A-UTE: Advection Informed, Uncertainty Aware Temperature Emulator

Hira Saleem; Cormac Purcell; Steven Sherwood; Flora Salim

physics.ao-phOct 29, 2024

A-UTE: Advection Informed, Uncertainty Aware Temperature Emulator

Authors: Hira Saleem, Cormac Purcell, Steven Sherwood, Flora Salim

Organizations: University of New South Wales, Sydney, NSW, Australia

Abstract

Physics-based Earth system models (ESMs) are essential for attributing climate change and generating scenario projections, yet their reliance on high-resolution numerical integration makes multi-decadal experiments expensive. In parallel, deep learning has delivered strong gains in short-range weather forecasting; however, auto-regressive roll-outs can accumulate error and become unstable when extended to decade-scale climate emulation. We introduce A-UTE: Advection Informed, Uncertainty Aware Temperature Emulator, aimed at stable multi-year emulation across heterogeneous climate models and grid resolutions. A-UTE is trained on various physics-based models at varying spatial resolutions to emulate temperature fields over a 10-year horizon. A-UTE formulates climate emulation as a forward-time stochastic dynamical system. We propose an auto-regressive ODE-SDE surrogate in which transport dynamics are constrained by an advection consistent ODE component, while a learned neural SDE term models coarse-grained variability and cross-model discrepancy at monthly resolution. We train A-UTE under negative log-likelihood objective for principled uncertainty estimates and probabilistic evaluation. Experiments across 20 climate models show that A-UTE improves long roll-out stability and accuracy relative to relevant baselines, advancing data-driven climate emulation with explicit physical structure and uncertainty-aware predictions.

Figures & tables

Figure 1. GCM vs. A-UTE (emulation). General Circulation Models (GCMs) numerically solve the coupled atmosphere physics to simulate climate evolution under prescribed forcings, producing physically grounded trajectories at high computational cost. A-UTE is a hybrid data-driven emulator, it uses a forced advection-based dynamical core to provide a coarse-grained transport prior, and a latent Neural SDE residual trained with a likelihood objective to produce calibrated probabilistic roll-outs at a fraction of the runtime, enabling efficient long-horizon scenario analysis and cross-model evaluation.

Figure 2. Architectural pipeline of A-UTE. Forcing maps and the initial temperature field are used as input for the physics-informed backbone (left), where an initial pseudo-velocity is inferred and the forced-advection equation is integrated with the dopri5 ODE solver to produce a transport-shaped backbone trajectory. This trajectory is refined by the latent neural SDE (right): an encoder $\mathcal{E}_{\theta}$ maps it to a latent state, an Itô SDE with drift $f_{\theta}$ and diffusion $g_{\theta}$ evolves the latent trajectory, and a decoder $\mathcal{D}_{\theta}$ returns the refined temperature field. The full pipeline is trained end-to-end under the latitude-weighted Gaussian negative log-likelihood objective.

Hyperparameters	Value
Epochs	30
Conv2d kernel size	3x3
Integrator	srk
Noise type	diagnol
ODE solver	dopri5
Pooling No.1	AvgPool2d (2,2)

Table 1. A-UTE Training Details

ConvLayers	Value
Encoder	4
Latent SDE - Drift $f_{\theta}$	11
Latent SDE - Diffusion $g_{\theta}$	1
Decoder	4
Velocity Refiner	2

Table 2. A-UTE ConvLayer Details

	A-UTE		U-Net		ConvLSTM		SFNO		NADE
Climate Model	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS
AWI-CM-1-1-MR	0.369	0.251	0.916	0.842	0.543	0.441	0.406	0.316	0.412	0.327
BCC-CSM2-MR	0.381	0.263	0.950	0.871	0.549	0.451	0.412	0.323	0.435	0.355
CAS-ESM2-0	0.420	0.306	1.431	0.921	0.542	0.439	0.451	0.362	0.432	0.330
CESM2-WACCM	0.333	0.233	1.154	0.932	0.556	0.459	0.463	0.365	0.461	0.351
CESM2	0.353	0.248	0.926	0.811	0.557	0.457	0.475	0.397	0.387	0.291

Table 3. 10-year auto-regressive roll-outs on 19 ClimateSet models under the held-out SSP2-4.5 scenario. Latitude-weighted RMSE (K) and CRPS (K); lower is better. CRPS for the deterministic baselines is computed from a 5-seed empirical ensemble (Section 5.2 ). Best entry per metric in bold.

