The input and output data provided in GEM-SA DEMO DATA are listed below.
Each subdirectory contains data for a single model. A file of input names is given with model1,
model2 and model3 but most of these are (not too meaningful) code variable names

model1:

Sebastian Lafont ran the forestETP vegetation model at 120 7-dimensional input points.
16 outputs were generated, so I emulated each one separately. The output files are out1.txt ...
out16.txt. Generally, these are all emulated very accurately.

1) annual GPP in mol/m2/an
2) determination coefficient (r2) between simulation and measurement
3) RMSE between simulation and measurement
4) June monthly GPP in mol/m2
5) Sept. monthly GPP in mol/m2
6) maixmale soil evaporation
7) light absorbtion
8) annual soil water content
9) annual ETP in mm
10) determination coefficient (r2) between simulation and measurement
11) RMSE between simulation and measurement
12) June monthly ETP in mm
13) annual soil evaporation
14) annual wet canopy evaporation
15) annual floor evaporation
16) transpiration

Most look like they have additive main effects, so not very interesting.
However, interactions can be seen in

out2, (MO.COVER, MO.TREEHT, MO.LAI)
out3, (MO.COVER, MO.LAI)
out10, (MO.TREEHT, MO.LAI)
out11 (MO.COVER, MO.TREEHT, MO.LAI, TREEHT.LAI)

out11 shows the strongest interaction effects and I think it would be a good example
to demonstrate the following strategy:
------------------------------------------------------------------------------------
1. compute main effects only

2. observe that the sum of main effect indices is significantly less than 100, and also that
there are total effect indices which don't seem to match the corresponding main effect proportion

3. run the analysis again, this time including joint effects for inputs which have the greatest
main effect contribution + those with total effect >> main effect
------------------------------------------------------------------------------------
this strategy captures all the significant joint effects in this case.
[NOTE: we can also get a sum of main effects that's significantly less than 100 if the
emulator is a poor fit, but in this case it is not]








model2:

Mat Disney ran the kuusk model (a light interception model) with 5 inputs. Both 100 and 150 point
datasets are included. Both provide a good emulator fit, although it's clear from the main effects
that the output has some non-stationary behaviour. The model has a clearly non-linear relationship with input 5
(lambda) which is clearly seen from the main effects. There is also an interaction effect
LAI.lambda (inputs 2 and 5) estimated as 6.91% from the dataset of 100 points.



model3:

An early version of SDGVM (sheffield dynamic global vegetation model) with 5 inputs and 70 data
points. The interesting feature here is that input 1 (leaf lifespan) is assigned the maximum
roughness parameter 99.9.. and it turned out this was due to a coding error which produced
discontinuities (something to watch out for!)



venicetoy1:

2 inputs, very boring smooth additive function.
Simple functional form

f(x1, x2) = (30.0+x1*sin(x1))*(4.0+exp(-x2));


venicetoy2:

I use this one in my presentation

f(x1, x2) = (1.0-exp(-1.0/(2*x2)))*
			(2300.0*x1^3 + 1900.0*x1^2 + 2092.0*x1 + 60.0)/
			(100.0*x1^3 + 500.0*x1^2 + 4.0*x1 + 20.0);

and provide 10, 15, 20 data points. There is a steep slope for small values of x1,
which is not properly picked up with the 10 points. Increasing the number of points
provides a much better fit, but the non-stationarity results in CVRMS-standardised
error that is much less than 1.
There is a small interaction effect (around 1 - 2%)



venicetoy3:

boring 2 inputs, 10 points, very smooth

f(x1, x2) = 4.9 + 21.15*x1 - 2.17*x2 - 15.88*x1*x1
		- 1.38*x2*x2 - 5.26*x1*x2
