A Differentiable Framework for Global Circulation Model Precipitation Bias Correction

Chaopeng Shen; Kamlesh Sawadekar; Kathryn Lawson; Peijun Li; Seth McGinnis

arxiv: 2604.23045 · v3 · submitted 2026-04-24 · 💻 cs.LG

A Differentiable Framework for Global Circulation Model Precipitation Bias Correction

Kamlesh Sawadekar , Seth McGinnis , Peijun Li , Kathryn Lawson , Chaopeng Shen This is my paper

Pith reviewed 2026-05-08 12:13 UTC · model grok-4.3

classification 💻 cs.LG

keywords precipitation bias correctionGCMdifferentiable learningextreme eventsclimate modelingparametric adjustmentCMIP6downscaling

0 comments

The pith

A differentiable framework learns parametric adjustments to correct GCM precipitation biases and reproduce extreme distributions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a differentiable bias-adjustment framework that learns a spatiotemporally adaptive parametric procedure for correcting precipitation outputs from General Circulation Models using historical CMIP6 data and Livneh observations. Traditional statistical methods have limited ability to capture patterns from large datasets or generalize, while direct machine learning can lack interpretability. By adjusting parameters rather than predicting corrected values directly, the approach targets precipitation's non-Gaussian, intermittent, and heavy-tailed properties. Results show effective correction of extreme magnitudes and distributions, accurate reproduction of quantiles across U.S. cities, and spatial patterns matching the LOCA2 product, along with partial trend preservation and bias reduction in unseen areas.

Core claim

The central claim is that a differentiable framework can learn a spatiotemporally adaptive parametric bias-adjustment procedure between historical CMIP6 GCM outputs and Livneh observations, which corrects the magnitude and distribution of extreme precipitation with strong upper-tail performance, reproduces quantile distributions across diverse U.S. cities, yields spatial patterns comparable to LOCA2, partially preserves future trends, and attenuates marginal biases in unseen regions.

What carries the argument

dCLIMBA, the differentiable bias-adjustment framework that learns a spatiotemporally adaptive parametric bias-adjustment procedure between GCM outputs and gridded observations rather than generating corrected precipitation directly.

Load-bearing premise

The learned parametric bias-adjustment procedure will generalize across GCMs and locations while preserving future trends without introducing artifacts.

What would settle it

Apply the trained procedure to a held-out GCM or future period and compare the corrected extreme quantiles, trend preservation, and absence of new artifacts against independent observations or reanalysis.

read the original abstract

Systematic biases in General Circulation Model (GCM) outputs limit their direct applicability in regional planning, making bias correction a technically demanding but necessary step for both short-term and long-term impact assessment. Correcting precipitation is particularly challenging due to its non-Gaussian distribution, intermittent nature, and heavy-tailed extremes. However, traditional statistical bias-correction methods have limited ability to learn systematic patterns from large datasets or generalize to new locations. While machine learning (ML) provides greater flexibility, it can produce unpredictable and difficult-to-interpret results, limiting generalization across GCMs and locations. In this study, we propose a differentiable bias-adjustment framework called dCLIMBA, that learns a spatiotemporally adaptive parametric bias-adjustment procedure, rather than corrected precipitation directly, between historical CMIP6 model outputs and a gridded observation-based dataset, Livneh. Results demonstrate that the proposed method corrects the magnitude and distribution of extreme precipitation with particularly strong performance in the upper tail. The quantile distribution of precipitation was well reproduced across diverse U.S. cities, and spatial patterns were comparable to those from the widely used LOCA2 statistical downscaling product. In addition, the framework showed partial future trend preservation and promising attenuation of marginal biases in unseen regions. This work presents a modular and efficient bias-correction approach. The differentiable approach provides an easy-to-use option for connecting atmospheric-model outputs to on-the-ground impacts.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

dCLIMBA introduces a differentiable parametric bias correction for GCM precipitation that handles historical extremes well but leaves the stationarity assumption for future application largely untested.