	A-UTE		U-Net		ConvLSTM		SFNO		NADE
Level	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS
$T_{1}$	0.437	0.347	0.924	0.876	0.653	0.598	0.466	0.354	0.452	0.342
$T_{3}$	0.408	0.274	0.865	0.772	0.656	0.582	0.482	0.392	0.492	0.398
$T_{4}$	0.355	0.222	0.796	0.622	0.592	0.462	0.491	0.365	0.382	0.323
$T_{5}$	0.349	0.221	0.899	0.731	0.536	0.401	0.379	0.399	0.421	0.392
$T_{6}$	0.379	0.240	0.738	0.598	0.582	0.421	0.459	0.387	0.451	0.372

Table 4. 10-year auto-regressive roll-outs on FV3GFS at six vertical levels. Latitude-weighted RMSE (K) and CRPS (K); lower is better. CRPS for the deterministic baselines is computed from a 5-seed empirical ensemble. Best entry per metric in bold.

Panel of Figure 3 — Figure 3. Time–mean maps for air temperature fields across vertical levels (2,3,4,5). Both the quantitative results in table 4 and visual results indicate stronger fidelity at higher levels (L4–L5) than at lower tropospheric levels (L2–L3), with reduced discretizations.

Source: AWI-CM-1-1-MR
	A-UTE		U-Net		ConvLSTM		SFNO		NADE
Target	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS	RMSE	CRPS
AWI-CM-1-1-MR	0.190	0.134	0.534	0.423	0.589	0.479	0.322	0.210	0.298	0.182
EC-Earth3	0.214	0.149	0.759	0.598	0.590	0.481	0.342	0.230	0.310	0.230
GFDL-ESM4	0.198	0.136	0.603	0.502	0.599	0.489	0.325	0.213	0.242	0.139
FGOALS-f3-L	0.225	0.155	0.659	0.553	0.587	0.477	0.351	0.240	0.336	0.215

Table 5. Zero-shot cross-model emulation at the 1-year roll-out horizon. Each block lists the source model in its first row (in-domain) and target models evaluated without fine-tuning. RMSE (K) and CRPS (K); lower is better. CRPS for the deterministic baselines reflects the 5-seed empirical construction. Best entry per metric in bold.

Figure 4. Ablation of the advection-forcing core across four CMIP6 models. Latitude-weighted RMSE for 10-year roll-outs comparing the full A-UTE (physics-informed) against the variant with the transport core removed and only the latent neural SDE retained (physics-uninformed). Removing the core raises RMSE on all four models, with the largest degradation on CAS-ESM2-0 and TaiESM1.

Climate Emulation

See all 20 papers

Sep 19, 2025

ArchesClimate: Probabilistic Decadal Ensemble Generation With Flow Matching

Internal variability is a dominant contributor to the uncertainty of predictions at the interannual to decadal timescale. A typical approach to separating the internal variability from forced climate responses is to generate large ensembles of simulations under different initial conditions. Due to the complexity of Earth System Models, generating these large ensembles is computationally expensive. In this work, we present ArchesClimate, a deep learning-based climate model emulator designed to reduce the cost of exploring internal variability at timescales ranging from monthly to decadal. ArchesClimate is trained on decadal hindcasts of the IPSL-CM6A-LR climate model. We train a flow matching model following ArchesWeatherGen, which we adapt to predict near-term climate. Once trained, the model generates states at a one-month lead time from the states of the two preceding months, and can be used to auto-regressively emulate climate model simulations. We show that for up to 10 years, these generations are stable and physically consistent. We also show that for several important climate variables, ArchesClimate generates simulations that are interchangeable with the IPSL model. This work suggests that climate model emulators could reduce the cost of generating large ensembles with climate models.

