Prof. Brian Powell is a Physical Oceanographer in
the Department of Oceanography at the University of Hawai‘i at
Mānoa and a Senior Fellow at the Cooperative Institute for Marine and
Atmospheric Research. His research focuses on the understanding and
prediction of the ocean environment, including physics and
biogeochemistry, applications of machine learning in state estimation
and prediction, and ocean observations and their design. He was the
recipient of the 2009 ONR Young Investigator award for his work on
prediction and state estimation. His lab welcomes people from any
background and discipline interested in the ocean and its climate.
Ideas within the lab group are shared and debated openly with all
discourse free of rigid hierarchies.
Physics and Ecosystems
Work in the Powell lab is aimed to understand the physics and
feedbacks that control the ocean, its geochemical cycle, and the
ecosystems that form in each niche. These physical processes
determine the health of everything from coral reefs to nearshore and
pelagic fish stocks that feed humans.
Over the past thirty years, the variability of extreme events has
likely increased (prolonged marine heat waves, potentially stronger
El Niño impacts, etc.); however, the climate itself has not
yet significantly shifted. Two crucial questions related to climate
change are: (i) how will interannual–to–decadal
variability change over the coming decades; (ii) and, when will
climate change emerge as the dominant signal?
As the variability in the physics increase towards a new climate
change signal, the impacts on the biogeochemistry and ecosystems may
be profound. These changes will be felt from the single–cell
planktons to the dinner–tables of families and reach from
science to economics to social convention.
Simulations and Machine Learning
To understand and forecast the ocean and its ecosystems, we rely
upon large–scale, high–performance computing. We develop
software tools for simulating the ocean and its environment, for
data analysis, and for combining observations and models in
mathematically consistent ways so that the models represent the
observations of the ocean.
Machine Learning refers to a broad range of methods that constrain a
model with observed behavior. From simple linear–regression to
deep neural–networks, there exist many ways to combine models
and data. Less than 1% of the ocean is observed, which means that
traditional, statistics–based machine learning is an
inappropriate tool: a naïve neural-network lacks the
observations to "learn" about the ocean. Physics-informed methods,
such as 4D-Variational Data Assimilation, Ensemble Kalman Filters,
and Physics-Informed Neural Networks (PINN) are the methods that are
required.
Our lab group strives to continue to be at the forefront of these
techniques and utilizing them to blur any boundaries between models
and observations. We look to quantify what observations actually
"see" in the ocean, which observations improve our forecasting and
understanding, and how machine–learning can supplement or
replace our estimations of small–scale processes in the ocean.
