UniGIO: Unified Generative Global In-situ Weather Modeling from Spatiotemporal Incomplete Observations

Songru Yang; Zili Liu; Tao Han; Ben Fei; Lei Bai; Chang Liu; Zhengxia Zou; Xiangyang Ji

cs.LGSep 2, 2026

UniGIO: Unified Generative Global In-situ Weather Modeling from Spatiotemporal Incomplete Observations

Authors: Songru Yang, Zili Liu, Tao Han, Ben Fei, Lei Bai, Chang Liu, Zhengxia Zou, Xiangyang Ji,

Organizations: Department of Aerospace Intelligent Science and Technology, School of Astronautics, Beihang University, and with the Key Laboratory of Spacecraft Design Optimization and Dynamic Simulation Technologies, Ministry of Education, Beihang University, Beijing 100191, China · Shanghai Artificial Intelligence Laboratory, Shanghai 200232, China · Department of Information Engineering, The Chinese University of Hong Kong, Hong Kong 999077, China · Department of Automation, Tsinghua University, Beijing 100084, China

Abstract

Global In-situ Observation (GIO) provides fine-scale, direct records of the global weather system from sparse point stations, making it an indispensable source for capturing localized and transient dynamics beyond the reach of satellite gridded data, and playing a critical role in key fields such as numerical weather prediction, disaster prevention, and agriculture. However, GIO exhibits strong spatiotemporal incompleteness, severely impairing accurate and real-time in-situ weather modeling. Unlike existing methods waiting for completed AI-ready data with extra introduced errors, in this work, we explore UniGIO, a novel generative framework for directly modeling global in-situ weather dynamics from native incomplete GIO. By generating missing data from observed ones annotated by masks, it unifies the coexisting forecasting, imputation, and generation under arbitrary missing ratios. Between the missing and observed, UniGIO captures station and region level complementarity through the Observation Mixer and Event Aligner, which diffuse discrete observations into continuous spaces where weather processes naturally span multiple stations. We further establish temporal dependencies with pattern shifts using the Adaptive Temporal Mixer, and track extreme events in chaotic local weather systems through a Mixture-ofExperts structure. Steady and extreme events are adapted in decoder by a Local Refiner. Extensive experiments on the up-todate largest global station weather dataset Weather-5K validate its SOTA performance with 11%, 12%, and 5% advantages on accuracy, fidelity, and extreme event capture, delivering a novel holistic solution for weather modeling in GIO networks.

Figures & tables

Fig. 1: Spatiotemporal incompleteness in GIO makes it difficult to model local weather systems and detect extreme events at the finest scale. We analyzed the station density, annual missing rate, and the AI-ready rate to demonstrate its severity. AI-ready rate: The ratio of non-missing to all samples when station data is spliced in windows.

Fig. 2: (a) Mask-driven observation-to-missing generation, (b) UniGIO achieves leading performance while maintaining lightweight in data and model scale

Fig. 3: Overall pipeline of UniGIO. The framework is organized as a lightweight CVAE, where key designs for GIO complementation are marked in blue and key designs for extreme pattern capture are marked in green .

Fig. 4: Detailed designs in Observation Mixer: (a) Station-Centric Field Complement to diffuse discrete station level observations into continuous local space with field and memory tokens; (b) Event Aligner corrects inter-station temporal delays via frequency-correlation alignment.

Fig. 5: Details for extreme pattern capture, assessment, and generation. (a) Uncertainty Score and Pattern Decoupled MoE suppress missing-induced state noise and decouple global similarity from local heterogeneity. (b) Extreme event Assessment and Generation use CVAE variance and sampling-noise control to estimate and generate extremes.

Fig. 6: UniGIO achieves SOTA performance on accuracy, extreme event capture, and fidelity across all time windows.

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	MAE $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	MSE $\downarrow$ ( $\times 10^{3}$ )	MAE $\downarrow$	MSE $\downarrow$
Deterministic	4.54 / 4.32	25.5 / 26.8	3.37 / 3.54	16.7 / 17.5	2.10 / 2.16	6.19 / 5.98	92.8 / 97.4	10.5 / 11.0	4.77 / 4.96	34.8 / 33.1
TimeGAN	5.51 / 10.3	37.8 / 76.5	3.91 / 6.98	22.4 / 41.5	2.51 / 4.33	8.31 / 14.7	103 / 147	12.5 / 17.9	5.77 / 8.69	49.8 / 78.5
TimeVAE	4.58 / 5.36	25.6 / 29.9	3.43 / 4.02	16.8 / 19.7	2.09 / 2.45	6.04 / 7.08	96.9 / 114	10.6 / 12.4	5.16 / 6.04	37.5 / 43.9
Diffwave	5.14 / 5.76	34.8 / 38.8	3.74 / 4.17	21.5 / 24.7	2.45 / 2.48	8.84 / 8.76	105 / 119	14.5 / 15.9	4.99 / 5.14	40.6 / 43.9
SSSD	4.98 / 5.39	33.1 / 37.9	3.88 / 4.25	23.9 / 27.0	2.37 / 2.69	8.65 / 9.52	107 / 111	15.4 / 16.1	5.00 / 5.78	41.6 / 48.1

