Science QuestionsHypotheses

1
Science Questions/Hypotheses

• Determining of the decoupling between nutrient
  supply and carbon export

2
What do marine biogeochemists need to know about
biology?

• Biology determines the timescale with which
  inorganic nutrients and carbon are transformed
  into organic material
• What determines an HNLC versus HC areas?
• There is a temporal decoupling between organic
  matter formation and its removal from the
  surface
• How long does it take to create sinking material?
• Should Iron Experiments have expected to see
  export?
• Biology determines the penetration of export
  through the surface ocean
• DOM vs POM, interactions with mineral, particle
  lability, particle sinking speeds (size and
  density)
• Elemental stoichiometry is variable
• N2-fixation, denitrification, carbon
  overconsumption, flexible physiology.

3
How can we move from a phenomenological model to
a truly mechanistic/predictive model?

• Jerrys Goal - There are a few things we think
  that we know, and one way to make projress is to
  leverage that information through the adjoint to
  learn about the multitude of things that we know
  nothing about

4
We can use the framework to test the best model
construction

• Model complexity improves portability
• Only simple models can be fully constrained
• In the simplicity extreme Model zooplankton
  cannot be well-constrained from existing
  observations
• Zooplankton dynamics are important
• Predator-Prey oscillations exist in nature
• Beyond zooplankton, we need to model bacteria,
  viruses, aggregation, and higher order
  top-down-control
• It is necessary to distinguish between
  phytoplankton functional groups
• It is necessary to distinguish between
  phytoplankton size classes
• It is necessity to include iron
• A single form of irradiance and co-nutrient
  limitation is more suitable than others
  (photo-adaptation, photo-inhibition, variable
  ChlN, optics/depth spectra)

5
The adjoint between-site comparison provides an
objective framework for establishing the
robustness of models

• Cost_AS Cost_EqPac ltlt Cost_ASEqPac

6
The empirical and mean models provide a means to
test whether our biological prejudice helps the
simulation

• Which model gets the mean?
• Which model gets the best data variance?

7
The testbed provides a quantitative framework to
understand the relative role of supply (i.e.
bio-optics, phytoplankton physiology) and loss
(i.e. ecology)

• Are initial results suggest that the adjoint
  allows us to control supply rather than loss is
  that because supply is the most important, or
  because the experiment is ill-concieved this
  will be tested partially with the cross-site
  robustness.
• Currently we are trying to parameterize the most
  certain parameters
• Perhaps we should try the opposite (as Jerry was
  trying), and fix the certain parameters and try
  to fit to the least certain ones.
• Should we attempt the optimization of only growth
  parameters versus only loss parameters?
• Assimilating fluxes only may give a very
  different priority to the optimization (grazing
  vs growth)
• Other experiments?

8
The timescales of ecosystem interactions provide
important constraints on the biogeochemical
imprint

• The adjoint can be used to estimate these rate
  constraint
• Growth and grazing rates
• Extent of decouplign between growth and grazing
• Evolution timescale of bloom to provide timing of
  export
• Sinking velocity to depth

9
What are the most important data constraints?

• Assimilating rates and fluxes versus standing
  stocks which are more important to model
  fidelity?
• Are export fluxes redundant in adjoint
  simulations forced by the nutrient concentrations
  and upwelling velocities?

10
Can we discern whether something non-resolved is
important?

• Are there some ways in which all of the models
  fail beyond our lack of understanding beyond the
  physics?
• Are we all compensating for the same model or
  data deficiency in the same way
• Is mass non-conservation apparent?

11
What data sets can we ask observationalists to
measure?

• What do we need to make the models robust?
• To what extent are the models failing through the
  mismatch by over-fitting the models to poor data?
• To what extent are we underestimating the power
  of complex models because we lack the data to
  constrain them?
• To what extent can we get around that through
  applying external, a priori contraints?