Benchmarking Generative Models for Weather Data Assimilation on Real Station Observations

Ruizhe Huang; Qidong Yang; Jonathan Giezendanner; Sherrie Wang

cs.LGSep 30, 2026

Benchmarking Generative Models for Weather Data Assimilation on Real Station Observations

Authors: Ruizhe Huang, Qidong Yang, Jonathan Giezendanner, Sherrie Wang

Organizations: Massachusetts Institute of Technology, Cambridge, MA, USA

Abstract

Weather reanalysis products rely on computationally intensive numerical weather predictions followed by data assimilation that corrects the forecast toward observations. Deep generative models offer a cheaper alternative that shifts much of this cost from inference to offline training. However, existing generative approaches have been evaluated on synthetic observations or under different datasets and evaluation schemes, making it unclear which design choices actually improve real-world data assimilation. We present the first controlled benchmark of generative weather data assimilation on real weather station observations. Using 11,849 NOAA MADIS stations across the contiguous United States and four weather variables, we evaluate methods while holding the dataset, observation operator, and deep learning architecture fixed. The benchmark compares the major design choices, including diffusion versus flow matching, pixel versus latent-space formulations, and multiple inference-time conditioning strategies, against a classical 3D-Var baseline. The benchmark reveals three clear conclusions. First, learned generative priors outperform the Gaussian prior of 3D-Var (35.7% vs. 33.3% RMSE reduction over ERA5) despite using no ERA5 background field at inference. Second, full-gradient guidance consistently outperforms stop-gradient and initial-noise optimization. Third, other choices provide little measurable benefit: diffusion and flow matching perform nearly identically under matched conditions, and latent-space variable mixing does not help. We further evaluate both dense and sparse station settings and find advantages from generative AI and full-gradient guidance more pronounced under sparsity. Together, these results identify which components of generative weather data assimilation improve performance on real station observations and establish a standardized benchmark for future work.

Figures & tables

Figure 1 : Conceptual illustration of generative data assimilation: starting from sparse station observations on the four target variables, a generative prior trained on ERA5 produces a full gridded analysis consistent with the observations.

	Diffusion Model	Flow Matching
Network output	Denoiser $D_{\theta}\approx\mathbf{x}_{1}$	Velocity $\mathbf{v}_{\theta}\approx\mathbf{x}_{1}-\mathbf{x}_{0}$
Training schedule	$\ln\sigma\sim\mathcal{N}(P_{\text{mean}},P_{\text{std}}^{2})$	$t\sim\mathcal{U}(0,1)$
Score function	$(s\,D_{\theta}-\mathbf{x})/(s^{2}\sigma^{2})$	$(t\,\mathbf{v}_{\theta}-\mathbf{x})/(1-t)$
Sampler	EI predictor + Langevin corrector	Adaptive ODE / Euler / Euler + corrector

Table 1 : Comparison of diffusion model and flow matching.

P_{\text{mean}}

and

P_{\text{std}}

parameterize the EDM noise-sampling distribution ( Appendix B ).

Figure 2 : Overview of the data assimilation strategies compared in this work. (a) Classical methods : 3D-Var and its latent variant minimize a background-plus-observation cost with no generative prior. (b) Diffusion-SDA ( Section 3.3.1 ): per-step guidance through the denoiser $D_{\theta}$ , optionally followed by a Langevin corrector, toward the observation-consistent region $\mathcal{H}(\mathbf{x})=\mathbf{y}^{*}$ . (c) Flow Guidance ( Section 3.3.2 ): the flow-matching analogue, backpropagating the observation loss through the terminal extrapolation $\hat{\mathbf{x}}_{1}(\mathbf{x}_{t})$ . (d) FlowDPS [ 28 ] : iterates $K$ optimization steps on the terminal estimate $\hat{\mathbf{x}}_{1}^{(K)}(\mathbf{x}_{t})$ and resamples to the next ODE step. (e) Latent Flow Guidance : Flow Guidance inside the latent space of an autoencoder $\mathcal{E}/\mathcal{D}$ ; the observation loss is evaluated in pixel space via the decoder. (f) D-Flow ( Section 3.3.3 ): no per-step guidance; an outer loop updates the initial noise $\mathbf{x}_{0}^{(0)}\to\mathbf{x}_{0}^{(k)}$ by backpropagating through the full ODE until $\mathbf{x}_{1}$ satisfies the observations. Numbered circles on panel (c) mark the four design axes ablated in Appendix C : (1) guidance schedule $\lambda(t)$ , (2) stop gradient, (3) terminal extrapolation $\hat{\mathbf{x}}_{1}$ , (4) sampler; a fifth axis (pixel vs. latent) is captured by the (c)/(e) contrast.

