures through the low-frequency constraints, which I suppose refers to this low-pass version. In the appendix at line 794, it says the model learns the filtered component. Could you please disambiguate?

  3. I would advise simplifying the math block in Figure 1, PhysLora, maybe just keep only the loss terms rather than all the math for vorticity, energy, etc.

  4. I would argue that conducting experiments is a standard requirement to prove the hypothesis made when formulating the core argument of the paper and to substantiate the technical contributions; therefore, the last contribution is a bit debatable, but this may be a personal view.

1 comment

Thank you for your suggestions. We clarify filtering and low-frequency supervision, explain the presentation of physical constraints, and refine the contribution statement.

Q1. Low-Pass and High-Pass Filtering

The LP and HP operators in Equations (2) and (5) are implemented directly in the spatial domain on the latitude–longitude grid, without FFT or SHT. Specifically, LP uses channel-wise Gaussian convolution with a 9×9 kernel, a standard deviation of 2, and a stride of 1. Replication padding of four grid cells on all sides preserves spatial dimensions. HP is obtained by subtracting the low-pass component from the input: HP(z) = z − LP(z). This Gaussian filtering imposes no hard cutoff frequency; the kernel width and standard deviation control the smoothing scale, and the frequency responses of LP and HP overlap in the transition region.

Q2. Low-Frequency Supervision

The LP operator defined above is explicitly used to construct the Base supervision target. We first compute the residual XL−XHX^L-X^H, then apply LP to obtain the low-frequency residual used to supervise Base. Thus, “recovering large-scale structures” in the main text and “learning the filtered component” in the appendix describe the same low-frequency residual supervision process.

Q3. Physical Constraints

Thank you for the suggestion. We agree that simplifying the presentation can improve readability. At the same time, we hope to retain the specific physical diagnostics, corresponding losses, and their physical basis to help readers less familiar with the domain understand Phy-LoRA’s physical constraints more intuitively.

Q4. Main Contribution Statement

We agree that experiments are necessary to validate a method and should not be presented as an independent contribution. We have therefore revised the final contribution to state a concrete performance finding: ClimLoRA significantly outperforms state-of-the-art deep learning downscaling models in global multivariable joint downscaling under both the in-domain ERA5-to-ERA5 and cross-source MPI-ESM-to-ERA5 settings.

Reviewer_Xe6bOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ClimLoRA, a frequency-aware Dual-LoRA framework for climate downscaling that aims to decouple fine-scale reconstruction from physical consistency optimization. The method first trains a low-frequency residual reconstruction model, then sequentially introduces Fine-scale Reconstruction LoRA (FSR-LoRA) and Physics LoRA (Phy-LoRA). FSR-LoRA employs high-frequency reconstruction and spherical harmonic spectral constraints, while Phy-LoRA uses meteorological structural constraints, including kinetic energy, divergence, vorticity, and spatial gradients. The framework also introduces adjustable inference coefficients to control the contributions of the two LoRA branches. Experiments on ERA5-to-ERA5 and MPI-ESM-to-ERA5 downscaling tasks demonstrate improvements in reconstruction accuracy and frequency-domain consistency compared with the evaluated baselines.

Critical strengths

1.Practically meaningful problem formulation. The paper addresses an important challenge in climate downscaling: balancing fine-scale reconstruction accuracy with physical structural consistency. Separating these objectives into independent parameter spaces is a reasonable approach to reducing potential optimization interference, particularly for global meteorological fields.

2.Relatively comprehensive experimental evaluation. The paper evaluates both ERA5-to-ERA5 in-domain and MPI-ESM-to-ERA5 cross-source downscaling, covering multiple meteorological variables and spatial resolutions. It also provides component ablations and a comparison with joint optimization under matched trainable parameter counts and training budgets, offering useful empirical evidence for the proposed design.