read the letter

The main thing to know is that this paper proposes dCLIMBA, a framework that learns a spatiotemporally adaptive parametric adjustment between CMIP6 outputs and Livneh observations instead of predicting corrected precipitation directly or relying on fixed statistical corrections. It reports solid performance on historical data for extreme precipitation magnitudes and distributions, especially in the upper tail, with quantile matches across U.S. cities and spatial patterns close to LOCA2. Partial future trend preservation and some bias reduction in unseen regions are also claimed. The differentiable setup is presented as modular and efficient for connecting models to impacts. What is new is the specific combination of differentiable optimization with a parametric form that adapts in space and time, which sits between rigid traditional methods and opaque ML predictors. The historical results on extremes and quantiles look like a genuine step forward for that period. The soft spot is the leap to future use. The method is trained on 1980-2014 data and then applied forward, but the abstract only notes partial trend preservation without quantitative trend-error numbers, cross-GCM hold-outs, or checks that the parametric shape remains suitable when underlying biases change under forcing. If the learned parameters mostly encode historical location-specific fixes rather than a transferable mapping, future quantiles could still get distorted even if the past fit is good. This is a real concern for any bias-correction claim that targets long-term projections. The paper is aimed at climate impact researchers and regional planners who need better precipitation corrections, particularly for extremes. Readers working on ML for climate post-processing would find the parametric differentiable angle practical. It deserves a serious referee because the core idea is fresh, the historical evidence is encouraging, and the problem it targets matters. Reviewers could push for clearer validation on the future generalization and the exact parametric choices. I would send it to peer review with those points in mind.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces dCLIMBA, a differentiable framework that learns a spatiotemporally adaptive parametric bias-adjustment procedure to correct precipitation outputs from CMIP6 GCMs against the Livneh gridded observational dataset. It claims effective correction of extreme precipitation magnitude and distribution (especially upper tail), accurate reproduction of quantile distributions across U.S. cities, spatial patterns comparable to the LOCA2 downscaling product, partial preservation of future trends, and attenuation of marginal biases in unseen regions.

Significance. If the central claims hold under rigorous validation, the work offers a useful middle path between rigid traditional statistical bias-correction methods and black-box ML approaches, with the differentiability providing modularity and potential for end-to-end optimization in climate-impact pipelines. The emphasis on parametric learning rather than direct output correction is a conceptual strength that could improve interpretability and generalization.

major comments (3)

Abstract: The claim of 'partial future trend preservation' is stated at a high level without quantitative trend-error metrics, cross-GCM hold-out tests, or explicit checks that the learned parametric form remains valid when GCM bias structures shift under forcing; this assumption is load-bearing for applicability to long-term projections.
Abstract/Results: No equations, loss functions, training details, error metrics (e.g., RMSE, quantile scores), or ablation studies are supplied to support the reported performance on extremes, quantiles, and generalization, preventing verification of the central empirical claims.
Methods (parametric form): The exact functional shape of the spatiotemporally adaptive parametric adjustment is not specified, so it is impossible to evaluate whether the learned mapping encodes transferable physical corrections or merely historical location-specific offsets that could distort future quantiles.

minor comments (2)

The acronym dCLIMBA is introduced without expansion on first use.
Figure captions and axis labels for quantile and spatial plots should include explicit units and reference periods for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful and constructive report. The comments highlight important areas where additional clarity and evidence will strengthen the manuscript. We address each major comment below and commit to revisions that directly respond to the concerns while preserving the core contributions of the work.

read point-by-point responses

Referee: Abstract: The claim of 'partial future trend preservation' is stated at a high level without quantitative trend-error metrics, cross-GCM hold-out tests, or explicit checks that the learned parametric form remains valid when GCM bias structures shift under forcing; this assumption is load-bearing for applicability to long-term projections.

Authors: We agree the abstract statement is high-level. The results section demonstrates preservation via side-by-side trend maps for 2070-2100 under SSP5-8.5, showing that corrected fields retain the sign and broad spatial structure of raw GCM trends. To address the concern, we will add quantitative metrics (e.g., MAE on linear trend slopes and correlation of trend fields) and a cross-GCM hold-out experiment in the revised results. We will also include a brief sensitivity discussion on how the parametric form may behave under changing bias structures, noting that this remains an assumption requiring further validation in future work. revision: partial
Referee: Abstract/Results: No equations, loss functions, training details, error metrics (e.g., RMSE, quantile scores), or ablation studies are supplied to support the reported performance on extremes, quantiles, and generalization, preventing verification of the central empirical claims.

Authors: We acknowledge that the current manuscript version presents performance claims without sufficient supporting detail in the main text. We will revise the Methods and Results sections to explicitly include: (1) the differentiable loss function (a quantile-matching objective with tail emphasis), (2) training hyperparameters and optimization procedure, (3) concrete error metrics (RMSE, quantile scores, and extreme-value statistics), and (4) ablation results on the spatiotemporal adaptation components. These additions will be placed in the main body or a new summary table to enable direct verification. revision: yes
Referee: Methods (parametric form): The exact functional shape of the spatiotemporally adaptive parametric adjustment is not specified, so it is impossible to evaluate whether the learned mapping encodes transferable physical corrections or merely historical location-specific offsets that could distort future quantiles.

