Objective assessment of five processoriented dynamic acidification models at
Hubbard Brook
Koji Tominaga, Julian Aherne, and Shaun Watmough
Trent University, Ontario, Canada
COLLABORATORS
Mattias Alveteg: Lund University, Sweden
Jack Cosby: University of Virginia
Charlie Driscoll [Jana Keikbusch]: Syracuse University
Thorjorn Larssen: Norwegian Institute for Water Research, Norway
Max Posch: Co-ordination Centre for Effects, The Netherlands
IES HB COS MEETING, MILLBROOK, NY, USA
14 APRIL 2008
Objectives
•! Bring more
confidence to
the models to
support policy
•! Develop a
multiple-model
simulation
framework
!!Data-model
fusion techniques
(such as MCMC)
are central in
meeting these
objectives
!!Hubbard Brook
provides excellent
data sets
Three approaches
1.! “Level-playing field”
approach
–!
Monte Carlo, input mapping,
input uncertainty (following
Rose et al.) [data " model
structure]
2.! “Best-fit” approach
–!
model-independent
parameter estimation, GaussMarquardt-Levenberg, MHMCMC, misfit measures,
equifinality space [model "
parameter space]
Approach 1
Data
Model
Approach 2
Approach 3
3.! Multiple-model approach
–!
forecasting uncertainty; use
of multiple models provide
more informative predictions
Multiple-model
simulation framework
Models
•! PnET-BGC
•! MAGIC
•! SAFE
•! SMART
•! VSD
Credits
•! North-eastern States Research Cooperative
•! Hubbard Brook Experimental Forests (Dr. Scott
Bailey, US Forest Service)
•! Shared Hierarchical Academic Computing
Network, Ontario
Thank you very much for
listening (and comments)!
Extra slides for questions…
955
54.5
20
0.432
1
4.5
Precisely
known inputs
5
15
-2
Imprecisely
known inputs
5 10
Give them many
possible numbers
Predicted variable
of interest
The model
**The result is an accumulation
of many possible cases
2
-20
10
25
Input mapping
•! Ensures equivalent model inputs for all models
•! A set of formulae:
–!‘coarse’ the most detailed (chemically and
spaciotemporally) data
–!prepares model input files using the same original
data
Input mapping
algorithm
Fundamental
inputs
Required inputs
for Model A
Required inputs
for Model B
Required inputs
for Model C
(after Rose et
al. 1991b)
Run MC on
selection of
fundamental
inputs
Iterations
(i.e., sets of
fundamental
inputs)
Fundamental
inputs
-!Watershed
characteristics
-!Process parameters
-!Driving variables
Input mapping
Models
MAGIC
inputs
PnET-BGC
inputs
VSD
inputs
SMART
inputs
SAFE
inputs
MAGIC
PnET-BGC
VSD
SMART
SAFE
MAGIC
pred’ns
PnET-BGC
pred’ns
VSD
pred’ns
SMART
pred’ns
SAFE
pred’ns
MAGIC
pred’ns
PnET-BGC
pred’ns
C
SMART
pred’ns
SAFE
pred’ns
D
B
E
A
-50
VSD
pred’ns
0
50
100
ANC in 2050 [µeq/L]
(after Rose et
al. 1991b)
Selected
variables
to compare
Graphing
and
statistical
analysis
e.g., base saturation
and ANC
PEST
Parameter Estimation software
Input
data
MAGIC
Imprecisely
Known parameters
Calibration
over
Output
data
Not good
Good
Important consideration: Predefine quantitatively what is
considered well-fit
Reproduced
The past?
PEST
?
Dissolved organic carbon
Behavioural spaces
B
A
C
D
E
Weathering rates
Some chemical measure
Experiment with the models
Observation
1970
1980
Some chemical measure
Experiment with the models
Model
1970
1980
Some chemical measure
Experiment with the models
Prediction
1970
1980
2005
Some chemical measure
Experiment with the models
What has
actually
happened
The model did
its best here
1970
1980
The real
examination for the
model accuracy
2005
Some chemical measure
Experiment with the models
All models do
well here
1970
1980
Which model was
most accurate?
2005