Critical weaknesses
  1. Insufficient validation of the proposed physics-aware decoupling. The physical losses mainly match diagnostic quantities derived from reference fields, and Figure 3(b) evaluates largely the same quantities without independent physical-consistency tests. Moreover, the comparison with joint optimization changes both the training strategy and parameter separation, making it difficult to attribute the improvements specifically to decoupling. These limitations weaken the evidence supporting the paper's central methodological claim.

  2. Unclear computational efficiency of Dual-LoRA. Although both LoRA branches use rank 64, the paper does not report actual trainable parameter counts, total training costs, memory consumption, or inference latency. Without quantitative comparisons against standard LoRA or full fine-tuning, it remains unclear whether the additional branches and training stages provide a meaningful performance–efficiency trade-off.

Questions for authors
  1. How much of the improvement comes specifically from separating the LoRA parameter spaces rather than sequential training? Can the authors provide comparisons with sequential single-LoRA and jointly optimized Dual-LoRA under matched conditions?

  2. Beyond improvements in physical diagnostic errors, how does the proposed Phy-LoRA ensure physical consistency? Have the authors evaluated independent physical properties, such as conservation-related constraints?

  3. Given that both LoRA branches use rank 64, what are the actual trainable parameter counts, training costs, memory requirements, and inference latency? How do these compare with standard LoRA and full fine-tuning?

  4. Since the method uses a fixed diffusion timestep and deterministic residual prediction, what advantages does the diffusion-based formulation offer over a conventional residual U-Net?

  5. How are the low-pass and high-pass operators in Equations (2) and (5) implemented? What cutoff frequencies and boundary-handling strategies are used, and how sensitive are the results to these choices?

1 comment

Thank you for your helpful questions. We provide adapter comparisons, independent physical evaluation, and computational cost measurements, and clarify the diffusion formulation and filtering implementation.

Q1. Effectiveness of Sequential Dual-LoRA

We compare sequential single-LoRA training (Single), jointly trained Dual-LoRA (Joint), and sequential Dual-LoRA (ClimLoRA) under matched parameter and training budgets. ClimLoRA achieves the lowest RMSE and LFD across all variables. For V10, T2M, and MSL, ClimLoRA reduces RMSE by 1.22%, 1.73%, and 1.62% relative to Single, and by 1.03%, 1.17%, and 3.34% relative to Joint, respectively. Single updates previously learned reconstruction parameters during the second stage, while Joint may introduce interference between objectives. These results suggest that sequential training with parameter freezing helps preserve reconstruction and reduce interference.

MethodU10 RMSE ↓U10 LFD ↓V10 RMSE ↓V10 LFD ↓T2M RMSE ↓T2M LFD ↓MSL RMSE ↓MSL LFD ↓
Single0.269211.19880.263111.15800.386611.927416.22019.3581
Joint0.268111.19730.262611.15390.384411.915116.50919.3857
ClimLoRA0.267711.19550.259911.13900.379911.891515.95719.3229

Q2. Additional Physical Error Metrics

We additionally evaluate the horizontal temperature advection error, A_T, which is excluded from the training loss and measures wind–temperature coupling. ClimLoRA reduces A_T by 11.93% relative to Base and 7.02% relative to FSR-LoRA, suggesting improved physical consistency beyond the training constraints. In Figure 3(b, c) of the revised main text, ClimLoRA outperforms Single and Joint across all six physical diagnostics and reduces horizontal kinetic energy, divergence, and vorticity errors by 9.64%, 3.18%, and 6.52% versus FSR-LoRA, respectively.

MetricBaseFSR-LoRAClimLoRASingleJointGeoFAR
A_T (K/day) ↓2.05681.94831.81151.84161.83931.9254

Q3. Parameter Counts and Computational Costs

ClimLoRA’s two rank-64 branches add 7.272M parameters in total. During the FSR stage, Base is frozen and only FSR-LoRA’s 3.636M parameters are updated. During the Phy stage, both Base and FSR-LoRA are frozen, and only Phy-LoRA’s 3.636M parameters are updated. Standard LoRA and full fine-tuning update 3.636M and 4.526M parameters, respectively. Training takes approximately 25, 17, and 15 hours for ClimLoRA, standard LoRA, and full fine-tuning, respectively, with peak GPU memory usage of 18.17, 16.94, and 16.77 GiB. Under the same conditions, inference over all 1,464 samples from 2020 takes approximately 60, 40, and 35 minutes, respectively.