Authors: We agree that the precise functional form must be stated explicitly. The adjustment takes the form P_corrected(x,y,t) = a(x,y,t) * P_GCM(x,y,t) + b(x,y,t), where the spatially and temporally varying parameters a and b are produced by a small neural network conditioned on location, time, and GCM precipitation. This structure is chosen to permit both scaling and offset corrections that can adapt locally while remaining parametric. We will add this equation and a short discussion of its generalization properties (supported by the unseen-region results) to the Methods section. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; derivation relies on external data fitting

full rationale

The paper introduces dCLIMBA as a differentiable framework that learns a spatiotemporally adaptive parametric bias-adjustment procedure directly from historical CMIP6 outputs paired with the external Livneh observation dataset. Performance claims (quantile reproduction, upper-tail correction, spatial comparability to LOCA2, partial future trend preservation) are presented as empirical outcomes of this training process rather than any derivation that reduces outputs to inputs by construction. No equations, self-citations, uniqueness theorems, or ansatzes are described that would make the central results tautological or force predictions via fitted parameters renamed as independent forecasts. The approach is self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; all details are at the level of high-level method description.

pith-pipeline@v0.9.0 · 5562 in / 1100 out tokens · 42726 ms · 2026-05-08T12:13:42.443090+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

Across six CMIP6 models over CONUS, δ CLIMBA reduces marginal and extreme-precipitation biases and largely preserves multiscale spatial structure

work page
[2]

δ CLIMBA partially preserves future precipitation trends and shows some spatial generalization to unseen regions

work page
[3]

black boxes

Introduction General Circulation Models (GCMs) are essential tools for projecting future atmospheric regime change, but their outputs are subject to systematic biases that limit their direct use in regional impact assessments. These models simulate interactions between the atmosphere, hydrosphere, cryosphere, and biosphere, and are used by policymakers an...

work page 2014
[4]

3.1 Quantile Comparison Quantile comparisons across five climatically-diverse U.S

Results δ CLIMBA was evaluated using a series of experiments to test its generalizability across different GCMs at different locations, including its ability to adjust biases through precipitation indices, conserve spatial structures, and capture long-term trends. 3.1 Quantile Comparison Quantile comparisons across five climatically-diverse U.S. cities de...

work page 2001
[5]

Should we apply bias correction to global and regional climate model data?

Conclusions Bias correction or bias adjustment is a domain that cannot be addressed or optimized by a single metric. Every GCM output carries different kinds of biases, and therefore an exhaustive evaluation of not only marginal biases but their temporal, spatial, and scale-dependent inherent characteristics is important. With its composite quantile loss,...

work page doi:10.5194/bg-20-2671-2023 2023

[1] [1]

Across six CMIP6 models over CONUS, δ CLIMBA reduces marginal and extreme-precipitation biases and largely preserves multiscale spatial structure

work page

[2] [2]

δ CLIMBA partially preserves future precipitation trends and shows some spatial generalization to unseen regions

work page

[3] [3]

black boxes

Introduction General Circulation Models (GCMs) are essential tools for projecting future atmospheric regime change, but their outputs are subject to systematic biases that limit their direct use in regional impact assessments. These models simulate interactions between the atmosphere, hydrosphere, cryosphere, and biosphere, and are used by policymakers an...

work page 2014

[4] [4]

3.1 Quantile Comparison Quantile comparisons across five climatically-diverse U.S

Results δ CLIMBA was evaluated using a series of experiments to test its generalizability across different GCMs at different locations, including its ability to adjust biases through precipitation indices, conserve spatial structures, and capture long-term trends. 3.1 Quantile Comparison Quantile comparisons across five climatically-diverse U.S. cities de...

work page 2001

[5] [5]

Should we apply bias correction to global and regional climate model data?

Conclusions Bias correction or bias adjustment is a domain that cannot be addressed or optimized by a single metric. Every GCM output carries different kinds of biases, and therefore an exhaustive evaluation of not only marginal biases but their temporal, spatial, and scale-dependent inherent characteristics is important. With its composite quantile loss,...

work page doi:10.5194/bg-20-2671-2023 2023