Method	Model class	Space	DA mechanism	$M$
3D-Var	Classical	Pixel	Variational optimization	1
Latent 3D-Var	Classical	Latent	Variational optimization	1
Diffusion-SDA	Diffusion	Pixel	Full-gradient guidance	16
Flow Guidance	Flow matching	Pixel	Full-gradient guidance	16
FlowDPS	Flow matching	Pixel	Stop-gradient guidance	16
D-Flow	Flow matching	Pixel	Initial-noise optimization	1

Table 2 : Summary of data assimilation methods. Columns: model class (classical, diffusion, or flow matching), operating space (pixel or learned latent), DA mechanism (how observations are injected), and default ensemble size

M

Figure 3 : Dataset split and regional station density under the two benchmarks. Top block (Dense, 11,849 stations): U.S. map with stations colored by partition (Train: 8,294 / Val: 1,777 / Test: 1,778) and dashed boxes marking three zoom regions, followed by California, Midwest, and New England panels in Lambert Conformal projection with identical physical extent ( $800\times 800$ km). Bottom block (Sparse, 1,000 stations): same layout for a uniformly subsampled set (Train: 700 / Val: 150 / Test: 150). Matching extents enable fair density comparison across regions and between the dense and sparse benchmarks.

Method	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Mem	Total
	$u_{10}$		$v_{10}$		$T_{2}$		$T_{d,2}$		Average		Cost
3D-Var	51.7	35.2	51.8	35.1	46.0	29.2	49.5	33.6	49.7	33.3	0.17	4
3D-Var (w/o ERA5 loss)	66.0	15.0	66.2	13.7	61.8	9.4	64.8	12.3	64.7	12.6	0.20	36
Diffusion-SDA	52.7	37.8	52.2	38.6	39.9	29.6	47.6	35.6	48.1	35.4	6.12	246
Diffusion-SDA (PO)	48.3	37.9	47.9	38.3	29.5	19.9	36.7	25.0	40.6	30.3	6.11	82
D-Flow	50.7	34.7	49.6	35.8	33.3	23.0	39.8	28.5	43.4	30.5	36.01	2563

Table 3 : Dense benchmark on the

11{,}849

-station network. RMSE improvement (%) over ERA5 at training (Tr) and held-out test (Te) stations; “Average” is the unweighted mean across the four variables. “Mem” is peak GPU memory (GB); “Total” is wall-clock time (min) for

360

snapshots on a single RTX8000. Higher improvement and lower cost are better. “PO” = predictor-only (no Langevin corrector).

Method	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Mem	Total
	$u_{10}$		$v_{10}$		$T_{2}$		$T_{d,2}$		Average		Cost
3D-Var	49.0	14.1	49.9	15.4	46.6	11.7	48.8	14.1	48.6	13.8	0.11	3
Latent 3D-Var ( $S{=}2$ , $G{=}1$ )	71.2	14.6	71.5	16.4	65.7	13.3	69.2	15.5	69.4	15.0	0.25	11
Latent 3D-Var ( $S{=}2$ , $G{=}4$ )	68.1	15.2	68.8	16.4	56.6	13.6	63.2	15.5	64.2	15.2	0.34	13
Diffusion-SDA	69.3	25.7	69.0	28.1	46.6	14.2	60.0	17.9	61.2	21.5	6.12	242
Diffusion-SDA (PO)	62.0	28.6	61.1	28.6	30.3	$-2.4$	42.6	$-2.8$	49.0	13.0	6.10	83

Table 4 : Sparse benchmark,

1{,}000

randomly subsampled stations. Columns as in Table 3 . “SG” applies the stop-gradient approximation; “Euler-128” replaces the adaptive dopri5 solver with a

128

-step fixed-step Euler solver; “PO” = predictor-only; “PC” adds a Langevin corrector.

Panel of Figure 4 — Figure 4 : Macro RMSE improvement ( $\Delta\%$ ) over ERA5 for Flow Guidance, Diffusion-SDA, and 3D-Var, as a function of mean 5-NN distance to assimilated stations, for (a) the dense and (b) the sparse benchmark. Lines show the median across $360$ time steps; shaded bands span the 10th–90th percentile range. Vertical dashed lines mark the median 5-NN distance of a random CONUS location under each benchmark ( $48$ km dense; $142$ km sparse). In the sparse benchmark, beyond ${\sim}150$ km temperature and dewpoint improvements from the generative methods turn negative while wind improvements remain positive.