TABLE I: Quantitative Accuracy Comparison on Curated / Native Masks (120h)

{\dagger}

means re-implemented

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$
Deterministic	0.07 / 0.07	0.35 / 0.33	0.12 / 0.11	0.50 / 0.53	0.01 / 0.01	0.10 / 0.20	$\backslash$	$\backslash$	0.08 / 0.08	0.40 / 0.38
TimeGAN	0.10 / 0.19	0.46 / 0.48	0.21 / 0.28	0.58 / 0.60	0.03 / 0.06	0.19 / 0.26	$\backslash$	$\backslash$	0.19 / 0.19	0.50 / 0.51
TimeVAE	0.11 / 0.12	0.47 / 0.47	0.19 / 0.19	0.60 / 0.60	0.02 / 0.02	0.14 / 0.14	$\backslash$	$\backslash$	0.13 / 0.13	0.52 / 0.51
Diffwave	0.12 / 0.12	0.39 / 0.40	0.19 / 0.19	0.53 / 0.54	0.10 / 0.10	0.22 / 0.22	$\backslash$	$\backslash$	0.15 / 0.15	0.44 / 0.44
SSSD	0.12 / 0.13	0.40 / 0.42	0.20 / 0.21	0.53 / 0.55	0.12 / 0.12	0.25 / 0.26	$\backslash$	$\backslash$	0.15 / 0.16	0.44 / 0.46

TABLE II: Quantitative F1-SEDI Comparison on Curated / Native Masks (120h)

{\dagger}

means re-implemented

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure		CroS $\downarrow$
Baselines	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$ ( $\times 10^{3}$ )	FID $\downarrow$	STE $\downarrow$	CroS $\downarrow$
Deterministic	4.76 / 4.95	6.21 / 6.42	4.73 / 4.50	2.16 / 2.05	4.54 / 4.31	1.66 / 1.74	3.88 / 4.08	3.98 / 4.18	5.40 / 5.13	1.58 / 1.66	3.20 / 3.36
TimeGAN	1.22 / 0.95	5.53 / 5.38	1.11 / 0.89	2.32 / 2.29	1.08 / 0.78	1.50 / 1.52	0.95 / 0.72	3.80 / 3.80	2.97 / 2.51	2.53 / 2.51	3.10 / 3.17
TimeVAE	1.28 / 1.20	6.57 / 6.22	1.88 / 1.78	2.42 / 2.31	1.64 / 1.56	1.80 / 1.71	1.38 / 1.31	4.18 / 3.94	3.37 / 3.20	1.67 / 1.59	4.45 / 4.33
Diffwave	0.92 / 0.87	6.01 / 5.69	0.53 / 0.50	3.04 / 2.89	0.46 / 0.44	1.81 / 1.72	0.67 / 0.75	4.24 / 4.03	2.21 / 2.11	3.02 / 2.87	3.13 / 3.14
SSSD	0.82 / 0.78	6.05 / 5.75	0.51 / 0.48	3.77 / 3.58	0.48 / 0.36	2.01 / 1.91	0.64 / 0.71	4.65 / 4.42	1.28 / 1.20	2.54 / 2.41	3.16 / 3.11

TABLE III: Quantitative Fidelity Comparison on Curated / Native Masks (120h)