Q4. One-Step Residual Diffusion

Previous work has shown that residual diffusion can help recover fine-scale structures that are smoothed by deterministic U-Net regression [1]. OSEDiff further demonstrates that one-step diffusion can recover details while reducing the computational cost of iterative sampling [2]. Motivated by these findings, we adopt a one-step residual formulation combined with frequency constraints to balance fine-scale reconstruction and inference efficiency.

Q5. Low-Pass and High-Pass Filtering

The low-pass operator (LP) in Equations (2) and (5) is implemented as a channel-wise spatial Gaussian convolution on the latitude–longitude grid, with a 9×9 kernel, a standard deviation of 2, and a stride of 1. The high-pass operator is defined as HP(z) = z − LP(z). Neither operator involves FFT or SHT. We use replication padding of four grid cells on all sides to preserve spatial dimensions. No hard cutoff frequency is imposed; instead, the kernel width and standard deviation control the smoothing scale. The frequency responses of LP and HP overlap in the transition region.

References

[1] Residual Corrective Diffusion Modeling for km-Scale Atmospheric Downscaling, Commun. Earth Environ., 2025.

[2] One-Step Effective Diffusion Network for Real-World Image Super-Resolution, NeurIPS, 2024.

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

This paper proposes ClimLoRA, a frequency-aware Dual-LoRA framework for climate downscaling. It separates fine-scale reconstruction and physical structure correction into two LoRA branches, trained sequentially on top of a low-frequency residual Base model. Experiments on ERA5-to-ERA5 and MPI-ESM-to-ERA5 downscaling demonstrate competitive reconstruction accuracy and frequency-domain consistency.

Critical strengths
  1. The paper addresses an important problem of balancing fine-scale reconstruction and physical consistency in climate downscaling.
  2. The proposed framework combines decoupled LoRA updates, spherical harmonic spectral constraints, and spherical physical operators in a reasonably motivated design.
Critical weaknesses
  1. Limited methodological novelty. Frequency-aware reconstruction, spherical representations, physics-based losses, and LoRA-based adaptation have been explored previously. Although joint optimization is compared, the benefits of parameter decoupling versus sequential training are not fully isolated. The authors should better justify the unique contribution of Dual-LoRA.
  2. Insufficient physical consistency validation. The physics loss mainly matches reference-field diagnostics rather than explicitly enforcing physical conservation laws. Moreover, evaluation largely relies on the same diagnostics used during training, leaving the claimed physical consistency insufficiently validated.
Questions for authors
  1. Can the authors compare single-LoRA sequential training, jointly trained Dual-LoRA, and sequential Dual-LoRA under matched budgets?
  2. Can independent physical consistency metrics be provided beyond the diagnostics directly optimized during training?
  3. The original MPI-ESM resolution is 5.625°, but the inputs are interpolated to 1°/2° for 4×/8× downscaling. How is the effective resolution interpreted, and how is the physical correspondence between MPI-ESM and ERA5 samples ensured?
  4. The one-step diffusion formulation resembles deterministic residual regression. What advantages does it provide over a conventional residual U-Net?
  5. Why is the spherical harmonic transform truncated at degree 64?
4 comments

Thank you for your comments. We add controlled adapter comparisons and independent physical validation, and clarify data resolution, sample correspondence, and modeling choices.

