Generalizable Neural Downscaling of Earth System Model Wind Fields via Continuous Dynamics Modeling

Chenxi Yu; Jianan Wei; Hanlin Kong; Hao Sun; Bian He; Wenguan Wang

physics.ao-phSep 26, 2026

Generalizable Neural Downscaling of Earth System Model Wind Fields via Continuous Dynamics Modeling

Authors: Chenxi Yu, Jianan Wei, Hanlin Kong, Hao Sun, Bian He, Wenguan Wang

Organizations: State Key Lab of Brain-Machine Intelligence, Zhejiang University · Gaoling School of Artificial Intelligence, Renmin University of China · State Key Laboratory of Earth System Numerical Modeling and Application, Institute of Atmospheric Physics, Chinese Academy of Sciences

Abstract

Accurate high-resolution wind field simulations are critical for resolving fine-scale atmospheric dynamics, yet the simulation of wind fields in Earth System Models (ESMs) remains limited by coarse spatial resolution and systematic biases. To address this, data-driven down scaling techniques have been widely used to enhance coarse-resolution ESM outputs. However, existing methods are typically tied to fixed discretizations, limiting generalization across models with different native resolutions. Here we formulate global near-surface wind downscaling as an operator-learning problem on continuous atmospheric state fields and develop a downscaling neural operator that maps coarse-scale fields to fine-scale counterparts across heterogeneous discretizations. The operator learning-based neural downscaling framework outperforms dominant baselines, recovers fine-scale physical structures, and generalizes to previously unseen ESMs and future climate scenarios without retraining, while preserving long-term wind projection trends. These findings establish a generalizable paradigm for high-resolution climate downscaling across diverse simulation outputs and future scenarios.

Figures & tables

Figure 1: Overview of our work. (i) Climate downscaling is formulated as a mapping between continuous state fields $\bm{a}$ and $\bm{b}$ defined over the domain $\Omega$ , such that $\bm{b}=G(\bm{a})$ , where $G$ denotes the theoretical downscaling operator. (ii) A neural operator $G_{\boldsymbol{\theta}}$ is trained to approximate the underlying downscaling operator $G$ . The HR gridded fields are first progressively coarsened to construct a hierarchy of LR gridded fields, which are used to represent the LR state field $\bm{a}$ during training; the neural operator then learns the LR-to-HR function mapping such that $\bm{b}(\bm{x})\approx(G_{\boldsymbol{\theta}}(\bm{a}))(\bm{x})$ , and the predicted HR state field is discretized on the target HR grids to compute the loss against the HR reference values for optimizing $G_{\boldsymbol{\theta}}$ . (iii) When applying the pretrained neural operator $G_{\boldsymbol{\theta}^{*}}$ to bias-corrected LR outputs $\bm{a(x_{\textbf{LR}})}$ from ESMs, which are interpreted as representation of the underlying LR state field, it directly generates HR gridded fields $\bm{b(x_{\textbf{HR}})}$ discretized on the target HR grids.

Methods	Scale factor	Input grid	MAE		RMSE		95 ${}^{\text{th}}$ PE		5 ${}^{\text{th}}$ PE
Methods	Scale factor	Input grid	$UV$	Speed	$UV$	Speed	$UV$	Speed	$UV$
SD (Interp.)	$\times$ 3	$0.75^{\circ}$	0.142	0.161	0.313	0.367	0.137	0.190	0.140
SD (Clim.)			0.113	0.126	0.244	0.290	0.107	0.137	0.111
Ours			0.046	0.048	0.079	0.083	0.028	0.032	0.028
SD (Interp.)	$\times$ 4	$1^{\circ}$	0.210	0.237	0.411	0.475	0.200	0.279	0.207
SD (Clim.)			0.169	0.190	0.322	0.377	0.158	0.204	0.165

Table 1: Continuous state field mapping performance comparison across the neural downscaling framework and SD baselines.