{\dagger}

means re-implemented

Baselines	Window	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	Window	MAE $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	MSE $\downarrow$ ( $\times 10^{3}$ )	MAE $\downarrow$	MSE $\downarrow$
TimeGAN	72h	5.24 / 9.42	33.2 / 63.4	3.33 / 5.71	16.3 / 28.7	2.49 / 4.20	8.05 / 13.8	97.0 / 136	11.4 / 15.9	4.55 / 6.49	31.1 / 45.0
	240h	5.88 / 11.2	43.5 / 88.8	4.51 / 8.18	28.9 / 54.4	2.82 / 5.16	10.4 / 19.8	111 / 165	13.9 / 20.8	6.66 / 10.2	63.9 / 101
TimeVAE	72h	4.41 / 3.36	23.2 / 17.7	3.00 / 2.28	12.9 / 9.87	2.03 / 1.54	5.52 / 4.20	93.0 / 70.9	10.1 / 7.66	4.24 / 3.23	25.8 / 19.6
	240h	4.88 / 11.6	29.7 / 70.6	3.94 / 9.34	21.8 / 51.6	2.20 / 5.22	6.75 / 16.0	103 / 244	11.4 / 27.2	6.00 / 14.2	50.4 / 119
UniMTD †	72h	2.00 / 2.03	8.96 / 10.4	2.04 / 1.99	9.94 / 10.7	1.36 / 1.28	4.05 / 4.21	66.2 / 61.1	8.21 / 7.08	2.65 / 2.35	19.8 / 17.2

TABLE IV: Quantitative Accuracy Comparison on Curated / Native Masks (72h/240h)

{\dagger}

means re-implemented

Baselines	Window	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	Window	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$
TimeGAN	72h	0.15 / 0.23	0.49 / 0.51	0.28 / 0.34	0.65 / 0.66	0.04 / 0.08	0.24 / 0.29	/	/	0.35 / 0.35	0.62 / 0.63
	240h	0.04 / 0.06	0.39 / 0.40	0.10 / 0.14	0.48 / 0.50	0.02 / 0.05	0.15 / 0.20	/	/	0.03 / 0.04	0.36 / 0.36
TimeVAE	72h	0.15 / 0.16	0.50 / 0.51	0.27 / 0.26	0.66 / 0.66	0.03 / 0.02	0.19 / 0.19	/	/	0.33 / 0.33	0.63 / 0.63
	240h	0.03 / 0.03	0.40 / 0.39	0.05 / 0.06	0.50 / 0.50	0.02 / 0.01	0.09 / 0.10	/	/	0.01 / 0.01	0.34 / 0.34
UniMTD †	72h	0.48 / 0.49	0.75 / 0.75	0.37 / 0.45	0.68 / 0.71	0.17 / 0.24	0.37 / 0.46	/	/	0.51 / 0.55	0.70 / 0.72

TABLE V: Quantitative F1-SEDI Comparison on Curated / Native Masks (72h/240h)

{\dagger}

means re-implemented

Baselines	Window	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure		CroS $\downarrow$
Baselines	Window	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$ ( $\times 10^{3}$ )	FID $\downarrow$	STE $\downarrow$	CroS $\downarrow$
TimeGAN	72h	0.47 / 0.39	5.67 / 5.54	0.42 / 0.34	2.16 / 2.15	0.42 / 0.32	1.47 / 1.49	0.90 / 0.91	3.78 / 3.78	1.18 / 1.06	2.03 / 2.00	3.42 / 3.50
TimeGAN	240h	3.24 / 2.61	5.76 / 5.65	3.01 / 2.38	2.57 / 2.51	3.01 / 2.28	1.53 / 1.54	3.42 / 2.57	3.91 / 3.89	5.65 / 4.77	3.26 / 3.23	2.74 / 2.82
TimeVAE	72h	0.28 / 0.26	5.34 / 5.11	0.75 / 0.67	2.52 / 2.27	0.39 / 0.37	1.24 / 1.19	0.89 / 0.86	3.06 / 3.11	1.56 / 1.69	1.31 / 1.44	4.53 / 4.77
TimeVAE	240h	2.74 / 2.43	6.72 / 6.74	3.04 / 3.37	2.48 / 2.42	4.11 / 3.80	1.85 / 1.80	4.06 / 3.93	4.26 / 4.13	6.15 / 4.90	1.68 / 1.61	4.24 / 4.25
UniMTD ${\dagger}$	72h	0.19 / 0.22	3.07 / 2.03	0.31 / 0.17	1.53 / 1.16	0.26 / 0.30	1.02 / 0.66	0.56 / 0.67	2.61 / 1.59	0.15 / 0.12	1.26 / 0.87	2.56 / 1.52

TABLE VI: Quantitative Fidelity Comparison on Curated / Native Masks (72h/240h)

{\dagger}

means re-implemented

Method

TimeGAN

TimeVAE

Diffwave

SSSD

Diffusion-TS

UniMTD

SportsTraj

UniGIO-M

UniGIO

ESFM-S

Chronos2

WSSM

iTransformer

Corrformer

Params.

