2025 Theses Doctoral

Machine learning-based modeling of atmospheric boundary layer processes

Hora, Gurpreet Singh

The atmospheric boundary layer (ABL) plays a crucial role in regulating Earth's surface-atmosphere exchanges, shaping weather patterns and climate dynamics. Although traditional approaches—including experimental investigations, observational studies, and computational modeling—have significantly advanced the understanding of ABL turbulence, each comes with its own inherent limitations. Experimental and observational methods are constrained by data sparsity, while process-resolving simulations are hindered by their prohibitive computational cost. Machine learning (ML) methods have rapidly emerged as a powerful tool for addressing these challenges.

This dissertation explores opportunities offered by ML to (i) reconstruct the three-dimensional (3D) turbulent flows from sparse planar measurements; (ii) develop surrogate models of atmospheric flows in urban areas for efficient multi-query analyses; and (iii) enhance momentum surface-flux parameterizations over the ocean.

To address the data sparsity challenge, a physics-informed ML framework is introduced to reconstruct the 3D structure of wall-bounded turbulence from planar measurements taken at a specified wall-normal location. The method is based on a variational autoencoder approach and is valid for channel flow simulations at moderate Reynolds numbers.

Physics knowledge is embedded through soft constraints in the loss function and hard constraints in the network architecture, enhancing model robustness and integrating inductive biases alongside observational data. The model effectively reconstructs 3D instantaneous flow fields, accurately capturing coherent structures and area-aggregate flow statistics. Additionally, it shows promise in improving upon state-of-the-art reconstruction methods, such as linear stochastic estimation, addressing the data sparsity challenge and offering a potential pathway toward more advanced tools for studying ABL exchange processes.

Next, we propose an ML-based emulator of large-eddy simulation (LES) of ABL flow over idealized canopies for use in multi-query applications such as uncertainty quantification and inverse modeling. Multi-query analyses are a fundamental tool for studying real-world ABL processes due to inherent uncertainties in boundary conditions, initial conditions, and model parameters. However, these analyses require numerous LES evaluations, which are impractical. A multi-layer perceptron (MLP)-based surrogate model is proposed as an alternative to LES, mapping the mean approaching wind angle and canopy geometry to the corresponding flow statistics.

The study focuses on optimizing the MLP architecture and evaluating its performance through the lens of turbulence theory.
Results demonstrate that the MLP surrogate accurately captures airflow spatial variability (such as wake regions and shear layers separating from buildings) and area-aggregate turbulent flow statistics, providing a viable pathway for rapid predictions of turbulence statistics in urban environments.

Expanding beyond land-atmosphere ABL modeling, ML techniques are further applied to enhance the characterization of ocean-atmosphere interactions. Most operational weather and climate models do not resolve complex multi-physics and multi-scale phenomena, such as airflow separation, spray, and bubble generation, which occur at the interface and rely on surface parameterization schemes. A convolutional neural network-based surface flux parameterization is designed to predict spatially distributed surface stresses, leveraging airflow velocity aloft and wave characteristics.

Trained and evaluated on high-resolution laboratory wave profiles and surface viscous stress measurements, the model accurately predicts skin friction drag, capturing key features such as peak viscous stress near wave crests, its rapid decline thereafter, and area-aggregate momentum fluxes over unseen wave conditions.

Although the findings are limited to laboratory experiments, the proposed ML method shows potential for integration as a surface parameterization in LES models, though further refinement is required for applicability to realistic wave fields. All in all, this dissertation highlights the important role that ML can play in ABL research, facilitating the interpretation of measurements and our ability to model ABL processes, spurring advances in the field.