Date: Mon, 01 Oct 2001 10:54:23 -0700 (PDT)
From: Peter Salameh <peter@gaslab.ucsd.edu>
Subject: revised A24 CFC-12 data
To: lebel@ldeo.columbia.edu, dongha@ocean.washington.edu,
mwarner@ocean.washington.edu, bullister@pmel.noaa.gov,
wise@pmel.noaa.gov, rfine@rsmas.miami.edu, bsmeth@ldeo.columbia.edu
Cc: rfw@gaslab.ucsd.edu, ricky@gaslab.ucsd.edu

Hello everyone,

The WOCE A24 CFC-12 data have been re-integrated and water
concentrations have increased by about 3 to 5%.  This problem
resulted from trapping which held more than the typical amount
of N2O, resulting in a large tailing N2O peak with CFC-12
eluting on this tail.

We were able to diagnose this problem by plotting ratios of
CFC-12 concentrations calculated by area to those calculated by
height, and found that the area results were about 3% lower than
the height results.  We next experimented with an exponentially
curved baseline fit, rather than a straight baseline, and found
that areas were increased by about 4% and heights by about 1%.

The integrator had to be modified to allow fitting of
exponentially tailing baselines.  This was made more complicated
by the need to be able to re-calculate baselines for peaks with
integrations which were adjusted manually (graphically).
Another important consideration was that the curve fit not
degrade precision -- and in fact early experiments with
polynomial fits results in much poorer precision.  With the
exponential fit described below, however, the precisions are
equal to or even slightly better than with the straight baseline.

The modifications to the integrator have been completed and seem
to work very well.  The integrator first determines whether the
peak is on an exponential tail -- since we don't want to fit an
exponential to standard gas runs, as their baselines are flat --
and then fits an exponential to a specified amount of baseline
before the CFC-12 peak.

The fit is:

     log(y - y0) = a + b*t

Where vectors y and t are the levels and times of the tailing
part of chromatogram being fitted (before the CFC-12 peak), and
y0 is the "zero level" of the tailing exponential N2O peak,
determined from a minimum point in the chromatogram within some
specified time range after the end of the CFC-12 peak.  Once
fitted by least squares, the resulting exponential baseline is:

    y = y0 + exp(a)*exp(b*t)

Attached are example chromatograms showing a typical CFC-12 peak
integrated with a straight and curved baselines ("straight.ps"
and "curved.ps").

Also attached are two sets of comparison plots -- one set as a
function of CFC-12 concentration in pmol/kg ("compare_conc.ps"),
and the other set as a time series ("compare_time.ps").  These
show four ratios for the CFC-12 water data (where "new" refers
to the new curved baseline fit):

    Area   (new) / Area (old)
    Height (new) / Height (old)
    Conc. by area (new) / Conc. by area (old)
    Conc. by area (new) / Conc. by height (new)

Although the new/old ratios show a trend with concentration --
as we would expect a much larger percentage effect for smaller
peaks -- there does not appear to be a significant trend with
time.  Also the results calculated by area are now nearly
identical to those calculated by height, where before they
differed by 3% (this ratio is a very good diagnostic of
potential biases in the results).

The updated data are attached in file "woce-a24.cfc.curved".  A
final plot has been made by comparing the data in this file by
those in the original file prepared for the CFC consortium
("compare_woce.ps").  The CFC-12 (new)/CFC-12 (old) ratio shows
exactly what we expect.  The second plot in this figure shows
the CFC-11 (new) - CFC-11 (old) DIFFERENCE, which I expected to
be 0.000.  However many CFC-11 values have changed by about
0.001 pmol/kg, a handful of values changed by as much as 0.005
pmol/kg, and a couple of point by about 0.015 pmol/kg (out of 7
pmol/kg).  I am investigating -- possible reasons for the CFC-11
differences are: 1) slight numerical differences between the
LINUX PC now being used versus the old Sun SPARC which was used
to process the original data set, 2) a possible slight change to
the CFC-11 integration parameters between the original set and
the new set, 3) although I strive not to change the fundamental
core of the integrator, a slight change in the algorithm could
also account for these small changes in CFC-11.

This email is rather large with all the attachments (6 in total).
Please let me know if you are able to print all the PostScript
figures, and especially if you are able to read the updated data file.

Regards, Peter