3.4M

2.3M

4.1M

4.3M

4.9M

10.1M

7.3M

0.9M

3.4M

115M

120M

1.8M

0.9M

55.6M

TABLE VII: Parameter Comparison of Unified and Task-specific GIO models.

Fig. 7: Due to space constraints, only wind rate and sea-level pressure are displayed for imputation and generation tasks

Fig. 8: UniGIO achieves SOTA performance on Forecasting, Imputation, and Generation tasks, even beating Foundation Models.

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ ( $\times 10^{3}$ )/90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$
Variable-Length Forecasting
TimeGAN	7.40 / 0.11	53.7 / 0.46	5.15 / 0.19	30.7 / 0.59	3.23 / 0.02	10.8 / 0.18	130.6 /	16.00 /	7.45 / 0.17	65.8 / 0.50
TimeVAE	5.85 / 0.10	32.77 / 0.47	4.38 / 0.17	21.51 / 0.61	2.65 / 0.02	7.56 / 0.13	123.9 /	13.6 /	6.52 / 0.12	46.9 / 0.51
UniMTD ${\dagger}$	2.35 / 0.40	11.3 / 0.70	2.27 / 0.35	11.7 / 0.65	1.59 / 0.22	5.29 / 0.35	76.3 /	10.7 /	2.79 / 0.41	19.5 / 0.67
SportsTraj ${\dagger}$	3.32 / 0.30	17.2 / 0.55	2.61 / 0.32	12.1 / 0.64	1.94 / 0.17	6.14 / 0.30	88.9 /	11.7 /	2.59 / 0.42	12.9 / 0.71

TABLE VIII: Quantitative comparison under variable-length, fixed-length, and noisy forecasting settings.

{\dagger}

denotes re-implemented methods, and

\star

denotes forecasting with an imputed look-back window.

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ ( $\times 10^{3}$ )/90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$
HRES	1.76 / -	7.39 / -	1.85 / -	7.94 / -	1.48 / -	4.53 / -	63.8 /	7.15 /	0.86 / -	2.68 / -
ESFM	1.76 / -	- / -	1.86 / -	- / -	1.40 / -	- / -	46.3 /	- /	2.55 / -	- / -
Chronos2	2.05 / 0.30	9.5 / 0.58	2.06 / 0.31	12.3 / 0.65	1.47 / 0.08	4.51 / 0.24	64.4 /	8.65 /	2.13 / 0.52	11.3 / 0.72
UniGIO fixed	1.28 / 0.63	3.68 / 0.83	1.32 / 0.52	4.29 / 0.78	1.15 / 0.20	2.88 / 0.45	58.6 /	7.63 /	1.05 / 0.76	2.82 / 0.87

TABLE IX: Quantitative Comparison with Finetuned Observation and Time Series Foundation models on 24h Forecast Task

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ ( $\times 10^{3}$ )/90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$
TimeGAN	8.84 / 0.14	64.84 / 0.42	6.10 / 0.23	36.4 / 0.57	3.74 / 0.03	12.48 / 0.20	152 /	18.5 /	8.94 / 0.20	80.6 / 0.49
TimeVAE	6.77 / 0.16	37.7 / 0.47	5.14 / 0.22	25.3 / 0.60	3.10 / 0.02	8.87 / 0.14	145 /	15.9 /	7.71 / 0.14	56.3 / 0.52
UniMTD †	1.78 / 0.51	6.69 / 0.77	1.61 / 0.49	6.20 / 0.75	1.21 / 0.19	3.13 / 0.42	59.5 /	7.01 /	1.89 / 0.62	9.23 / 0.78
SportsTraj †	1.74 / 0.60	3.80 / 0.84	1.65 / 0.58	3.88 / 0.82	1.52 / 0.30	3.17 / 0.49	73.9 /	7.87 /	1.25 / 0.82	1.93 / 0.91
SSSD	5.49 / 0.22	35.21 / 0.51	4.51 / 0.26	28.23 / 0.59	2.52 / 0.17	8.85 / 0.27	116.82 /	16.46 /	5.67 / 0.31	47.54 / 0.57

