Invitation: PhD Dissertation Defense of Md Badrul Hasan
Hello ME Community,
You are invited to join the PhD Dissertation Defense of Md Badrul Hasan, on Thursday, June 4, beginning at 1:00pm. The defense will be presented in person in the Engineering Building room 210-I (Mechanical Engineering Conference Room).
Advisor: Dr. Meilin Yu
Title: Invariance-Embedded Machine Learning Sub-Grid-Scale Stress Models for Meso-Scale Hurricane Boundary Layer Simulations: Model Development, a priori Assessment, and a posteriori Tests with WRF
Abstract: Accurately simulating the turbulent dynamics of hurricane boundary layers is critical to improving hurricane-intensity forecasts. It is also critical for understanding extreme weather systems. Traditional sub-grid-scale (SGS) models, such as the standard Smagorinsky scheme, cannot represent the bidirectional energy transfer between large-scale flow features and small-scale energetic eddies in extreme weather systems. Contemporary observational and high-resolution modeling studies have shown that backscatter plays a dynamically significant role in hurricane eyewalls and spiral rainbands. This dissertation develops and evaluates an invariance-embedded machine learning (ML) framework for SGS stress modeling to improve the accuracy of large-eddy simulations (LES) of hurricane boundary-layer flows. It integrates physical (pointwise Galilean and rotational) and geometric (spatially relative rotation and reflection) invariant features into a hybrid ML architecture. The architecture couples classification and regression models to predict a signed Smagorinsky coefficient. It captures both the forward energy cascade and backscatter in a single computationally tractable form.
The model development is supported by systematic a priori performance assessment. Multiple ML classifiers (logistic regression, support vector machines, random forest, gradient boosting, and neural networks), together with multiple regression models (Bayesian-regularized and ADAM-trained neural networks), have been evaluated in this work. Ensemble neural networks with embedded invariances outperform the dynamic Smagorinsky reference in a priori tests over the LES eyewall annulus. A complementary direct regression of the deviatoric SGS stress components provides a backup pathway that does not rely on the eddy-viscosity ansatz. It also clarifies the role of optimizer choice in SGS-stress neural-network training.
Building on these a priori results, the dissertation establishes the practical feasibility of deploying the ML SGS framework inside the Weather Research and Forecasting (WRF) model. Two complementary studies are presented. An a posteriori evaluation in idealized adiabatic 2 km tropical-cyclone simulations demonstrates that ML-predicted eddy-viscosity fields can be injected into WRF without compromising numerical stability. The injection produces a stronger vortex than the static Smagorinsky baseline across three different viscosity-floor settings. A separate combined semi-a priori and a posteriori study retains the signed C_s field. It compares ML predictions against offline dynamic Germano-Lilly and effective WRF reference coefficients. It then injects the positive part of the predicted field into WRF in two architectural variants. The dynamic reference exhibits a persistent strain-weighted backscatter fraction of 56% to 65%. The ML predictions preserve the qualitative properties of the dynamic reference at the distributional level. A four-case a posteriori comparison against the baseline, the dynamic reference, the ML full-domain regression, and the architecturally faithful classify-then-regress variant produces a clear and physically interpretable intensity hierarchy. Across both studies, the ML closures consistently outperform the static Smagorinsky baseline by selectively reducing dissipation in regions where the resolved flow does not sustain it, rather than by adding additional mixing. Together, these evaluations bridge data-driven SGS modeling and operational weather prediction. They establish a practical foundation for fully coupled, backscatter-admitting implementations of invariance-embedded ML SGS closures inside mesoscale hurricane simulations.