Q1. Effectiveness of Sequential Dual-LoRA

We compare sequential single-LoRA training (Single), jointly trained Dual-LoRA (Joint), and sequential Dual-LoRA (ClimLoRA) under matched parameter and training budgets. ClimLoRA achieves the lowest RMSE and LFD across all variables. For V10, T2M, and MSL, ClimLoRA reduces RMSE by 1.22%, 1.73%, and 1.62% relative to Single, and by 1.03%, 1.17%, and 3.34% relative to Joint, respectively. Single updates previously learned reconstruction parameters during the second stage, while Joint may introduce interference between objectives. These results suggest that sequential training with parameter freezing helps preserve reconstruction and reduce interference.

MethodU10 RMSE ↓U10 LFD ↓V10 RMSE ↓V10 LFD ↓T2M RMSE ↓T2M LFD ↓MSL RMSE ↓MSL LFD ↓
Single0.269211.19880.263111.15800.386611.927416.22019.3581
Joint0.268111.19730.262611.15390.384411.915116.50919.3857
ClimLoRA0.267711.19550.259911.13900.379911.891515.95719.3229

Q2. Additional Physical Error Metrics

We additionally evaluate the horizontal temperature advection error, A_T, which is excluded from the training loss and measures wind–temperature coupling. ClimLoRA reduces A_T by 11.93% relative to Base and 7.02% relative to FSR-LoRA, suggesting improved physical consistency beyond the training constraints. In Figure 3(b, c) of the revised main text, ClimLoRA outperforms Single and Joint across all six physical diagnostics and reduces horizontal kinetic energy, divergence, and vorticity errors by 9.64%, 3.18%, and 6.52% versus FSR-LoRA, respectively.

MetricBaseFSR-LoRAClimLoRASingleJointGeoFAR
A_T (K/day) ↓2.05681.94831.81151.84161.83931.9254

Q3. Data Sources and Preprocessing

MPI-ESM1-2-HR has a native resolution of approximately 100 km, rather than 5.625°. The 4×/8× factors describe input-to-output grid-spacing ratios (1°/2° to 0.25°). Interpolation changes the sampling grid without adding physical information or improving effective resolution.

We harmonize variables, units, vertical levels, and coordinates, then pair samples at shared six-hourly timestamps. This pairing does not guarantee synchronized weather states between MPI-ESM and ERA5, but provides a common reference for cross-source comparison and evaluation. Our setup and preprocessing follow ClimaX [1].

Q4. One-Step Residual Diffusion

Previous work has shown that residual diffusion can help recover fine-scale structures that are smoothed by deterministic U-Net regression [2]. OSEDiff further demonstrates that one-step diffusion can recover details while reducing the computational cost of iterative sampling [3]. Motivated by these findings, we adopt a one-step residual formulation combined with frequency constraints to balance fine-scale reconstruction and inference efficiency.

Q5. Spherical Harmonic Truncation

Lower spherical harmonic cutoffs limit fine-scale representation [4]. In our preliminary experiments, increasing the truncation further yielded limited performance gains while continuing to increase computational cost. We therefore retain 64 spherical harmonic degrees, ℓ = 0, …, 63, to balance the representation of spatial detail on the sphere with computational efficiency.

References

[1] ClimaX: A Foundation Model for Weather and Climate, ICML, 2023.

[2] Residual Corrective Diffusion Modeling for km-Scale Atmospheric Downscaling, Commun. Earth Environ., 2025.

[3] One-Step Effective Diffusion Network for Real-World Image Super-Resolution, NeurIPS, 2024.

[4] Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks, ICLR, 2024.

Thank you for the detailed response and the additional experiments. The controlled Dual-LoRA comparative experiments, along with the newly introduced temperature advection metric, have addressed some of my questions and concerns. However,

  1. First, beyond the additional experiments, the authors have not explicitly addressed the weaknesses I initially raised. I remain unclear regarding the specific contributions and the physical consistency of the approach.

  2. Since temperature advection is a function of the wind field and temperature gradient, variables that are already constrained during the training process. It is questionable whether this can serve as robust evidence for independent physical consistency.