Methods	MAE (pool)	PCC		ME				95 ${}^{\text{th}}$ PE				5 ${}^{\text{th}}$ PE
Methods	$UV$	$UV$	Speed	$UV$	%	Speed	%	$UV$	%	Speed	%	$UV$	%
BCSD (Interp.)	0.114	0.993	0.985	0.249	55.50	0.306	55.24	0.432	54.98	0.523	56.74	0.422	55.10
BCSD (Clim.)	0.099	0.997	0.992	0.192	65.77	0.254	62.87	0.396	58.70	0.453	62.51	0.386	58.96
Ours	0.048	0.996	0.996	0.195	65.15	0.184	73.08	0.322	66.47	0.309	74.42	0.296	68.48

Table 2: Statistical comparison in CMIP6 [ 9 ] modeling of ERA5 performance across the neural downscaling framework and BCSD baselines.

Figure 2: Qualitative comparison of daily near-surface wind fields. a , Daily near-surface wind field from raw GFDL-ESM4 [ 47 ] simulation. b , Bias-corrected (BC) wind field on the native resolution. c , Near-surface wind field from the target ERA5 reanalysis. d,e,f , Downscaled fields from our neural operator ( d ), BCSD (Clim.) ( e ), and BCSD (Interp.) ( f ), respectively. The local regions highlighted by red boxes and labeled A and B in these panels are enlarged and shown in g and h , respectively.

Figure 3: RAPSD comparison of neural downscaling framework and BCSD baselines. Radial averaged power spectral density (RAPSD) of the zonal wind component $U$ , meridional wind component $V$ , and wind speed $S$ from the downscaled fields. The lower panel shows the enlarged views of the small-scale regime (spatial scales $<100$ km). The shaded region denotes $\pm 1$ standard deviation over all time steps during the evaluation period.

Figure 4: Bias correction capability comparison of neural downscaling framework and BCSD baselines. a , MAE distribution across speed percentiles for global near-surface wind components and wind speed relative to ERA5, with the shaded region representing the interquartile range of errors (25 ${}^{\text{th}}$ –75 ${}^{\text{th}}$ percentile). b , The relative frequency of near-surface wind components and wind speed with respect to ERA5. c , The global distribution of mean EKE bias.

Figure 5: Predictive uncertainty of the neural downscaling framework for daily near-surface wind fields from the GFDL-ESM4 simulations. a,b , The ensemble mean ( a ) and standard deviation ( b ) of the downscaled fields from our framework with stochastic forward passes for the near-surface wind speed, respectively.

Figure 6: Future climate scenario projections downscaled by the neural downscaling framework. a,b , Spatial patterns of future global near-surface wind speed trends produced by our framework ( a ) and the original GFDL-ESM4 [ 47 ] simulations ( b ) under the SSP119, SSP245, and SSP585 scenarios. c , Latitudinally averaged wind speed trends of our framework and the original GFDL-ESM4 simulations over the global domains for the SSP119, SSP245, and SSP585 scenarios, from left to right.

Figure 7: Overview architecture of the downscaling neural operator. a , The overall architecture of the downscaling neural operator. b , The detailed architecture of the Fourier integral operator $F_{\boldsymbol{\theta}_{3}}$ , i.e. , the U-shaped FNO, which consists of an encoder-decoder architecture with skip-connections in which convolutional layers are replaced by FNO [ 66 ] blocks. c , The detailed architecture of single FNO block.

Figure S.1: Global MAE distribution comparison for state mapping across the neural downscaling framework and SD baselines. From left to right, MAE of our framework, SD (Clim.), and SD (Interp.), respectively. From top to bottom, MAE of the downscaled fields from coarsened ERA5 reanalysis data with spatial resolution of 1.5 ${}^{\circ}\times$ 1.5 ∘ , 1.25 ${}^{\circ}\times$ 1.25 ∘ , 1 ${}^{\circ}\times$ 1 ∘ , and 0.75 ${}^{\circ}\times$ 0.75 ∘ to HR ERA5 reanalysis with spatial resolution of 0.25 ∘ .

Figure S.2: RAPSD comparison for state mapping across the neural downscaling framework and SD baselines. The RAPSD of the zonal wind component $U$ ( a ), meridional wind component $V$ ( b ), and wind speed $S$ ( c ) from coarsened ERA5 reanalysis data with spatial resolution of 1.5 ${}^{\circ}\times$ 1.5 ∘ , 1.25 ${}^{\circ}\times$ 1.25 ∘ , 1 ${}^{\circ}\times$ 1 ∘ , and 0.75 ${}^{\circ}\times$ 0.75 ∘ to HR ERA5 reanalysis with spatial resolution of 0.25 ∘ for our framework and SD baselines. The lower panel shows the enlarged views of the small-scale regime (spatial scales $<100$ km). The shaded region denotes $\pm 1$ standard deviation over all time steps during the evaluation period.

Figure S.3: Qualitative comparison of raw ESM simulations and downscaled fields for the same date. a,b , The raw coarse-scale ESM simulations for the MPI-ESM1-2-HR [ 48 ] ( a ), and MRI-ESM2-0 [ 49 ] ( b ) simulations, respectively. c,d , The downscaled fields from our framework for the MPI-ESM1-2-HR ( c ), and MRI-ESM2-0 ( d ) simulations.

Figure S.4: RAPSD comparison of neural downscaling framework and BCSD baselines across different ESMs. a , RAPSD of the zonal wind component $U$ , meridional wind component $V$ , and wind speed $S$ for the MPI-ESM1-2-HR [ 48 ] simulation. b , RAPSD of $U$ , $V$ , and $S$ for the MRI-ESM2-0 [ 49 ] simulation. The lower panel shows the enlarged views of the small-scale regime (spatial scales $<100$ km). The shaded region denotes $\pm 1$ standard deviation over all time steps during the evaluation period.

Figure S.5: Global distribution of the 95 ${}^{\text{th}}$ percentile wind speed error relative to ERA5 for the neural downscaling framework, BCSD baselines, and GFDL-ESM4 [ 47 ] simulations. The global distribution of the 95 ${}^{\text{th}}$ percentile wind speed error relative to ERA5 for our framework ( a ), BCSD (Clim.) ( b ), BCSD (Interp.) ( c ), and GFDL-ESM4 simulations ( d ), respectively.

Figure S.6: Relative frequency distributions of near-surface wind fields downscaled by the neural downscaling framework and BCSD baselines across different ESMs. a , The relative frequency distribution of the near-surface wind fields for MPI-ESM1-2-HR [ 48 ] simulations and their downscaled fields with respect to ERA5. b , The relative frequency distribution of the near-surface wind fields for MRI-ESM2-0 [ 49 ] simulations and their downscaled fields.

Figure S.7: Double-ITCZ bias reduction performance of bias-correction preprocessing (BC) and neural downscaling framework for different ESMs. a , The divergence of the near-surface wind field for the ESM simulations (GFDL-ESM4 [ 47 ] , MPI-ESM1-2-HR [ 48 ] , and MRI-ESM2-0 [ 49 ] ). b , The divergence of the wind field for the bias-corrected ESM simulations. c , The divergence of the wind field for the downscaled fields from our framework. d , The divergence of the wind field for the ground truth ERA5 reanalysis data.