TABLE X: Quantitative Comparison on Imputation Task

{\dagger}

means re-implemented

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ ( $\times 10^{3}$ )/90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$
TimeGAN	3.81 / 0.12	23.9 / 0.47	2.80 / 0.20	15.2 / 0.59	1.87 / 0.03	6.15 / 0.20	77.7 /	9.46 /	4.19 / 0.19	34.5 / 0.51
TimeVAE	3.46 / 0.11	19.3 / 0.47	2.57 / 0.18	12.5 / 0.60	1.59 / 0.02	4.63 / 0.15	72.7 /	7.96 /	3.89 / 0.13	28.2 / 0.52
UniMTD †	2.60 / 0.33	14.15 / 0.67	2.83 / 0.23	17.2 / 0.57	1.57 / 0.10	5.31 / 0.27	75.2 /	9.58 /	4.39 / 0.24	43.2 / 0.50
SportsTraj †	2.10 / 0.41	9.06 / 0.73	2.04 / 0.34	9.35 / 0.69	1.63 / 0.20	5.55 / 0.35	80.8 /	11.6 /	1.46 / 0.71	5.19 / 0.84
Diffusion-TS	4.23 / 0.13	33.06 / 0.42	3.06 / 0.19	20.44 / 0.54	1.94 / 0.15	7.80 / 0.25	88.58 /	13.59 /	4.22 / 0.21	38.68 / 0.49

TABLE XI: Quantitative Comparison on Generation Task

{\dagger}

means re-implemented

Settings	Temperature		Wind Rate		Wind Direc.		Temperature		Wind Rate
	MAE $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	MSE $\downarrow$ ( $\times 10^{3}$ )	99.5th $\uparrow$	90.0th $\uparrow$	99.5th $\uparrow$	90.0th $\uparrow$
World wide	1.88	8.03	1.52	5.13	75.3	10.4	0.44	0.75	0.16	0.34
All stations	3.14 / 2.22	20.5 / 11.3	1.81 / 1.62	6.91 / 5.77	80.9 / 84.9	11.7 / 12.2	0.13 / 0.29	0.46 / 0.64	0.08 / 0.12	0.20 / 0.23
90% stations	2.28 / 2.07	15.9 / 9.67	1.83 / 1.52	7.48 / 4.77	74.0 / 82.1	10.2 / 11.4	0.22 / 0.27	0.60 / 0.68	0.08 / 0.08	0.25 / 0.24
60% stations	2.06 / 1.84	8.76 / 7.99	1.68 / 1.36	6.11 / 3.78	73.4 / 80.9	10.1 / 11.1	0.32 / 0.30	0.64 / 0.71	0.14 / 0.10	0.27 / 0.25
30% stations	2.04 / 1.70	8.20 / 7.17	1.65 / 1.35	5.76 / 3.05	72.7 / 79.8	10.4 / 10.7	0.36 / 0.33	0.70 / 0.72	0.13 / 0.16	0.28 / 0.27

TABLE XII: Quantitative Evaluation on cross-region (South America / Africa) Generation

Fig. 9: Distribution coverage of results against GT across baselines and different mask settings. Less red area is better. TSNE plots look unclustered because GIO is largely continuous with no ‘category’ between randomly selected samples for visualization.

Fig. 10: Wind rate with multiple extreme events and stable sea-level pressure under different sampling noise intensities

Fig. 11: Ablation study results (right) and the representative expert specializations across variables, locations, uncertainties, and temporal patterns (left), different experts are denoted by colors

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure
	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ ( $\times 10^{3}$ )/90.0th $\uparrow$	MAE $\downarrow$ /99.5th $\uparrow$	MSE $\downarrow$ /90.0th $\uparrow$
UniGIO 3.4M	1.45 / 0.55	5.35 / 0.81	1.50 / 0.46	5.99 / 0.75	1.27 / 0.22	3.63 / 0.39	64.6 /	8.45 /	1.02 / 0.79	3.67 / 0.88
UniGIO-M 0.9M	1.47 / 0.52	5.89 / 0.76	1.55 / 0.45	6.16 / 0.75	1.33 / 0.19	3.74 / 0.37	66.0 /	8.59 /	1.09 / 0.77	3.82 / 0.86
expert 32 4.5M	1.41 / 0.59	4.71 / 0.81	1.45 / 0.49	5.46 / 0.76	1.27 / 0.20	3.49 / 0.41	65.1 /	8.32 /	0.84 / 0.83	2.94 / 0.91
sample 0	1.40 / 0.45	4.14 / 0.66	1.41 / 0.37	5.24 / 0.58	1.15 / 0.10	3.02 / 0.27	56.4 /	6.05 /	0.73 / 0.76	1.97 / 0.77
sample 1.5	1.84 / 0.55	7.54 / 0.81	1.83 / 0.49	8.20 / 0.76	1.77 / 0.27	4.91 / 0.49	86.4 /	13.40 /	1.29 / 0.81	4.59 / 0.89