  3. Furthermore, I suggest that the authors clarify the source of the MPI-ESM input data: is it derived from the original MPI-ESM1-2-HR grid, or from data resampled to a 5.625° grid? There appears to be an inconsistency between the original manuscript and the response.

In summary, while the response strengthens the experimental evidence, my concerns regarding the methodological innovation and physical validation persist. Therefore, I am inclined to maintain my original score.

Thank you for the further feedback.

  1. Main Contributions and Physical Consistency. Our core contribution is a staged Dual-LoRA training strategy that separately learns fine-scale reconstruction and physical structure correction, improving both reconstruction accuracy and physical structural consistency. To address the limited attention to physical structure in existing methods that primarily emphasize numerical reconstruction, we introduce physical constraints. Our experiments show that these constraints improve physical structure while further reducing reconstruction errors.

  2. Significance of Physical Diagnostic Errors. Our objective is to demonstrate that introducing physical constraints reduces reconstruction errors while improving physical consistency. The original constraints separately target wind-field structure and temperature gradients, whereas temperature advection measures their coupling and is not directly included in the training objective. Its reduced error therefore provides complementary evidence that physical constraints improve cross-variable coupling structures.

  3. Source of MPI-ESM Data. We use data from the native MPI-ESM1-2-HR grid, whose nominal atmospheric resolution is approximately 100 km, and directly regrid these data to a 1° input grid. We previously made an incorrect conversion of the native 100 km resolution to 5.625°, and we have corrected this error.

Thank you for your detail suggestions. Our previous presentation did not sufficiently clarify our contributions or the meaning of physical consistency. We clarify these points below.

Methodological contributions:

  1. We introduce physical structural constraints based on kinetic energy, divergence, and vorticity to improve downscaling accuracy and physical consistency.
  2. To address interference between reconstruction and physical objectives during joint optimization, we propose a sequential dual-LoRA optimization strategy.

Physical consistency:

  1. Our definition follows Saccardi et al. [1]: diagnostic quantities such as kinetic energy, divergence, and vorticity are used to assess how accurately predictions recover the physical structures of the reference fields.
  2. Regarding independent evaluation metrics, Physics-Informed Diffusion Models [2] likewise evaluates generated samples using the physical residuals constrained during training.
  3. Using the same physical diagnostics for training and evaluation does not, by itself, imply insufficient validation. We evaluate physical consistency on held-out test samples to assess generalization beyond the training data.

References:

[1] Saccardi et al. Assessing the Geographic Generalization and Physical Consistency of Generative Models for Climate Downscaling. NeurIPS 2025 AI for Science Workshop.

[2] Bastek, Sun, and Kochmann. Physics-Informed Diffusion Models. ICLR, 2025.

Reviewer_138TOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ClimLoRA proposes a staged framework for global, multivariable climate downscaling. A deterministic residual U-Net first learns low-frequency corrections; a fine-scale LoRA then learns high-frequency reconstruction with spherical harmonic spectral supervision; finally, a physics LoRA matches spatial diagnostics while earlier parameters remain frozen. Two inference coefficients adjust the resulting corrections. The paper reports lower RMSE and Log Frequency Distance than the evaluated baselines on ERA5-to-ERA5 and MPI-ESM-to-ERA5 tasks, together with improved physical diagnostic errors relative to the Base model and joint optimization. Its main contribution is the proposed coordination of reconstruction and physical structure objectives through sequential adapter training.

Critical strengths
  1. The reported reconstruction improvements are consistent across variables and experimental settings. In Table 1, ClimLoRA improves T2M RMSE from GeoFAR’s 0.4038 to 0.3799 and MSL RMSE from 17.8289 to 15.9568, corresponding to approximately 5.9% and 10.5% reductions. Tables 2 and 4 also report the lowest RMSE across all five evaluated variables. These results provide a substantive empirical motivation for the framework, although the interpretation of the cross-source experiments requires clarification below.