Graham Clyne, Guillaume Couairon, Guillaume Gastineau +2

ARCHES, INRIA, Paris, France · UMR LOCEAN, IPSL, Sorbonne Universit´e, IRD, CNRS, MNHN, Paris, France · Department of Computer Science, University of Colorado Boulder, USA

Neural Surrogate Modeling Climate Science Climate Modeling Conditional Generative Modeling Autoregressive Modeling Climate Emulation

Ask PDF

Sep 28, 2026

CasEm: A Cascade Architecture for Long-Horizon Neural Emulation

Autoregressive neural emulators can drift or diverge over long rollouts despite accurate short-term predictions. We introduce Cascaded Emulation (CasEm), a one-way rollout architecture that augments an existing full-state backbone with an independently evolving model of physically specified aggregates. Its forecasts guide corrections to full-state predictions, without feedback from the backbone to the aggregate model. Effective guidance requires aggregates that cover substantial backbone error, remain accurately predictable, and support useful full-state corrections. We derive a finite-horizon error bound that clarifies these three factors and use empirical diagnostics to guide subsystem selection. Across four ODE/PDE benchmarks, CasEm reduces long-horizon rollout errors across diverse backbones and suppresses the trend toward error divergence in both diffusion tasks using Fourier neural operator backbones. In global climate emulation, CasEm with a regional total-water subsystem reduces 10-year full-state time-mean error by 66.6% and 46.3% for frozen ACE and Spherical DYffusion backbones, respectively, while adding less than 3% to inference time.

Zhaoyi Li, Jingtao Ding, Shihua Li

School of Automation, Southeast University · Beijing Zhongguancun Academy · Department of Earth System Science, Tsinghua University

Neural Surrogate Modeling PDE Surrogate Modeling

Ask PDF

May 24, 2026

Samudra 2: Scaling Ocean Emulators across Resolutions

Ocean general circulation models (OGCMs) are essential to climate science but computationally expensive, limiting ensemble size and forcing scenarios. Neural emulators promise orders-of-magnitude speedups, yet existing ocean emulators have not combined fine spatial resolution with multi-year autoregressive rollouts. Samudra, the first autoregressive neural ocean emulator to produce multi-decade global rollouts, is limited to

1^\circ

resolution and exhibits two long-horizon failure modes: variance collapse, the loss of temporal variability, and imprinting artifacts, in which velocity patterns leak into deep-ocean fields. We present Samudra 2, which introduces a wider U-Net backbone with modified ConvNeXt-style blocks and a reduced block-internal expansion factor, together with a dynamic loss that reweights output channels according to their prediction errors, strengthening gradients for slow-evolving deep-ocean fields. At

1^\circ

, Samudra 2 increases upper-ocean global-mean temperature

R^2

from 0.56 to 0.87 and reduces deep-ocean temperature error by roughly sevenfold. The same architecture scales to

1/2^\circ

and

1/4^\circ

over approximately 8-year autoregressive rollouts, recovering mesoscale eddies and sharp western boundary currents. Running on a single GPU, Samudra 2 enables larger ensembles for sea-level projections, ocean heat uptake, and climate variability studies. All artifacts are publicly available: project page, code, checkpoints, documentation.

Yuan Yuan, Jesse Rusak, Alexander Merose +5

Courant Institute School of Mathematics, Computing, and Data Science, New York University, New York, NY, USA · Open Athena AI Foundation, Inc., New York, NY, USA · Program in Atmospheric and Oceanic Sciences, Princeton University, Princeton, NJ, USA

Neural Surrogate Modeling Climate Science Time Series Forecasting Climate Modeling Computational Fluid Dynamics

Ask PDF

Abstract

Climate Emulation

Explore similar work

ArchesClimate: Probabilistic Decadal Ensemble Generation With Flow Matching

CasEm: A Cascade Architecture for Long-Horizon Neural Emulation

Samudra 2: Scaling Ocean Emulators across Resolutions