TABLE XIII: Ablation Study on Accuracy and F1-SEDI

Baselines	Temperature		Dewpoint		Wind Rate		Wind Direc.		Sea-Level Pressure		CroS $\downarrow$
Baselines	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$	FID $\downarrow$	STE $\downarrow$ ( $\times 10^{3}$ )	FID $\downarrow$	STE $\downarrow$	CroS $\downarrow$
UniGIO 3.4M	0.17	2.48	0.25	1.45	0.28	1.07	0.60	2.53	0.11	0.79	1.46
UniGIO-M 0.9M	0.19	2.47	0.29	1.46	0.30	1.09	0.64	2.63	0.05	0.71	1.49
expert 32 4.5M	0.14	2.45	0.21	1.43	0.23	0.97	0.55	2.46	0.02	0.68	1.43
sample 0	0.29	2.77	0.41	1.48	0.54	1.10	0.94	2.83	0.19	0.72	1.54
sample 1.5	0.27	3.09	0.45	1.87	0.66	1.57	1.25	4.11	0.26	1.15	1.56

TABLE XIV: Ablation Study on Fidelity

Songru Yang received the B.S. degree from Beihang University, Beijing, China in 2023. He is pursuing the Ph.D. degree in the Department of Aerospace Intelligent Science and Technology, School of Astronautics, Beihang University. His research interests include machine learning, parttern recognition and AI4S.

Table 26

Weather Forecasting

Weather forecasting predicts future atmospheric conditions, from local storms to global patterns, using physical models, observations, and machine learning.

See all 176 papers

WeatherNext 3: Increasing resolution and performance of global weather models with raw observations
Sep 3, 2026Stephan Rasp, Boris Babenko, Dominic Masters +22Probabilistic Forecasting Weather Forecasting
1Google DeepMind · 2Google Research · 3Google
Benchmarking Generative Models for Near-Surface Data Assimilation on Real Station Observations
Sep 30, 2026Ruizhe Huang, Qidong Yang, Jonathan Giezendanner +1Flow Matching Data Assimilation
Massachusetts Institute of Technology, Cambridge, MA, USA
Diffusion Precipitation Generation with Medium-range Weather Forecasting
Oct 9, 2026Guowen Li, Yang Liu, Mengxuan Chen +11Precipitation Forecasting Weather Forecasting
Tsinghua University · The Chinese University of Hong Kong · The Hong Kong University of S & T +5

Sep 3, 2026J/K move · Enter open · S save

WeatherNext 3: Increasing resolution and performance of global weather models with raw observations

Stephan Rasp, Boris Babenko, Dominic Masters, Andrew El-Kadi+21

1Google DeepMind · 2Google Research · 3Google

Probabilistic Forecasting Weather Forecasting

Open paper AskPDF arXiv

State-of-the-art AI weather models have shown impressive medium-range forecast skill and computational efficiency, but suffer two key shortcomings: their forecasts have lower spatial and temporal resolution than the best physics-based models and they are exclusively initialized with and trained on analysis data. As a result, they cannot directly make use of observations, and any biases in the analysis are inherited by the forecast. WeatherNext 3 addresses these shortcomings and establishes a new state-of-the-art for probabilistic medium-range forecasting skill. First, WeatherNext 3 generates new forecasts every hour (rather than every 6 hours like traditional global models) by ingesting low-latency geostationary satellite data. Second, WeatherNext 3's temporal and spatial resolution are on par with physics-based global models, with hourly time steps and 0.1 degree resolution for single-level variables, including solar radiation and cloud cover. Third, WeatherNext 3 moves beyond traditional analysis variables by learning to predict satellite-derived precipitation estimates, as well as tropical cyclone and station observations. Modelling sparse station data allows WeatherNext 3 to make 2m temperature and dewpoint predictions at any location and time, conditioned on local geographical features, with substantially lower error than competing global models, even when evaluated against unseen stations. Together, WeatherNext 3's capabilities move operational AI-based weather forecasting beyond emulating the traditionally distinct stages of data assimilation, forecasting and post-processing, which helps to further push the frontier of performance and granularity for global weather prediction.

Abstract

Weather Forecasting

Explore similar work

WeatherNext 3: Increasing resolution and performance of global weather models with raw observations

Benchmarking Generative Models for Near-Surface Data Assimilation on Real Station Observations

Diffusion Precipitation Generation with Medium-range Weather Forecasting