BEGIN:VCALENDAR VERSION:2.0 PRODID:Linklings LLC BEGIN:VTIMEZONE TZID:Europe/Stockholm X-LIC-LOCATION:Europe/Stockholm BEGIN:DAYLIGHT TZOFFSETFROM:+0100 TZOFFSETTO:+0200 TZNAME:CEST DTSTART:19700308T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:+0200 TZOFFSETTO:+0100 TZNAME:CET DTSTART:19701101T020000 RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260522T162632Z LOCATION:Bldg. 6 - Room 102 DTSTART;TZID=Europe/Stockholm:20260701T113000 DTEND;TZID=Europe/Stockholm:20260701T120000 UID:submissions.pasc-conference.org_PASC26_sess175_pap117@linklings.com SUMMARY:Physics-Aware Multi-Task Learning for Atmospheric Turbulence Param eterization: Auxiliary Tasks versus Architectural Conditioning DESCRIPTION:Sambit Kumar Panda, Todd R. Jones, and Muhammad Shahzad (Unive rsity of Reading); Bryan N. Lawrence (University of Reading, National Cent re for Atmospheric Science); and Anna-Louise Ellis (Met Office)\n\nDynamic subgrid-scale (SGS) turbulence parameterizations in Large Eddy Simulation (LES) achieve superior physical fidelity but impose 2–4× computational ov erhead compared to static schemes, creating a critical bottleneck for high -resolution atmospheric modeling on HPC systems. Neural network based emul ation offers a pathway to comparable accuracy at reduced computational cos t, but realizing this potential requires architectures that generalize rel iably across diverse atmospheric conditions and variable grid configuratio ns.
We systematically compare two physics-aware multi-task learning str ategies for emulating Smagorinsky-based SGS closure in the UK Met Office N ERC Cloud Model (MONC): a baseline approach using Richardson number predic tion 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 thr ee neural architectures
(multi-layer perceptron (MLP), MLP with residua l blocks (ResMLP) and Tabular Transformer (TabTransformer)) trained on mix ed-resolution, multi-regime atmospheric data (66% coarse tropical
conve ction, 34% fine shallow cumulus), we find that uncertainty-based task weig hting consistently outperforms manual tuning and dynamic weighting alterna tives. The simple MLPs with Richardson
conditioning provide the best ro bustness-accuracy trade-off under distribution shift during inference, and the architectural complexity amplifies cross-regime failures despite impr oving in-distribution metrics. Notably, models maintain physical constrain t compliance even when predictive accuracy degrades substantially, suggest ing that the data coverage limitations, rather than any fundamental physic s incompatibility, drive the cross-regime transfer failures.
All result s represent offline validation on static simulation data. Ongoing work foc uses on online MONC integration to assess numerical stability, energy cons ervation, and computational performance under coupled feedback dynamics.\n \nSession Chair: Thorsten Kurth (NVIDIA Inc.)\n\n END:VEVENT END:VCALENDAR