Figure 5 : Rocky Mountains case study for $T_{2}$ . Top row: geographic context (left) and satellite view (right) of the region of interest. Middle row: (a) ERA5 background, (b) Flow Guidance, (c) Diffusion-SDA, (d) 3D-Var. Bottom row: per-station errors (analysis $-$ observation) for each method, with mean absolute error (MAE) and RMSE annotated. ERA5 shows a near-uniform ${\sim}6.5$ K cold bias across the region; all three methods substantially reduce the error, with the generative methods (RMSE $3.3$ – $3.5$ K) modestly outperforming 3D-Var (RMSE $4.0$ K).

Figure 6 : Great Lakes case study for $u_{10}$ . Panels as in Figure 5 . ERA5 overestimates wind speed by ${\sim}4.5$ m/s across the region. Flow Guidance and Diffusion-SDA reduce the error at the shoreline test stations (RMSE 2.39 m/s, 53.0% reduction), while 3D-Var retains larger residual errors (RMSE $2.81$ m/s, $44.8\%$ reduction).

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Interpolation	$u_{10}$	$v_{10}$	$T_{2}$	$T_{d,2}$	Average
ERA5 Baseline
Nearest	1.90 $\pm$ 0.31	1.95 $\pm$ 0.28	2.51 $\pm$ 0.31	2.72 $\pm$ 0.51	2.27 $\pm$ 0.35
Bilinear	1.88 $\pm$ 0.31	1.93 $\pm$ 0.28	2.43 $\pm$ 0.30	2.60 $\pm$ 0.45	2.21 $\pm$ 0.34
Bicubic	1.89 $\pm$ 0.31	1.94 $\pm$ 0.28	2.44 $\pm$ 0.31	2.63 $\pm$ 0.47	2.22 $\pm$ 0.34
Flow Guidance
Bilinear	1.15 $\pm$ 0.17	1.18 $\pm$ 0.18	1.75 $\pm$ 0.18	1.70 $\pm$ 0.21	1.44 $\pm$ 0.18

Appendix

Table 5: Effect of interpolation scheme on ERA5 baseline and Flow Guidance assimilation RMSE (mean

\pm

std, validation stations). Bicubic is selected based on validation performance. Bold marks the best value per column within each panel.

	$u_{10}$	$v_{10}$	$T_{2}$	$T_{d,2}$
Training std (physical units, m s -1 or K)	3.45	3.84	11.70	11.67
ERA5 validation RMSE (physical)	1.89	1.94	2.44	2.63
ERA5 validation RMSE (normalized)	0.547	0.504	0.209	0.225

Appendix

Table 6 : ERA5 baseline error in physical and normalized units.

Variant	Covariance $\bm{\Sigma}$	Notes
DPS [ 9 ]	$\bm{\Sigma}_{y}$	Ignores uncertainty in $\hat{\mathbf{x}}_{1}$ .
SDA / $\Pi$ GDM [ 47 ]	$\bm{\Sigma}_{y}+\mathbf{H}\,\bm{\Sigma}_{0}(t)\,\mathbf{H}^{\top}$	$\bm{\Sigma}_{0}=\gamma\sigma^{2}\mathbf{I}$ (SDA) or $r_{t}^{2}\mathbf{I}$ ( $\Pi$ GDM).
MMPS [ 2 ]	$\bm{\Sigma}_{y}+\mathbf{H}\,\mathbb{V}[\mathbf{x}_{1}\mid\mathbf{x}_{t}]\,\mathbf{H}^{\top}$	Exact posterior covariance; conjugate-gradient solves per step.
NDTM [ 39 ]	per-step variational optimization	Most general; high per-step cost.

Appendix

Table 7 : Common Gaussian likelihood approximations

p(\mathbf{y}^{*}\mid\mathbf{x}_{t})\approx\mathcal{N}(\mathbf{y}^{*};\,\mathcal{H}(\hat{\mathbf{x}}_{1}),\,\bm{\Sigma})

used in posterior-guidance methods. We use the SDA form throughout.