ESMs	Methods	MAE (pool)	PCC		ME				95 ${}^{\text{th}}$ PE				5 ${}^{\text{th}}$ PE
ESMs	Methods	$UV$	$UV$	Speed	$UV$	%	Speed	%	$UV$	%	Speed	%	$UV$	%
MPI- ESM1-2- HR	BCSD (Interp.)	0.115	0.993	0.986	0.250	60.13	0.289	60.82	0.421	58.77	0.456	62.47	0.401	56.94
	BCSD (Clim.)	0.097	0.996	0.992	0.198	68.42	0.241	67.27	0.388	61.91	0.396	67.42	0.370	60.36
	Ours	0.041	0.996	0.996	0.202	67.79	0.189	74.34	0.327	67.91	0.294	75.75	0.299	67.96
MRI- ESM2-0	BCSD (Interp.)	0.147	0.991	0.983	0.275	62.00	0.309	69.93	0.440	68.75	0.509	72.12	0.449	65.18
	BCSD (Clim.)	0.126	0.995	0.991	0.214	70.45	0.251	75.55	0.403	71.42	0.435	76.21	0.411	68.14

Table S.1: Statistical comparison in CMIP6 (MPI-ESM1-2-HR [ 48 ] and MRI-ESM2-0 [ 49 ] ) modeling of ERA5 performance across the neural downscaling framework and BCSD baselines.

ESMs	Methods	MAE (MKE)	%	MAE (EKE)	%
GFDL-ESM4	BCSD (Interp.)	1.081	57.05	1.289	55.33
	BCSD (Clim.)	0.855	66.03	1.243	56.92
	Ours	0.856	65.98	0.816	71.73
MPI-ESM1-2-HR	BCSD (Interp.)	1.090	62.85	1.184	58.49
	BCSD (Clim.)	0.885	69.84	1.144	59.89
	Ours	0.894	69.55	0.833	70.81

Table S.2: Statistical comparison in specific kinetic energy decomposition (MKE and EKE) for CMIP6 (GFDL-ESM4 [ 47 ] , MPI-ESM1-2-HR [ 48 ] , and MRI-ESM2-0 [ 49 ] ) modeling of ERA5 across the neural downscaling framework and BCSD baselines.

Figure S.8: Comparison of temporal consistency across the neural downscaling framework and BCSD baselines. a , Temporal autocorrelation and biases with respect to ERA5 of the zonal wind component $U$ for the our framework and the BCSD baselines. b , Temporal autocorrelation and biases of the meridional wind component $V$ .

Figure S.9: Uncertainty quantification of neural downscaling framework for the downscaled fields from GFDL-ESM4 [ 47 ] simulations. a,b , The ensemble mean ( a ) and standard deviation ( b ) of the downscaled fields from our framework with stochastic forward passes for the wind speed. c , The CRPS for the ensemble of downscaled fields from our framework with stochastic forward passes for the wind speed. d , The CRPSS for the ensemble of downscaled fields from our framework with stochastic forward passes relative to BCSD (Clim.) baseline ensemble.

Figure S.10: Future climate scenario projection trends under the SSP119, SSP245, and SSP585 scenarios over the global, land and ocean domains. a , Latitudinally averaged wind speed trends of the GFDL-ESM4 simulations without bias correction over the global, land, and ocean domains. b , Latitudinally averaged wind speed trends of the downscaled fields obtained with our framework over the global, land, and ocean domains for three different future scenarios.

ESMs	Methods	MAE (pool)	PCC		ME				95 ${}^{\text{th}}$ PE				5 ${}^{\text{th}}$ PE
ESMs	Methods	$UV$	$UV$	Speed	$UV$	%	Speed	%	$UV$	%	Speed	%	$UV$	%
GFDL- ESM4	BCSD (Interp.)	0.114	0.993	0.987	0.254	54.59	0.300	55.57	0.411	57.37	0.510	57.68	0.427	54.87
	BCSD (Clim.)	0.099	0.996	0.994	0.198	64.54	0.246	63.47	0.374	61.26	0.441	63.35	0.391	58.72
	Ours	0.048	0.996	0.997	0.202	63.96	0.180	73.28	0.315	67.39	0.332	72.46	0.318	66.39

Table S.3: Statistical comparison in SSP245 future climate scenario projections across the neural downscaling framework and BCSD baselines from GFDL-ESM4 [ 47 ] simulations.

