geco-bern / rpmodel Goto Github PK

Hi,

Connected to #11, one way you get a mismatch between a_j and a_c is when iabs is zero.

a_j

fact_jmaxlim <- vcmax * (ci + 2 * gammastar)/(kphio * iabs * (ci + kmm))
a_j <- kphio * iabs * (ci - gammastar)/(ci + 2 * gammastar) * fact_jmaxlim

Because fact_jmaxlim has vcmax in the numerator and iabs in the denominator, this gives 0/0 and hence both it and a_j are NaN when iabs is zero.

a_c

In contrast:

a_c <- vcmax * (ci - gammastar)/(ci + kmm)

Here iabs only features in numerator via vcmax and so is 0/real = 0.

In practice, I guess the equation for a_j should be set to zero when there is no photosynthetically active light, and that needs to be trapped to allow the test in #11 to work correctly.

Dear Beni,

According to the original elevation dependence of atmospheric pressure formula proposed by (Berberan-Santos et al., 1997; Equ.40), kT0 better to be set as 288.15K rather than 298.15K, since it better represent the global annual mean surface temperature of the Earth. 288.15K was also originally proposed by Berberan-Santos et al., 1997, in its Equation 40.

To solve this, it is suggested to change kTo <- 288.15K, in patm function at rpmodel/R/subroutines.R.

Where parallel to this issue, also changing kTo <- 288.15K in calc_patm function at ingestr/R/calc_patm.R, since we will rely on this function to download vpd in site-simulations.

Much thanks
Yunke

Sorry about these bugreps! I'm chasing down some numerical warnings in my python implementation that cropped up using real world inputs with more extreme ranges in test inputs.

There is particular behaviour in the equations for density_h20 that create aberrant results. Specifically, temperatures between about -44 and -46 show extreme numerical instability. The code below shows how the mean and sd of the predicted density go haywire for small changes in temperature in this region:

t <- seq(-88, 58, length=1001)
r <- runif(100, -0.5,0.5)
sd_t <- sapply(t, function(xx) sd(density_h2o(xx + r, p=101325)))
plot(sd_t ~ t, type='l')
rho_t <- sapply(t, function(xx) mean(density_h2o(xx + r, p=101325)))
plot(rho_t ~ t, type='l')

This isn't really a flaw in the equation - it is just being extrapolated way outside the intended range of use and happens to have behaviour that causes all sorts of downstream numerical issues.

I think the best fix here is probably a more general constraint on temperature inputs to the pmodel - so that no predictions are made for sub-zero temperatures, given how little the freezing point changes under environmental temperature and pressures.

Hi Beni,

There's an issue with the correction to filter out negative values in calc_ftemp_kphio in version 1.1.2 . The max() function finds the global maximum of the inputs, which of course collapses matrix and array inputs down to a scalar (and gets hung up on missing values and NaN).

> tc <- array(c(NA, -9:28, NaN), dim=c(4,5,2))
> calc_ftemp_kphio(tc)
[1] NA

Substituting pmax(ftemp, 0) preserves the shape of ftemp and calculates maxima for each matrix/array entry.

Hello Beni,
I found a condition where your model crashes. On fluxnet sites, I compute VPD from Qair, press and Temperature.
In dew conditions we can have negative vpd, with qair>qsat.
In calc_optimal_chi you have sqrt(vpd), which leads to NaN values for vpd<0. You can easily correct that by using sqrt(max(vpd,0)).
Cheers,

There is a problem with the test for assimilation == GPP when using C4=TRUE:

> ret <- rpmodel(tc = 20, vpd = 1000, co2 = 400, ppfd = 30, elv = 0, c4=TRUE, fapar=1)
Error in rpmodel(tc = 20, vpd = 1000, co2 = 400, ppfd = 30, elv = 0, c4 = TRUE,  : 
  rpmodel(): Assimilation and GPP are not identical.
> ret <- rpmodel(tc = 20, vpd = 1000, co2 = 400, ppfd = 30, elv = 0, c4=FALSE, fapar=1)
>

The lines are:

    assim <- ifelse(a_j < a_c, a_j, a_c)
    if (any(abs(assim - gpp/c_molmass) > 0.001)) 
        stop("rpmodel(): Assimilation and GPP are not identical.")

I can't immediately see what is wrong - it isn't NA issues as in #11 and the values genuinely are unequal. Having said that, I think all.equal is more robust here as well.

Also - the first line there creating assim: the previous code has just tested that a_j == a_c, so presumably a_j or a_c could be used directly rather than going to the computational expense of combining them?

Hi,

There is something odd going on with the code for setting kphio here:
https://github.com/computationales/rpmodel/blob/master/R/rpmodel.R#L305

  if (do_ftemp_kphio) {
    kphio <- ftemp_kphio( tc, c4 ) * kphio
  } else {
    if (c4){
      kphio <- ftemp_kphio( 15.0, c4 ) * kphio
    }
  }

If the temperature correction is turned off, C4 plants are forced to be reduced by a scaling factor ftemp_kphio( 15.0, c4 ) = 0.2816. That seems like an error?