	$u_{10}$		$v_{10}$		$T_{2}$		$T_{d,2}$		Avg		Cost
Configuration	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Mem	Total
No-guidance window ( $\lambda=2\times 10^{5}$ , constant elsewhere)
None (default)	53.2	38.2	53.0	38.8	42.9	29.9	49.1	35.9	49.6	35.7	3.05	380
$[0.0,0.2]$	53.3	38.0	53.1	38.7	42.3	28.3	48.7	34.4	49.3	34.9	3.05	203
$[0.2,0.4]$	53.1	38.3	52.9	38.8	42.7	29.7	48.9	35.7	49.4	35.6	3.05	393
$[0.4,0.6]$	52.9	38.3	52.7	38.9	42.5	29.6	48.6	35.7	49.2	35.6	3.05	388

Appendix

Table 8 : Ablation: guidance schedule. Top : effect of removing guidance from 20%-wide windows (constant schedule,

\lambda{=}2{\times}10^{5}

). Bottom : comparison of four schedule shapes ( Figure 7 ), each at its optimal

\lambda

. All use dopri5, single-step terminal extrapolation, without stop gradient. RMSE improvement (%) over ERA5 at training (Tr) and test (Te) stations; total wall-clock cost (min) for 360 snapshots.

Figure 7 : The four guidance schedules compared in Table 8 . The constant schedule maintains uniform guidance strength throughout the trajectory; cosine-increasing concentrates guidance near

t{=}1

; cosine-decreasing concentrates it near

t{=}0

; and the SDA-derived schedule ( Equation 31 ) peaks at intermediate

t

and vanishes at both endpoints.

	$u_{10}$		$v_{10}$		$T_{2}$		$T_{d,2}$		Avg		Cost
Method	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Mem	Total
w/o SG (default)	53.2	38.2	53.0	38.8	42.9	29.9	49.1	35.9	49.6	35.7	3.05	380
w/ SG	60.9	34.1	61.0	34.9	54.6	27.8	58.1	32.0	58.6	32.2	0.37	77
FlowDPS (terminal-space optimization)
FlowDPS, $K=5$	52.7	35.5	52.5	35.7	27.3	2.1	38.2	12.0	42.7	21.3	0.25	28
FlowDPS, $K=10$	57.8	34.5	57.7	35.2	45.2	16.9	51.5	22.6	53.0	27.3	0.25	31

Appendix

Table 9 : Ablation: stop gradient (SG) and FlowDPS. RMSE improvement (%) over ERA5 at training (Tr) and test (Te) stations, and total wall-clock cost (min) for 360 snapshots. Without SG, gradients backpropagate through the velocity network. FlowDPS-

K

denotes FlowDPS with

K

optimization steps per ODE step.

Extrapolation	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Mem	Total
	$u_{10}$		$v_{10}$		$T_{2}$		$T_{d,2}$		Avg		Cost
Single-step (default)	53.2	38.2	53.0	38.8	42.9	29.9	49.1	35.9	49.6	35.7	3.05	380
RK4 two-step	55.7	37.6	55.4	38.3	46.4	31.0	51.7	36.4	52.3	35.8	25.21	10,659

Appendix

Table 10 : Ablation: terminal extrapolation. RMSE improvement (%) over ERA5 at training (Tr) and test (Te) stations, plus total wall-clock cost (min) for 360 snapshots.

Sampler	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Tr	Te	Mem	Total
	$u_{10}$		$v_{10}$		$T_{2}$		$T_{d,2}$		Avg		Cost
Flow Guidance dopri5 (PO, default)	53.2	38.2	53.0	38.8	42.9	29.9	49.1	35.9	49.6	35.7	3.05	380
Flow Guidance Euler-128 (PO)	49.2	38.6	48.8	39.0	34.7	24.1	40.8	29.6	43.4	32.8	2.92	53
Flow Guidance Euler-128 (PC)	48.9	38.7	48.7	39.3	35.4	26.9	42.9	34.5	44.0	34.9	2.93	208
Diffusion-SDA PC (default)	52.7	37.8	52.2	38.6	39.9	29.6	47.6	35.6	48.1	35.4	6.12	246
Diffusion-SDA PO	48.3	37.9	47.9	38.3	29.5	19.9	36.7	25.0	40.6	30.3	6.11	82

Appendix

Table 11 : Ablation: ODE solver and corrector. RMSE improvement (%) over ERA5 at training (Tr) and test (Te) stations, plus total wall-clock cost (min) for 360 snapshots. Flow Guidance dopri5 uses

\lambda{=}2{\times}10^{5}

; Euler-128 variants use

\lambda{=}10^{5}

. “PO” = predictor-only; “PC” adds a Langevin corrector between predictor steps.

Figure 8 : Perturbation amplification (

\log_{10}

scale) as a function of injection time

t

, averaged over

16

trajectories with

\delta{=}0.01

. Top row (a,b): Isolate score quality by fixing the flow ODE trajectory. Flow PO/PC use the flow-derived score (

\tau{=}0.02

); Hybrid PO/PC use the diffusion denoiser’s score (

\tau{=}0.1

) on the same trajectory. Bottom row (c,d): Isolate trajectory by fixing the diffusion score. Hybrid PO/PC integrate the flow ODE; Diffusion PO/PC integrate the VP-SDE (

\tau{=}0.1

). Left panels show guidance-direction perturbations; right panels show random-direction perturbations. Insets zoom into

t\in[0.95,1.0]

Figure 9 : Accuracy–compute trade-off for Flow Guidance in pixel vs. latent space.

x

-axis: total wall-clock cost for

360

snapshots on

1

GPU (log scale).

y

-axis: test-set RMSE improvement over ERA5, averaged over

\{u_{10},v_{10},T_{2},T_{d,2}\}

. Colors denote downsampling depth

S

; marker shapes denote the attention grouping (circle:

G{=}1

, variable-plus-spatial-mixing; square:

G{=}4

, spatial-mixing-only). Lines connect configurations of the same architecture across guidance strengths

\lambda

. Peak memory (approximately constant within each

S

-group) is annotated per group. Reported runtimes for

G{=}4

S{\geq}3

are inflated by a PyTorch grouped-convolution inefficiency that a mathematically-equivalent batch-reshape implementation removes (

1.5

–

4\times

faster).

3D-Var variant	Tr (%)	Te (%)
Pixel (no AE)	49.7	33.3
Latent ( $S{=}2$ , $G{=}1$ )	53.4	33.3
Latent ( $S{=}2$ , $G{=}4$ )	55.5	33.5

Appendix

Table 12 : 3D-Var accuracy in three nonlinear regimes. RMSE improvement (%) over ERA5 at training (Tr) and test (Te) stations, averaged across the four variables. The three variants are indistinguishable on Te.

$G$	$S$	Compr.	$u_{10}$	$v_{10}$	$T_{2}$	$T_{d,2}$	MAE
4	2	0.99 $\times$	0.0069	0.0078	0.0175	0.0177	0.0125
4	3	3.85 $\times$	0.0819	0.0808	0.1482	0.1735	0.1211
4	4	14.7 $\times$	0.1982	0.2024	0.2688	0.3322	0.2504
1	2	0.99 $\times$	0.0063	0.0067	0.0213	0.0218	0.0140
1	3	3.85 $\times$	0.0533	0.0573	0.1494	0.1897	0.1124
1	4	14.7 $\times$	0.1307	0.1377	0.3185	0.3720	0.2397

Appendix

Table 13 : Autoencoder reconstruction MAE on test fields in normalized units. “Compr.” is the spatial compression factor.

Figure 10 : Per-variable RMSE improvement vs. mean 5-NN distance, dense benchmark.

Figure 11 : Per-variable RMSE improvement vs. mean 5-NN distance, sparse benchmark.

Figure 12 : Los Angeles basin case study for $T_{2}$ . Top row: geographic context (left) and satellite view (right) of the region of interest. Middle row: (a) ERA5 background, (b) Flow Guidance, (c) Diffusion-SDA, (d) 3D-Var. Bottom row: per-station errors (analysis $-$ observation) for each method, with MAE and RMSE annotated. ERA5 overestimates mountain temperatures and underestimates coastal urban temperatures; all three methods achieve nearly identical corrections on this case (RMSE $3.6$ – $3.7$ K, $36$ – $37\%$ reduction).

Figure 13 : Florida peninsula case study for $u_{10}$ . Panels as in Figure 12 . Flow Guidance and Diffusion-SDA reduce wind RMSE by ${\sim}72\%$ on this snapshot versus $66\%$ for 3D-Var, a ${\sim}6$ pp gap that mirrors the Great Lakes finding at smaller magnitude.