  2. The treatment of spherical geometry is technically relevant to global downscaling. The spectral objective uses spherical harmonics, and Appendix A.4.2 includes the geometric correction terms required for spherical divergence and vorticity. Table 3 separately compares SHT with FFT and spherical with planar physical operators, reporting improvements from both choices. These comparisons support the value of adapting the representation and diagnostics to the domain.

  3. The paper makes a useful effort to examine the training strategy and inference controls. Section 4.4 explicitly matches trainable parameter count and total training budget when comparing joint and sequential optimization. Appendix A.5.2 reports coefficient sensitivity, including degradation under excessive correction. These analyses provide preliminary evidence beyond the main baseline comparisons and help identify the operating range of the proposed controls.

Critical weaknesses
  1. The validity of the cross-source pairing is unresolved. Appendix A.1 constructs MPI-ESM/ERA5 pairs using identical six-hourly timestamps but does not identify the MPI-ESM experiment, ensemble member, or any initialization or nudging that would synchronize weather states. If these are free-running CMIP6 simulations, matching timestamps does not establish correspondence between individual simulated and observed weather events. This pairing problem is explicitly discussed in related downscaling research (Schmidt et al., 2025 (https://arxiv.org/html/2412.15361v4)). Under that condition, pointwise RMSE would not directly measure recovery of the input simulation’s corresponding high-resolution weather. Dataset provenance and evidence of coarse-scale anomaly alignment are necessary to interpret Tables 2 and 4. Also, both input grids originate from the same stated 5.625° fields: although Appendix A.5.1 correctly defines 4× and 8× as grid-spacing ratios, these settings retain the same underlying source information. This limits the interpretation of 8× as a substantially harder information-recovery task. Training and testing on the same source pair also supports a narrower generalization claim than transfer to an unseen source.

  2. The comparison does not isolate parameter separation from sequential training. Freezing earlier parameters in Eqs. (3)–(4) preserves their values, while the additional adapter can still change the reconstructed output through Eq. (13). Section 4.4 provides encouraging evidence that the complete sequential procedure outperforms the chosen joint procedure, but it changes both the parameter organization and the loss schedule. It therefore does not establish that separate adapter parameters specifically reduce optimization interference. A shared-adapter control using the same staged objectives and learning-rate schedule, with comparable capacity and training budget, would address this alternative explanation. Reporting variation across a small number of runs would also help interpret the approximately 0.695 versus 0.707 normalized RMSE difference in Figure 3(a). Alternatively, the mechanistic claim should be narrowed to the demonstrated advantage of the complete training procedure.

  3. The frequency evaluation does not establish recovery at the claimed fine scales. LFD in Eq. (16) aggregates squared differences between complex Fourier coefficients before applying a logarithm. With the full spectrum, the quantity inside the logarithm is proportional to spatial squared error by Parseval’s identity. The reported frequency mask changes that equivalence, but the resulting scalar still does not show where spectral improvements occur. Consequently, lower RMSE and LFD provide limited independent evidence for improved high-frequency structure. A degree-resolved spherical spectrum or spectral-error plot for the Base, FSR-LoRA, full model, and strongest baseline would directly test this central claim using existing predictions. The exclusion of the first and last unshifted FFT indices also needs justification. Latitude-weighted RMSE would be a useful secondary sensitivity check, given the explicit unweighted evaluation in Appendix A.3.

  4. The incremental physical benefit is incompletely demonstrated. Equation (12) supplies supervised matching of kinetic energy, wind derivatives, and scalar gradients. These are meaningful spatial diagnostics within the stated single-time-step scope. However, Figure 3(b) compares the complete method and joint optimization against the low-frequency Base; it does not report diagnostic errors for FSR-LoRA alone or for the strongest reconstruction baseline. Thus, the additional physical benefit attributable to Phy-LoRA remains unclear. Table 5 likewise reports reconstruction metrics without physical diagnostics, leaving the claimed adjustable reconstruction–physics trade-off untested. Absolute diagnostic errors for the existing intermediate and final models, together with a few coefficient settings, would resolve this concern without requiring temporal modeling or new datasets.

