Physics-aware multi-task learning for atmospheric turbulence parameterization: auxiliary tasks versus architectural conditioning

Panda, Sambit Kumar; Jones, Todd R.; Shahzad, Muhammad; Lawrence, Bryan N.; Ellis, Anna-Louise

Download

[thumbnail of pasc_2026_sambit_camera_ready_v1.pdf]

Text
- Accepted Version
· Restricted to Repository staff only
· The Copyright of this document has not been checked yet. This may affect its availability.

Advice

Please see our End User Agreement.

It is advisable to refer to the publisher's version if you intend to cite from this work. See Guidance on citing.

Tools

Lists

Panda, S. K., Jones, T. R. ORCID: https://orcid.org/0000-0002-7669-1499, Shahzad, M. ORCID: https://orcid.org/0009-0002-9394-343X, Lawrence, B. N. ORCID: https://orcid.org/0000-0001-9262-7860 and Ellis, A.-L. (2026) Physics-aware multi-task learning for atmospheric turbulence parameterization: auxiliary tasks versus architectural conditioning. In: PASC26 Conference, 29 Jun - 1 Jul 2026, Bern, Switzerland. doi: 10.1145/3815572.3815750 (In Press)

Abstract/Summary

Dynamic subgrid-scale (SGS) turbulence parameterizations in Large Eddy Simulation (LES) achieve superior physical fidelity but impose 2–4× computational overhead compared to static schemes, creating a critical bottleneck for high-resolution atmospheric modeling on HPC systems. Neural network based emulation offers a pathway to comparable accuracy at reduced computational cost, but realizing this potential requires architectures that generalize reliably across diverse atmospheric conditions and variable grid configurations. We systematically compare two physics-aware multi-task learning strategies for emulating Smagorinsky-based SGS closure in the UK Met Office NERC Cloud Model (MONC): a baseline approach using Richardson number prediction as auxiliary gradient regularization, and an Ri-conditioned approach that explicitly feeds predicted stability into coefficient (viscosity and diffusion) prediction heads. Evaluating 54 model configurations across three neural architectures (multi-layer perceptron (MLP), MLP with residual blocks (ResMLP) and Tabular Transformer (TabTransformer)) trained on mixed-resolution, multi-regime atmospheric data (66% coarse tropical convection, 34% fine shallow cumulus), we find that uncertainty-based task weighting consistently outperforms manual tuning and dynamic weighting alternatives. The simple MLPs with Richardson conditioning provide the best robustness-accuracy trade-off under distribution shift during inference, and the architectural complexity amplifies cross-regime failures despite improving in-distribution metrics. Notably, models maintain physical constraint compliance even when predictive accuracy degrades substantially, suggesting that the data coverage limitations, rather than any fundamental physics incompatibility, drive the cross-regime transfer failures. All results represent offline validation on static simulation data. Ongoing work focuses on online MONC integration to assess numerical stability, energy conservation, and computational performance under coupled feedback dynamics.

Altmetric Badge

Dimensions Badge

Item Type	Conference or Workshop Item (Paper)
URI	https://centaur.reading.ac.uk/id/eprint/129861
Identification Number/DOI	10.1145/3815572.3815750
Refereed	Yes
Divisions	Science > School of Mathematical, Physical and Computational Sciences > Department of Computer Science Science > School of Mathematical, Physical and Computational Sciences > NCAS Science > School of Mathematical, Physical and Computational Sciences > Department of Meteorology
Uncontrolled Keywords	Large Eddy Simulation, Machine Learning, Turbulence Parameterization, Multi-Task Learning, Atmospheric Physics
Download/View statistics	View download statistics for this item

Funded Project

Related URLs

Deposit Details

References

[1] A. R. Brown, S. Belcher, S. Milton, P. J. Wynn, and A. Hill. The Met Office NERC Cloud Model (MONC). Bull. Amer. Meteor. Soc., 101(11):E1752–E1767, 2020. [2] A. R. Brown, S. J. Abel, S. Belcher, N. Bloom, and P. J. Wynn. Leveraging modern computing for atmospheric simulation. In Proc. PASC, pages 1–8, 2018. [3] N. Brown et al. A highly scalable Met Office NERC Cloud model. In Proc. EASC, pages 132–137, 2015. [4] S. B. Pope. Turbulent Flows. Cambridge University Press, 2000. [5] J. C. Wyngaard. Atmospheric turbulence. Annu. Rev. Fluid Mech., 24:205–233, 1992. [6] D. K. Lilly. The representation of small-scale turbulence in numerical simulation experiments. In IBM Scientific Computing Symposium on Environmental Sciences, pages 195–210, 1967. [7] A. Pal. Deep learning emulation of subgrid-scale processes in turbulent flows. Geophys. Res. Lett., 47:e2020GL087005, 2020. [8] D. Kochkov et al. Machine learning-accelerated computational fluid dynamics. PNAS, 118(21):e2101784118, 2021. [9] S. Rasp, M. S. Pritchard, and P. Gentine. Deep learning to represent subgrid processes in climate models. PNAS, 115(39):9684–9689, 2018. [10] H. Mandler and B. Weigand. Generalization Limits of Data-Driven Turbulence Models. Flow Turb. Combust., 2024. [11] R. Molinaro et al. Generative AI for fast and accurate statistical computation of fluids. arXiv preprint arXiv:2409.18359, 2024. [12] M. Sakarvadia et al. The False Promise of Zero-Shot Super-Resolution in MachineLearned Operators. arXiv preprint arXiv:2510.06646, 2025. [13] H. Heuer et al. Beyond the Training Data: Confidence-Guided Mixing of Parameterizations in a Hybrid AI-Climate Model. arXiv preprint arXiv:2510.08107, 2025. [14] G. E. Karniadakis et al. Physics-informed machine learning. Nat. Rev. Phys., 3:422–440, 2021. [15] T. Beucler, M. Pritchard, S. Rasp, J. Ott, P. Baldi, and P. Gentine. Enforcing analytic constraints in neural networks emulating physical systems. Phys. Rev. Lett., 126(9):098302, 2021. [16] J. Yuval and P. A. O’Gorman. Stable machine-learning parameterization of subgrid processes for climate models with conservation laws. Nat. Commun., 11(1):3295, 2020. [17] A. Kendall, Y. Gal, and R. Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482–7491, 2018. [18] S. Liu, E. Johns, and A. J. Davison. End-to-end multi-task learning with attention. In CVPR, pages 1871–1880, 2019. [19] G. Bertoli, S. Mohebi, F. Ozdemir, J. Jucker, S. Rüdisühli, F. Perez-Cruz, M. Salzmann, and S. Schemm. Revisiting machine learning approaches for short- and longwave radiation inference in weather and climate models. J. Adv. Model. Earth Syst., 17(9):e2025MS004956, 2025.

CORE (COnnecting REpositories)

University Staff: Request a correction | Centaur Editors: Update this record

Funders:	University of Reading	The sponsoring bodies who contributed funding for the creation of this item. Example: NERC Example: The Royal Society of Chemistry A pick list of funders may appear as you type in the funder's name in full or as an acronym. Select a correct match to complete the field or type in a new entry in full. For new entries, the full name is preferred.
Projects:	Advancing the Frontiers of Earth Systems Prediction (AFESP) Funded by: University of Reading (N/A - ) UNSPECIFIED - UNSPECIFIED	Click Add to select your project (received at Reading) from an autocomplete list.

Date Deposited:	26 May 2026 11:05	Date item deposited into CentAUR
Last Modified:	26 May 2026 11:15	Date item last modified