	$u_{10}$			$v_{10}$			$T_{2}$			$T_{d,2}$			Average
Method	Tr	Va	Te	Tr	Va	Te	Tr	Va	Te	Tr	Va	Te	Tr	Va	Te
ERA5 (baseline)	1.87	1.89	1.88	1.93	1.94	1.91	2.40	2.44	2.38	2.62	2.63	2.53	2.20	2.22	2.18
3D-Var	0.90	1.21	1.21	0.93	1.26	1.24	1.29	1.76	1.68	1.31	1.76	1.66	1.11	1.50	1.45
Diffusion-SDA	0.88	1.16	1.16	0.92	1.18	1.17	1.43	1.73	1.67	1.36	1.69	1.61	1.15	1.44	1.40
Diffusion-SDA (PO)	0.96	1.16	1.16	1.00	1.19	1.18	1.68	1.93	1.90	1.63	1.95	1.87	1.32	1.56	1.52
D-Flow	0.92	1.21	1.22	0.97	1.23	1.22	1.59	1.87	1.82	1.55	1.84	1.79	1.26	1.54	1.51

Appendix

Table 14 : Absolute RMSE (train/validation/test) on the dense benchmark. Wind variables are in m s -1 ; temperature and dewpoint are in K.

Absolute RMSE
	$u_{10}$			$v_{10}$			$T_{2}$			$T_{d,2}$			Average
Method	Tr	Va	Te	Tr	Va	Te	Tr	Va	Te	Tr	Va	Te	Tr	Va	Te
ERA5	1.86	1.87	1.87	1.93	1.94	1.92	2.40	2.44	2.39	2.61	2.62	2.52	2.20	2.22	2.17
3D-Var	0.90	1.21	1.21	0.93	1.26	1.23	1.29	1.76	1.69	1.31	1.76	1.66	1.11	1.50	1.45
Diffusion-SDA	0.87	1.15	1.16	0.92	1.18	1.17	1.43	1.73	1.67	1.36	1.70	1.61	1.14	1.44	1.40
Flow Guidance	0.87	1.14	1.15	0.90	1.17	1.16	1.36	1.73	1.67	1.32	1.69	1.60	1.11	1.43	1.40

Appendix

Table 15 : Full-year benchmark on the dense station network (

8{,}664

snapshots, 2023). RMSE in physical units (top) and improvement over ERA5 (bottom).

Explore similar work

Multimodal Spatiotemporal Atmospheric Data Assimilation with Latent Video Flow-matching
Aug 5, 2026Dibyajyoti Chakraborty, Romit MaulikData Assimilation Era5 Reanalysis Data
Information Sciences and Technology The Pennsylvania State University University Park, Pennsylvania · School of Mechanical Engineering Purdue University West Lafayette, Indiana
ForcingDAS: Unified and Robust Data Assimilation via Diffusion Forcing
May 14, 2026Yixuan Jia, Siyi Chen, Yida Pan +9Data Assimilation Precipitation Nowcasting
University of Michigan, Ann Arbor · Michigan State University · Los Alamos National Laboratory
Generative Atmospheric Super-Resolution from Heterogeneous In Situ Observations through Composable Interfaces
Sep 24, 2026Yang Xu, Dibyajyoti Chakraborty, Haiwen Guan +2Atmospheric Dynamics Generative Super-Resolution
School of Mechanical Engineering, Purdue University, West Lafayette, Indiana, USA · College of Information Sciences and Technology, The Pennsylvania State University, University Park, Pennsylvania, USA

Aug 5, 2026, cs.LGJ/K move · Enter open · S save

Multimodal Spatiotemporal Atmospheric Data Assimilation with Latent Video Flow-matching

Dibyajyoti Chakraborty, Romit Maulik

Information Sciences and Technology The Pennsylvania State University University Park, Pennsylvania · School of Mechanical Engineering Purdue University West Lafayette, Indiana

Data Assimilation Era5 Reanalysis Data Posterior Sampling

Open paper AskPDF arXiv

Data assimilation (DA) uses Bayesian inference to update the state of a numerical forecast model with observed data. In this study, we propose a fundamentally different, unified approach to atmospheric data assimilation. We use latent video flow-matching to sample temporally consistent trajectories from a prior trained using ERA5 reanalysis (69 variables over an 8-day window). We also use posterior sampling to assimilate real observation sources, such as those from the NOAA Integrated Global Radiosonde Archive and the Integrated Surface Database. Because the prior generates a continuous trajectory, it naturally propagates information between observed and unobserved frames. Therefore, we can perform various DA tasks, such as filtering and smoothing, simply by changing the observed frames. Moreover, we generate full-state ensemble forecasts directly from sparse observations, achieving performance competitive with state-of-the-art observation-to-forecast models.

Abstract

Explore similar work

Multimodal Spatiotemporal Atmospheric Data Assimilation with Latent Video Flow-matching

ForcingDAS: Unified and Robust Data Assimilation via Diffusion Forcing

Generative Atmospheric Super-Resolution from Heterogeneous In Situ Observations through Composable Interfaces