  5. Several implementation choices essential to the method remain unspecified. Appendix A.2 provides architecture, adapter ranks, optimization settings, and hardware, but I could not find concrete definitions and cutoffs for LP/HP, the degree weights w_l, the fixed physical normalization scales, or the allocation of training epochs across the Base and adapter stages. The numerical derivative stencils and treatment of polar singularities are also unspecified. These details affect the central design: for example, the HP cutoff must be understood relative to the reported SHT truncation at 64. The authors should also clarify whether R_f in Eq. (13) denotes the raw network residual or the effective residual after the high-pass correction defined before Eq. (5). These omissions prevent an unambiguous reconstruction of the algorithm from the supplied materials. Despite the promising reconstruction results, the unresolved cross-source evaluation and attribution of the central mechanism leave me leaning against acceptance on the current evidence.

Questions for authors
  1. Which exact MPI-ESM model, experiment, ensemble member, and data product were used? What makes its six-hourly states correspond to ERA5 weather states? Can you report coarse-scale input–target anomaly correlations after removing the seasonal climatology?

  2. Do you have a shared-adapter experiment using the same sequential objectives and learning-rate schedule as ClimLoRA? Please specify the stage budgets, joint-loss weights, capacity matching, and available variation across runs.

  3. Can you provide physical diagnostic errors and degree-resolved spectral errors for Base, FSR-LoRA, full ClimLoRA, and GeoFAR, using existing test predictions? How do physical diagnostic errors change across the inference coefficients already evaluated in Table 5?

  4. Can you give executable definitions or concise pseudocode for LP/HP, spectral weighting, physical normalization, derivative boundary handling, and the residuals combined in Eq. (13), including how filtering is applied during training and inference?

2 comments

Thank you for your comments. We clarify data pairing and implementation and expand the controlled comparisons and physical–spectral analyses.

Q1. Data Sources and Sample Alignment

We use CMIP6 MPI-ESM1-2-HR data with experiment ID historical, ensemble member r1i1p1f1, table ID 6hrPlevPt, native grid label gn, and six-hourly sampling. We harmonize the datasets and pair samples at shared six-hourly timestamps. This pairing does not guarantee synchronized weather states between MPI-ESM and ERA5, but provides a common reference for cross-source comparison and evaluation.

Our setup and preprocessing follow ClimaX’s MPI-ESM-to-ERA5 setting [1], as clarified in the updated Appendix A.1. We have not evaluated deseasonalized coarse-scale input–target anomaly correlations, a limitation to address in future work.

Q2. Effectiveness of Sequential Dual-LoRA

We compare sequential single-LoRA training (Single), jointly trained Dual-LoRA (Joint), and sequential Dual-LoRA (ClimLoRA) under matched parameter counts and training budgets. ClimLoRA achieves the lowest RMSE and LFD across all four variables. For V10, T2M, and MSL, ClimLoRA reduces RMSE by 1.22%, 1.73%, and 1.62% relative to Single, and by 1.03%, 1.17%, and 3.34% relative to Joint, respectively.

Single modifies learned reconstruction parameters, while Joint may incur objective interference. These results suggest that sequential training with parameter freezing helps preserve reconstruction and reduce interference.