ESMs	Methods	MAE (MKE)	%	MAE (EKE)	%
GFDL-ESM4	BCSD (Interp.)	1.077	56.23	1.239	58.57
	BCSD (Clim.)	0.848	65.56	1.191	60.17
	Ours	0.846	65.63	0.785	73.74

Table S.4: Statistical comparison in specific kinetic energy decomposition (MKE and EKE) for SSP245 future climate scenario projections across the neural downscaling framework and BCSD baselines from GFDL-ESM4 [ 47 ] simulations.

Model	$s$	$\times$ 3		$\times$ 4		$\times$ 5		$\times$ 6
Model	$s$	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
SD (Interp.)	-	0.1602	0.3694	0.2359	0.4775	0.2937	0.5618	0.3586	0.6492
SD (Clim.)	-	0.1256	0.2927	0.1883	0.3794	0.2372	0.4519	0.2929	0.5277
Ours	3	0.0376	0.0663	0.1227	0.2442	0.1847	0.3728	0.2480	0.5665
Ours	6	0.1110	0.3077	0.1477	0.3804	0.1819	0.3758	0.1131	0.1794
Ours	$\mathcal{U}(3,6)$	0.0484	0.0840	0.0664	0.1115	0.0936	0.1505	0.1167	0.1859

Table S.5: Ablation study on the downscaling neural operator with different training scale factor strategies.

Model	$\lambda_{2}$	$\times$ 3		$\times$ 4		$\times$ 5		$\times$ 6
Model	$\lambda_{2}$	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
SD (Interp.)	-	0.1602	0.3694	0.2359	0.4775	0.2937	0.5618	0.3586	0.6492
SD (Clim.) [ 42 ]	-	0.1256	0.2927	0.1883	0.3794	0.2372	0.4519	0.2929	0.5277
CNN [ 67 ]	0	0.0639	0.1289	0.0941	0.1795	0.1301	0.2395	0.1695	0.3050
SRNO [ 68 ]	0	0.0583	0.1155	0.0837	0.1530	0.1150	0.2019	0.1521	0.2631
DFNO [ 40 ]	0	0.0545	0.0946	0.0699	0.1173	0.0968	0.1556	0.1192	0.1899

Table S.6: State mapping capability of our downscaling neural operator with different model architectures and gradient loss weights.

Figure S.11: Comparison of global MAE distribution for state mapping with different training scale factor strategies. From left to right, MAE of our framework with continuous sampling range $s\sim\mathcal{U}(3,6)$ , and with discrete sampling range (3 and 6), respectively. From top to bottom, MAE of the downscaled fields from coarsened ERA5 reanalysis data with spatial resolution of 1.5 ${}^{\circ}\times$ 1.5 ∘ , 1.25 ${}^{\circ}\times$ 1.25 ∘ , 1 ${}^{\circ}\times$ 1 ∘ , and 0.75 ${}^{\circ}\times$ 0.75 ∘ to HR ERA5 reanalysis with spatial resolution of 0.25 ∘ .

Climate Downscaling

Climate downscaling translates coarse climate simulations and observations into finer regional detail for local impact and risk assessment.

See all 27 papers

Oct 17, 2026

Spherical Physics-informed Neural Operator with Multi-scale Coupling for Meteorological Downscaling

Meteorological downscaling is crucial for high-resolution regional climate forecasting and disaster early warning. While neural operators have emerged as a promising paradigm for modeling complex spatiotemporal mappings, existing frameworks often struggle with spherical manifold geometric distortions, inherent atmospheric multi-scale coupling mismatches, and lack of explicit atmospheric laws. We propose the Spherical Physics-informed Neural Operator, which utilizes a Spherical Laplacian Decomposition to partition atmospheric fields into hierarchical frequency components, maintaining exact point-wise correspondence across scales. To evaluate these representations at arbitrary locations, we introduce a localized spherical integral operator that approximates continuous kernel transforms via geometry-aware attention. Dynamical consistency is further enforced by embedding differentiable constraints into the learning process. Extensive experiments demonstrate that our framework attains superior accuracy and zero-shot generalization across various meteorological variables and unseen queries, representing a robust and interpretable solution for global-to-regional meteorological downscaling.

Yiqiang Ye, Yichi Wang, Haodi Zhang +5

College of Computer and Software Engineering, Shenzhen University, Shenzhen, China · School of Computer Science, Wuhan University, Wuhan, China

Climate Downscaling Physics-Informed Neural Operators Neural Operators Multi-Scale Feature Fusion

Ask PDF

Feb 13, 2026

High-Resolution Climate Projections Using Diffusion-Based Downscaling of a Lightweight Climate Emulator

The proliferation of data-driven models in weather and climate sciences has marked a significant paradigm shift, with advanced models demonstrating exceptional skill in medium-range forecasting. However, these models are often limited by long-term instabilities, climatological drift, and substantial computational costs during training and inference, restricting their broader application for climate studies. Addressing these limitations, Guan et al. (2024) introduced LUCIE, a lightweight, physically consistent climate emulator utilizing a Spherical Fourier Neural Operator (SFNO) architecture. This model is able to reproduce accurate long-term statistics including climatological mean and seasonal variability. However, LUCIE's native resolution (~300 km) is inadequate for detailed regional impact assessments. To overcome this limitation, we introduce a deep learning-based downscaling framework, leveraging probabilistic diffusion-based generative models with conditional and posterior sampling frameworks. These models downscale coarse LUCIE outputs to 25 km resolution. They are trained on approximately 14,000 ERA5 timesteps spanning 2000-2009 and evaluated on LUCIE predictions from 2010 to 2020. Model performance is assessed through diverse metrics, including latitude-averaged RMSE, power spectrum, probability density functions and First Empirical Orthogonal Function of the zonal wind. We observe that the proposed approach is able to preserve the coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution.

Haiwen Guan, Dibyajyoti Chakraborty, Moein Darman +3

1Information Sciences and Technology, The Pennsylvania State University, University Park, Pennsylvania · Department of Applied Mathematics, University of California, Santa Cruz, California · 3Allen Institute for Artificial Intelligence (AI2), Seattle, Washington +1

Climate Science Climate Modeling Conditional Diffusion Models Climate Downscaling Diffusion Models

Ask PDF

Mar 30, 2026

Downscaling weather forecasts from Low- to High-Resolution with Diffusion Models

We introduce a probabilistic diffusion-based method for global atmospheric downscaling implemented within the Anemoi framework. The approach transforms low-resolution ensemble forecasts into high-resolution ensembles by learning the conditional distribution of finer-scale residuals, defined as the difference between the high-resolution fields and the interpolated low-resolution inputs. The system is trained on reforecast pairs from ECMWF IFS, using coarse fields at 100 km to reconstruct fine-scale variability at 30 km resolution. The bulk of the training focuses on recovering small-scale structures, while fine-tuning in high-noise regimes enables the generation of extremes. Evaluation against the medium-range IFS ensemble target shows that the model increases probabilistic skill (FCRPS) for surface variables, reproduces target power spectra at small scales, captures physically consistent multivariate relationships such as wind-pressure coupling, and generates extreme values consistent with those of the target ensemble in tropical cyclones.

Joffrey Dumont Le Brazidec, Simon Lang, Martin Leutbecher +9

European Centre for Medium-Range Weather Forecasts (ECMWF)

Weather Forecasting Probabilistic Forecasting Diffusion Models

Ask PDF

Abstract

Climate Downscaling

Explore similar work

Spherical Physics-informed Neural Operator with Multi-scale Coupling for Meteorological Downscaling

High-Resolution Climate Projections Using Diffusion-Based Downscaling of a Lightweight Climate Emulator

Downscaling weather forecasts from Low- to High-Resolution with Diffusion Models