| dc.description.abstract |
<p>Sorted bedforms are spatial extensive (100 m-km) features present on many inner
continental shelves with subtle bathymetric relief (cm-m) and localized, abrupt variations
in grain size (fine sand to coarse sand/gravel). Sorted bedforms provide nursery habitat
for fish, are a control on benthic biodiversity, function as sediment reservoirs,
and influence nearshore waves and currents. Research suggests these bedforms are
a consequence of a sediment sorting feedback as opposed to the more common flow-bathymetry
interaction. This dissertation addresses three topics related to sorted bedforms:
1) Modeling the long-term evolution of bedform patterns, 2) Refinement of morphological
and sediment transport relations used in the sorted bedform model with `machine learning';
3) Development of a new sorted bedform model using these new `data-driven' components.</p><p> Chapter
1 focuses on modeling the long term evolution of sorted bedforms. A range of sorted
bedform model behaviors is possible in the long term, from pattern persistence to
spatial-temporal intermittency. Vertical sorting (a result of pattern maturation processes)
causes the burial of coarse material until a critical state of seabed coarseness is
reached. This critical state causes a local cessation of the sorting feedback, leading
to a self-organized spatially intermittent pattern, a hallmark of observed sorted
bedforms. Various patterns emerge when numerical experiments include erosion, deposition,
and storm events. </p><p> Modeling of sorted bedforms relies on the parameterization
of processes that lack deterministic descriptions. When large datasets exist, machine
learning (optimization tools from computer science) can be used to develop parameterizations
directly from data. Using genetic programming (a machine learning technique) and large
multisetting datasets I develop smooth, physically meaningful predictors for ripple
morphology (wavelength, height, and steepness; Chapter 2) and near bed suspended sediment
reference concentration under unbroken waves (Chapter 3). The new predictors perform
better than existing empirical formulations. </p><p> In Chapter 3, the new components
derived from machine learning are integrated into the sorted bedform model to create
a `hybrid' model: a novel way to incorporate observational data into a numerical model.
Results suggest that the new hybrid model is able to capture dynamics absent from
previous models, specifically, the two observed end-member pattern modes of sorted
bedforms (i.e., coarse material on updrift bedform flanks or coarse material in bedform
troughs). However, caveats exist when data driven components do not have parity with
traditional theoretical components of morphodynamic models, and I address the challenges
of integrating these disparate pieces and the future of this type of `hybrid' modeling.</p>
|
|