MethodU10 RMSE ↓U10 LFD ↓V10 RMSE ↓V10 LFD ↓T2M RMSE ↓T2M LFD ↓MSL RMSE ↓MSL LFD ↓
Single0.269211.19880.263111.15800.386611.927416.22019.3581
Joint0.268111.19730.262611.15390.384411.915116.50919.3857
ClimLoRA0.267711.19550.259911.13900.379911.891515.95719.3229

Experimental Settings. Single and ClimLoRA train the Base, fine-scale reconstruction, and physical optimization stages for 30, 10, and 10 epochs at learning rates of 10−410^{-4}, 10−410^{-4}, and 10−510^{-5}, respectively. Joint retains the first two stages and jointly optimizes normalized reconstruction and physical objectives (1:1) for the final 10 epochs. High-frequency and spherical harmonic spectral losses use weights of 0.6:0.4; wind-field structural and scalar-gradient losses use 1:0.4. All configurations add 7,272,448 adapter parameters: one rank-128 shared adapter for Single or two rank-64 branches for ClimLoRA and Joint. Methods share data splits, optimizer settings, and sample order for each seed; independent runs vary initialization and sampling order across seeds.

Q3. Physical and Spectral Evaluation

Physical Diagnostic Errors

As shown below, ClimLoRA outperforms Base, FSR-LoRA, and GeoFAR across all six diagnostics. Horizontal temperature advection error, ATA_T, is excluded from training and assesses wind–temperature coupling. ClimLoRA reduces it by 11.93% versus Base and 7.02% versus FSR-LoRA, suggesting improvements beyond the training constraints.

MethodHorizontal KE (m²/s²) ↓Divergence (10⁻⁷ s⁻¹) ↓Relative vorticity (10⁻⁷ s⁻¹) ↓Temperature gradient (10⁻⁶ K/m) ↓Pressure gradient (10⁻⁵ Pa/m) ↓Temperature advection (K/day) ↓
Base0.15362.97443.04772.530910.04252.0568
FSR-LoRA0.14942.89923.02562.34999.39611.9483
GeoFAR0.19303.29613.31802.271110.20711.9254
ClimLoRA0.13502.80692.82822.11819.01981.8115

Degree-Resolved Spectral Errors

Figure 7 in the revised main text compares MSL spectral errors for the same 128 samples from 2020 (see Appendix A.5.4). Over spherical harmonic degrees 20–400, FSR-LoRA improves on Base, and ClimLoRA further reduces errors. The curves converge at higher degrees, indicating scale-dependent gains from fine-scale reconstruction and physical correction.

Effects of Inference Coefficients

Figure 5 in the revised main text shows physical diagnostic errors with λfine=1\lambda_{\mathrm{fine}}=1 (see Appendix A.5.3). As λphy\lambda_{\mathrm{phy}} increases, all six errors decrease then rise, reaching minima at 1.2 for kinetic energy, divergence, and vorticity; 1.0 for temperature gradient and advection; and 0.8 for pressure gradient. The default 1.0 is near-optimal across all diagnostics.

References

[1] ClimaX: A Foundation Model for Weather and Climate, ICML, 2023.

Our Q4 response follows.

Q4. Implementation Details

LP/HP. LP uses channel-wise 9×99\times9 Gaussian convolution (σ=2\sigma=2, stride 1, four-cell replication padding); HP(z)=z−LP(z)\mathrm{HP}(z)=z-\mathrm{LP}(z). LP supervises Base residuals; HP filters FSR targets and output corrections, with identical output filtering during training and inference. Base outputs receive no additional inference filtering, and physical corrections bypass HP.

Spectral weighting. For ℓ=0,…,63\ell=0,\ldots,63, log-amplitude differences are weighted by wℓ=1+2(ℓ/63)2w_\ell=1+2(\ell/63)^2 and normalized by the number of valid spherical harmonic coefficients.

Physical normalization. Fixed scales use mean training-sample losses from the frozen FSR model; normalized spatial wind-field structural and scalar-gradient losses are weighted 1:0.41:0.4.

Derivative boundaries. Spherical derivatives use periodic central differences in longitude, central differences at interior latitudes, and second-order one-sided differences at latitude boundaries.

Residual combination. In Equation (13), Section 3.5 of the main text, effective residuals incorporate the corresponding filters; the two coefficients control fine-scale and physical corrections, respectively.

These details have been added to Appendix A.4.4.

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

Related papers

Discussion (0)

Sign in to comment.