Using vectors for all variables (temperature, vpd, co2, ppfd, fapar, patm) retuned the wrong number of outputs (something like 2 millions lines even though input vectors had a size of 1460).
A modification was made to line 336 of rpmodel.R . It now reads:
iabs <- fapar * ppfd
I don't know if it would work for all cases.

Unlike other inputs, if a sequence of CO2 values is provided, only the first value is used.

Expected behavior (this is what currently happens with, e.g., a VPD sweep):

> rpmodel( tc = 25, vpd = seq(500, 3000, 500), co2 = 400, fapar = 1, ppfd = 10, elv = 0, do_soilmstress = FALSE, do_ftemp_kphio = FALSE)$gpp
[1] 2.224166 2.115682 2.035845 1.970669 1.914766 1.865385

Actual behavior (when a CO2 sweep is used instead):

> rpmodel( tc = 25, vpd = 500, co2 = seq(350, 400, 10), fapar = 1, ppfd = 10, elv = 0, do_soilmstress = FALSE, do_ftemp_kphio = FALSE)$gpp
[1] 2.027605

Dear Stocker,

Following your 2020 "P-model v1.0: an optimality-based light use efficiency model for simulating ecosystem gross primary production", the deduced Eq. 12 different from yours.
Could you have a look please?

The following is the corresponding MATLAB code:

syms gamma c_i c_s C_i C_a A g_sw beta Gamma R_d Chi g_0 epsilon xi eta K D
% epsilo = 1.6

% K = 2 * Gamma
xi = sqrt( beta * (K + Gamma) / (epsilon * eta ))
Chi = Gamma/C_a + (1 - Gamma/C_a) / xi/(xi + sqrt(D))

C_i = simplify( Chi * C_a )

m = (C_i - Gamma) / (C_i + 2 * Gamma);
m = simplify(m)

The values used here:

https://github.com/computationales/rpmodel/blob/master/R/subroutines.R#L358

ftemp = -0.008 + 0.00375 * tc - 0.58e-4 * tc^2

are incorrect. Shirley has pointed out that the values have been scaled twice in error and that these should be:

ftemp = -0.064 + 0.03 * tc - 0.000464 * tc^2 

Stomatal conductance is currently calculated using the dummy value 1.0 for optimal chi. For calculating g_s, the true internal CO2 partial pressure is needed. This can be calculated for C4 plants using the same equations but a different beta cost, which Colin has suggested should be:

Both Lin et al 2015 and DeKauwe et al 2015 provide estimates for the $g_1$
parameter for C3 and C4 plants, with a ratio of C3/C4 values of around 3. The
$g_1$ parameter is equivalent to $\xi$ in the P model. Given that
$\xi \propto \surd\beta$, a reasonable default for C4 plants is that
$\beta = 146 / 9 \approx 16.222$

So, a new optimal chi function is needed that returns $\chi$ for use in appropriate calculations.

Hi,

There is a bug in soilmstress at the line:
https://github.com/computationales/rpmodel/blob/master/R/subroutines.R#L124

The problem only arises if a user provides soilm as a scalar but meanalpha as an array. That's an odd thing to want to do, but might arise if someone wants to see how the stress factor changes for a given soil moisture in different sites.

In this case - in the nested ifelse statements - the first test 'iterates' over the comparison of the single scalar soilm to x1. If meanalpha is an array, the output is therefore either 1 or the first element only of the no argument:

Browse[2]> ifelse(test = 2 > 4, yes = 1, no = runif(10))                                                                                                                         
[1] 0.5927906

A bit obscure, I admit!

The same issue as in #16 also affect the mechanics of do_ftemp_kphio:

https://github.com/computationales/rpmodel/blob/master/R/rpmodel.R#L287

  kphio <- ifelse(
    do_ftemp_kphio,
    ftemp_kphio( tc, c4 ) * kphio,
    kphio
  )

The argument do_ftemp_kphio is a logical value, so this structure guarantees that kphio will be truncated when do_ftemp_kphio is TRUE:

> kphio <- ifelse(TRUE, c('lots', 'of', 'numbers'), c('some', 'other','numbers'))                                                                             
> kphio                                                                                                                                               
[1] "lots"

I am working on fixes for these bugs - I'll submit some pull requests!

Hi,

This line introduces a problem with NA values:

https://github.com/stineb/rpmodel/blob/master/R/rpmodel.R#L437

In the presence of NAs (but no other problems) the test fails:

> any(c(FALSE, NA))  # returns NA
[1] NA
> if(any(c(FALSE, NA))) { stop('a_j!=a_c')}  # if() statement chokes,
Error in if (any(c(FALSE, NA))) { : missing value where TRUE/FALSE needed

With grid maps as inputs, NAs are pretty likely and are handled by all steps up to this one. One alternative here is all.equal, which handles NAs and allows a tolerance to be set:

> all.equal(c(1,2,3,4, NA), c(1,2,3,4.001, NA), tol=0.001)
[1] TRUE
> all.equal(c(1,2,3,4, NA), c(NA, 2,3,4.001, NA), tol=0.001)
[1] "'is.NA' value mismatch: 2 in current 1 in target"

It would probably be useful to handle $a_j != a_c$ and mismatched NA values separately though? So trap the output of that all.equal and report separately or perhaps use a separate xor(is.na(a_j), is.na(a_c))