YIP Closing the loop on joint physics- and data-driven modeling of marine boundary layer turbulence above waves
Abstract
Marine meteorological predictions are critical to guide real-time marine vessel and aviation decisionsin the ocean battlespace and for long-term Naval base resilience planning. Weather modelshave spatial resolutions of 0(1 km) to enable fast predictions, but there is wide continuum of importantspatial and temporal scales in the atmosphere. Small-scale interactions between windsand the wavy surface dictate exchanges of momentum, heat, and gases and marine boundary layercharacter. These processes are only represented in weather models through surface and boundarylayer turbulence parameterizations that have been independently developed and tuned using limiteddata. Model biases and uncertainties are problematic for regimes outside the scope of dataused for development and tuning, suchas new locations, conditions, or extreme events.The rapidly increasing availability of high-fidelity data that capture wind-wave interactionsfrom high resolution simulations and field observations has led toa paradigm shift towards the developmentof fast data-driven models. But leveraging such data to learn about uncertainty (inverseuncertainty quantification) in widely used weather models has been held back by the very highcomputational effort required. There is an urgent need to advance marine meteorological predictionsfor real-time Naval decision-making by lowering error and quantifying uncertainty, but thestate-of-the-art in quantifying marine meteorological modeling uncertainty is ensemble modelingthat depends on ad hoc choices in setup and does not leverage newly available data.This project builds foundational knowledge of the interactions between boundary layer turbulenceand surface parameterizations inmarine flows with waves, directly quantifies modelinguncertainties, and identifies observation locations that maximally reduce these uncertainties. Wewill use a synergistic combination of gradient-free optimization and machine learning to accelerateinverse uncertainty quantification by a thousand-fold, enabling its use on meteorological modelsfor the first time. The objectives are to use large eddy simulations across wind and wave regimes,considering methods that either average over or resolve the wave phase, along with mesoscale(weather) simulations to understand interactions between the surface and boundary layer turbulenceparameterizations. Data from phase-resolving large eddy simulations will be used to performinverse uncertainty quantification to estimate the joint probability distribution of mesoscalesurface and turbulence model parameters and model-form error for the first-time, enabled by machinelearning. An extremely efficient Bayesian experimental design method developed here willidentify sensing locations and time periods that maximally reduce modeling uncertainty.This project s outcome closes-the-loop on marine meteorological modeling by leveraging wavephase resolving simulations to quantify surface and boundary layer turbulence parameterizationuncertainty in fast-running operational models. An additional outcome quantifies which windand wave regimes lead to statistically significant differences between large eddy simulation andmesoscale weather models and will learn about interactions between surface and boundary layerparameterizations along with data-driven model augmentation terms to lower error. The uncertaintyquantification approach will be validated using field observations, and is naturally extensibleto investigate other physical processes or to quantify uncertainty in other numerical models.This research produces physics-informed data-driven modeling of wind-wave interactions startingfrom a physics-based model. This project provides physically-quantified confidence intervals,representing uncertainty from unresolved processes, on state-of-the-art predictions of marineboundary layer flows to enable decision making under uncertainty in the Ocean Battlespace.Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 14, 2024
- Source ID
- N000142512045
Entities
People
- Michael F Howland
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy