GGG2020-GGG2014DataDescription-R1-20250207.pdf

The Total Carbon Column Observing Network’s

GGG2020 Data Version:

Data Quality, Comparison with GGG2014, and Future Outlook

Debra Wunch

, Joshua Laughner

,

, Geoffrey C. Toon

, Coleen M. Roehl

, Paul O. Wennberg

Luis F. Mill ́an

, Nicholas Deutscher

, Thorsten Warneke

, David F. Pollard

, Dietrich Feist

Kimberly Strong

, Erin McGee

, S ́ebastien Roche

,

, Joseph Mendonca

Rigel Kivi

, Pauli Heikkinen

, Frank Hase

, Mahesh Kumar Sha

, Martine de Mazi`ere

Ralf Sussmann

, Markus Rettinger

, Nasrin Mostafavi Pak

Isamu Morino

, Voltaire Velazco

, David Griffith

Justus Notholt

, Christof Petri

, Matthias Buschmann

, Jonas Hachmeister

Stamatia Doniki

, Damien Weidmann

, Constantina Rousogenous

, Mihalis Vrekoussis

,

Hirofumi Ohyama

, Young-Suk Oh

, Omaira Garc ́ıa

John Robinson

, Manvendra Dubey

, Mingqiang Zhou

,

, Pucai Wang

Yao Te

, Pascal Jeseck

, Laura Iraci

, James Podolske

, Kei Shiomi

, Shuji Kawakami

February 7, 2025

University of Toronto, Toronto, ON, Canada.

California Institute of Technology, Pasadena, CA, USA.

Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA, USA.

School of Earth, Atmospheric and Life Sciences, University of Wollongong, Wollongong, NSW, Australia.

Institute of Environmental Physics (IUP), University of Bremen, Bremen, Germany.

National Institute of Water & Atmospheric Research Ltd (NIWA), Lauder, New Zealand

Institute for Physics of the Atmosphere, German Aerospace Center (DLR), K ̈oln, Germany.

Environmental Defense Fund, New York, NY, USA.

Environment and Climate Change Canada, Downsview, ON, Canada.

Finnish Meteorological Institute, Sodankyl ̈a, Finland.

IMK-ASF, Karlsruhe Institute of Technology, Karlsruhe, Germany.

Royal Belgian Institute for Space Aeronomy, Brussels, Belgium.

Karlsruhe Institute of Technology-Campus Alpin, Garmisch-Partenkirchen, Germany.

Earth System Division, National Institute for Environmental Studies, Tsukuba, Japan.

Deutscher Wetterdienst (DWD), Hohenpeissenberg, Germany.

Space Science and Technology Department, STFC Rutherford Appleton Laboratory, Didcot, Oxfordshire, UK

Climate and Atmosphere Research Centre (CARE-C), The Cyprus Institute, Nicosia, Cyprus

National Institute of Meteorological Sciences, Jeju-do, Korea.

Iza ̃na Atmospheric Research Center (IARC), State Meteorological Agency of Spain (AEMet), Santa Cruz de

Tenerife, Spain

Los Alamos National Laboratory, Los Alamos, NM, USA.

Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China

Laboratoire d’

Etudes du Rayonnement et de la Mati`ere en Astrophysique et Atmosph`eres, Paris, France.

NASA Ames Research Center, Moffett Field, CA, USA.

EORC Japan Aerospace Exploration Agency, Sapporo, Japan.

Abstract

New atmospheric gas retrievals using the GGG2020 version of the Total Carbon Column Observing Network

(TCCON) dataset are significantly improved compared with the previous GGG2014 data version. The GGG2020

retrievals have lower variability across the network, better spectral fits, and smaller biases. The exception to

this is X

4

, which is, on average across the network, 5.3 ppb lower in GGG2020 than in GGG2014, a difference

we mostly attribute to the

a priori

profiles used to perform the scaling. We discuss the remaining biases

and corrections, the TCCON approach of identifying, quantifying, and minimizing biases, and suggest future

work to further minimize the biases. The GGG2020 TCCON data are available from the TCCON archive at

http://tccondata.org

To cite this document, please use:

Wunch, D., et al. The Total Carbon Column Observing Network’s GGG2020 Data Version: Data Quality, Com-

parison with GGG2014, and Future Outlook. CaltechDATA. doi:10.14291/TCCON.GGG2020.DOCUMENTATION.R0,

2025.

1 Introduction

The Total Carbon Column Observing Network (TCCON) provides atmospheric column-averaged dry-air mole

fractions of CO

, CH

, CO, N

O, H

O, HDO, and HF to the scientific and satellite validation communities.

The TCCON is over two decades old: its first dedicated instruments, located at Park Falls, WI, USA and

Lauder, New Zealand, were installed in 2004. The TCCON has since expanded to 29 operational sites; the site

locations are plotted in Figure 1, and the site list with latitude, longitude, and altitude information is in Table

A1. Information about sites that were previously part of the network is given in Table A2.

The network consists of ground-based Fourier transform spectrometers that measure absorption of the direct

solar beam in the near and short wave infrared region of the spectrum. The spectral resolution of the TCCON

spectra is around 0.02 cm

−

(a maximum optical path difference (MOPD) of

∼

45 cm unapodized), which was

selected because it allows the absorption lines of interest to be resolved from the interfering absorptions, telluric

and solar. It also allows investigation of the vertical distribution of the gases from their lineshapes. The solid

angle, Ω, of the detected radiation was selected to satisfy Brault’s criterion [

Brault

, 1985;

Davis et al.

, 2001],

Ω

max

MOPD

π,

(1)

which chooses the Ω that maximizes the modulation depth at MOPD for the highest wavenumber of interest

(

max

). Ω = 2

−

cos

, where

is the angular radius of the field stop. The original optical setup of the

TCCON spectrometers covers a spectral range of 3800–15000 cm

−

with a circular field stop radius of around

= 0

5 mm and

= 418 mm, giving Ω

max

MOPD

= 3

03, which is close to

. The TCCON instrumentation

is described in various general and site-specific publications [

Washenfelder et al.

, 2006;

Deutscher et al.

, 2010;

Wunch et al.

, 2011a;

Kiel et al.

, 2016a;

Kivi and Heikkinen

, 2016;

Pollard et al.

, 2017;

Velazco et al.

, 2017;

Wang

et al.

, 2017;

Yang et al.

, 2020;

Pollard et al.

, 2021;

Weidmann et al.

, 2024], and previous versions of the software

are described in detail in

Wunch et al.

[2011a, 2015].

The TCCON data have been used in scientific investigations of the carbon cycle [e.g.,

Wunch et al.

, 2009, 2013,

2016, 2019;

Sussmann et al.

, 2012;

Sussmann and Rettinger

, 2020;

Keppel-Aleks et al.

, 2012, 2011;

Chevallier

et al.

, 2011;

Guerlet et al.

, 2013;

Deutscher et al.

, 2014;

Saad et al.

, 2016;

Byrne et al.

, 2018, 2020, 2021, 2023,

2024;

Yuan et al.

, 2019;

Babenhauserheide et al.

, 2020;

Labzovskii et al.

, 2021]. They have been used in the

development of improved spectroscopic models and line lists [e.g.,

Thompson et al.

, 2012;

Scheepmaker et al.

2013;

Reuter et al.

, 2011;

Tran et al.

, 2010;

Tran and Hartmann

, 2008;

Reuter et al.

, 2012;

Miller and Wunch

2012;

Long and Hodges

, 2012;

Hartmann et al.

, 2009;

Gordon et al.

, 2011, 2010;

Galli et al.

, 2012;

Mendonca

et al.

, 2016, 2017, 2019]. The TCCON has provided key datasets for the validation of satellite measurements

and satellite algorithm development [e.g.,

Morino et al.

, 2011;

Wunch et al.

, 2011b, 2017;

Butz et al.

, 2011;

Schneising et al.

, 2012;

Schepers et al.

, 2012;

Reuter et al.

, 2013;

Parker et al.

, 2011;

Oshchepkov et al.

, 2013;

Frankenberg et al.

, 2013;

Boesch et al.

, 2013;

Deng et al.

, 2014;

Inoue et al.

, 2016;

Kulawik et al.

, 2016;

Dupuy

et al.

, 2016;

Liang et al.

, 2017;

Hedelius et al.

, 2019;

Wang et al.

, 2020;

Nalli et al.

, 2020;

Hong et al.

, 2021;

Sha

et al.

, 2021;

Karbasi et al.

, 2022;

Taylor et al.

, 2022;

Velazco et al.

, 2019], and in the evaluation of carbon cycle

models [e.g.,

Basu et al.

, 2011;

Houweling et al.

, 2010;

Keppel-Aleks et al.

, 2013;

Mu et al.

, 2011;

Messerschmidt

et al.

, 2013;

Fraser et al.

, 2013;

Kulawik et al.

, 2016;

Ostler et al.

, 2016;

Feng et al.

, 2016;

Stanevich et al.

, 2020;

Bukosa et al.

, 2023]. Attempts have been made to infer vertical information from the TCCON measurements

[

Washenfelder et al.

, 2003;

Saad et al.

, 2014;

Wang et al.

, 2014;

Connor et al.

, 2016;

Roche et al.

, 2021;

Parker

et al.

, 2023].

The GGG2020 version of the TCCON data was publicly released in April 2022. This paper describes the

changes to the TCCON data for the GGG2020 data release relative to the GGG2014 version. The GGG2020

TCCON data and several previous versions (GGG2009, GGG2012, and GGG2014) are archived at the Caltech-

Data Archive (

http://tccondata.org

). Each TCCON dataset from each station has an unique digital object

identifier (DOI) that can be used to cite the data in scientific articles, and the DOI remains static as new data

are added to the time series. If revisions to a dataset are necessary, a new DOI will be minted. The dataset

citations for the GGG2020 version of the TCCON data are listed in Tables A1 and A2, and can also be accessed

using this online tool:

https://tccondata.org/metadata/siteinfo/genbib/

. The network has expanded sub-

stantially since the GGG2014 dataset was first released, adding 4 stations in Asia (Anmyondo, Korea; Hefei,

China; Xianghe, China; and Burgos, Philippines), two in Europe (Nicosia, Cyprus; and Harwell, UK), and 3 in

the Americas (East Trout Lake, Canada; Porto Velho, Brazil; and Calakmul, Mexico). The Porto Velho and

Calakmul instruments have not yet delivered their first datasets to the TCCON archive. A new station is also

being installed at Cambridge Bay, Canada.

The TCCON data processing software, called “GGG”, is centrally maintained at the California Institute of

Figure 1: A map showing the locations of the TCCON stations. The background image is the Blue Marble: Next

Generation, produced by Reto St ̈ockli, NASA Earth Observatory (NASA Goddard Space Flight Center).

Technology’s Jet Propulsion Laboratory and written in a combination of the FORTRAN 77 and FORTRAN 90

programming languages. Wrappers are written in a combination of shell scripting and Python. The software

is open-source with an Apache 2.0 license (

https://www.apache.org/licenses/LICENSE-2.0

), and can be

downloaded from GitHub [

https://github.com/TCCON/GGG

Toon et al.

, 2023]. The GGG2020 version of the

software is described in detail in

Laughner et al.

[2024a]. Each TCCON site uses the same version of the

software, and the processing procedure is consistent from site to site. The first step is to process the raw data

(interferograms) into spectra, using a subroutine called “I2S” (interferogram-to-spectrum).

A priori

profiles of

meteorological parameters (temperature and pressure) and atmospheric composition are generated using a new

algorithm described in detail in

Laughner et al.

[2023]. The spectra are then passed into the main nonlinear least

squares spectral fitting subroutine “GFIT” (gas fit) that iteratively scales the

a priori

atmospheric amounts to

generate forward-modeled spectra that best fit the measured data. The total vertical column amount, or

V C

gas

is defined as the integral of the mole fraction of the gas (

gas

(

)), multiplied by the total number density (

(

)),

from the altitude of the surface (

) to the top of the atmosphere:

V C

gas

∞

gas

(

)

(

)

(2)

The retrieved total column amounts of the gases are in units of molecules

−

and tend to be strongly

influenced by surface pressure (and hence topography). Column-averaged dry-air mole fractions (DMFs; denoted

gas

) are less sensitive to variations in surface pressure and atmospheric water vapour than the retrieved total

column amounts. This characteristic is advantageous for carbon cycle studies because it permits direct compar-

isons of the trace-gas measurements during different seasons, between sites, and with

in situ

measurements. To

calculate DMFs, the total column amount of the gas of interest is divided by the total column amount of dry

air, which we measure using co-retrieved oxygen (O

) multiplied by the dry-air mole fraction of O

, DMF

2

gas

V C

gas

V C

2

DMF

2

(3)

By ratioing the column amounts, systematic errors that are common to

gas

and O

mostly cancel. In GGG2020,

the dry-air mole fraction of O

is assumed to be a constant value of 0.2095. However, this value is known to

be decreasing slowly [e.g.,

Manning and Keeling

, 2006], and so the next version of GGG2020 (GGG2020.1) will

include a time-dependent

DMF

2

The column-averaged amount of dry air (X

luft

‡

) is a special case, and a useful quantity we use to examine

station-to-station biases, as it depends only on the surface pressure measurement (

) and oxygen measurement

(

V C

2

). The X

luft

value should therefore be identical at all sites. It is defined as:

luft

V C

luft

V C

2

DMF

2

(4)

V C

luft

{

}

air

dry

air

(5)

where

is Avogadro’s constant (6.022

molecules mol

−

{

}

air

is the column-averaged gravitational

acceleration, and

dry

air

(28.964 g mol

−

) is the mean molecular mass of dry air. In GGG2014, X

air

was explicitly

corrected for the influence of water by subtracting

2

2

dry

air

from equation 4

‡‡

. This was required because

the GGG2014

a priori

air mole fractions were written as the dry mole fraction despite the surface pressure being

enhanced by the atmospheric water content (equation 5). However, for more self-consistency, in GGG2020, the

a priori

profiles are written as the wet mole fraction, so this explicit water correction is no longer needed. For

an O

measurement with accurate spectroscopy, surface pressure, and H

O retrievals, X

luft

would have a value

of 1.0. Large (

∼

1%) deviations from the network-mean X

luft

at any site indicate serious problems such as an

‡

In GGG2014, this quantity was referred to as X

air

, but that nomenclature has changed to avoid ambiguity in the use of the term

“air” within the software.

‡‡

The parameters

m

2

(18.02 g mol

−

) and

m

dry

air

(28.964 g mol

−

) are the mean molecular masses of water and dry air, respectively.

error in surface pressure, solar zenith angle, various spectral defects caused by laser sampling errors [or “ghosts”,

as described in

Brault

, 1996;

Messerschmidt et al.

, 2010;

Dohe et al.

, 2013;

Wunch et al.

, 2015], poor optical

alignment, or detector nonlinearity.

The dry-air mole fractions are passed through a set of post-processing routines, which include an airmass

dependence correction used to remove any errors that depend on the airmass (e.g., types of spectroscopic errors),

and a bias correction which ties the TCCON data to the currently-accepted World Meteorological Organization

(WMO) trace gas scale through comparisons with WMO-calibrated

in situ

profile measurements obtained from

aircraft or balloons over TCCON sites. Figure 2 shows the number of

in situ

profiles available to us in GGG2014

and GGG2020 to perform the WMO scaling. These corrections are described in detail in

Laughner et al.

[2024a],

as is a detailed error budget.

Figure 2: These plots show the number of aircraft and AirCore profiles available to us for the GGG2014 scaling

and the GGG2020 scaling, per profile in the top row, and per day in the bottom row. The left column is for

, the middle column is for CH

, and the right column is for CO. For GGG2020, the bars are split into the

available profiles and those used to compute the WMO scaling factors for GGG2020. Note that we do not scale

the GGG2020 X

to match the

in situ

scale.

The purpose of this technical report is to describe the TCCON method of identifying, quantifying, and mini-

mizing biases within the TCCON dataset (

2), to describe the differences between the GGG2014 and GGG2020

versions of the TCCON data, and try to explain them in terms of the updates to the GGG algorithm (

3).

2 Evaluating and minimizing TCCON biases

Our approach to minimizing biases within the TCCON is to first understand the causes of the biases, and

subsequently model them using appropriate physical models where possible, filter them out of the public dataset

through a quality control process (

2.1), or reduce them by adjusting the instruments. We have successfully

reduced biases using physical models of laser sampling errors [

Dohe et al.

, 2013;

Wunch et al.

, 2015], nonlinearity

[see

4.1 in

Laughner et al.

, 2024a], and some spectroscopic uncertainties [e.g.,

Mendonca et al.

, 2016, 2017, 2019].

However, we still need to perform an empirical airmass-dependent bias correction, because the airmass

dependence we see in our retrievals is caused by several insufficiently understood errors. In GGG2020, the

airmass dependence correction is much more successful at consistently reducing airmass dependencies at all

TCCON stations than it was in GGG2014 (Fig. 3). The increased network consistency is from improvements

in the H

O spectroscopy itself within the CO

and CH

windows and from the work by

Mendonca et al.

[2016,

2017, 2019] to improve the model of the absorption line shapes within GGG to include speed-dependence and

line mixing effects. Future work will attempt to reduce the residual airmass dependence, eventually to zero.

Figure 3: This figure shows the difference between the residual airmass dependence of retrieved X

(top row)

and X

(bottom row) for select TCCON stations for GGG2014 (left column) and GGG2020 (right column).

These plots were created with data that had airmass dependence corrections applied. In GGG2014, the airmass

dependence correction was insufficient for several sites, particularly those with larger water columns. In GGG2020,

after several spectroscopic improvements in CO

, CH

, H

O, and O

, the residual airmass dependence is substan-

tially smaller and more consistent across the network.

Because the TCCON consists of many instruments, there are biases that are consistent across the network,

which we will call “network biases,” and others that can vary from instrument to instrument, which we will

call “instrument biases.” Examples of network biases in the TCCON data include spectroscopy, absorption line

models,

a priori

profile generation methods, choice of fitting windows, and many other algorithmic choices. To

mitigate the network class of biases, we work toward assessing and minimizing these biases within the retrieval

algorithm itself. Our spectroscopy uses a “greatest hits” approach [

Toon et al.

, 2016] in which we combine

the best available line lists for each window and gas of interest. In addition to the updated line lists, we have

substantially improved our model of the spectral absorption line shapes for CO

, CH

, and O

[e.g.,

Mendonca

et al.

, 2016, 2017, 2019]. We improved the

a priori

profiles for GGG2020 using an age of air climatology, surface

in situ

measurements, and the GEOS-FPIT meteorological model, which is described in detail in

Laughner et al.

[2023, 2024b]. Finally, we run benchmarking tests after changes to the retrieval algorithm to ensure that the

improvements have the intended consequences.

Examples of instrument biases include instrument optical alignment, solar pointing accuracy, surface pressure

measurement accuracy, detector nonlinearity, laser sampling errors, beamsplitter optical flatness, and so on. For

this class of biases, we rely on several diagnostics to identify problems and decide upon the most effective

correction procedure. For example, in GGG2020, we diagnose the severity of detector nonlinearity using a

parameter calculated in the I2S program from the low-pass filtered interferogram near zero path difference

(ZPD) [see

6 and Fig. 6(b) of

Keppel-Aleks et al.

, 2007]. Under nonlinear conditions, this low-pass filtered

interferogram has a small dip near ZPD (see Fig. 4b), and its amplitude is recorded as a parameter we call

“DIP” [see Fig. 4 in this paper, and

4.1 in

Laughner et al.

, 2024a]. If nonlinearity is identified in a set of spectra

(i.e., the rolling median value of “DIP” is larger in magnitude than 5

−

), an offline algorithm provided by

Heikkinen et al.

[2023] is used to compute the coefficients that correct the nonlinearity, and those coefficients are

subsequently embedded into I2S to produce corrected spectra.

Solar pointing and instrument alignment errors are diagnosed using retrieved parameters in GGG2020. For

solar pointing errors, we use the retrieved solar-gas stretch (“S-G”) to identify periods when the solar tracking is

off-centre. However, to actually resolve the problem, mechanical intervention is required; that is, a realignment

of the solar tracker coupling into the spectrometer. GGG2020 currently models the FTS instrument as perfectly

optically aligned, so we diagnose instrument alignment errors (instrument line shape, or ILS) by placing a 10-

cm long glass cell in the solar beam filled with a precisely known amount of HCl gas [roughly

∼

5 hPa,

Hase

et al.

, 2013]. The subsequent retrieval of HCl, given the known cell pressure, provides a diagnostic of the ILS

of the instrument. Like solar tracking, however, a mechanical intervention is required to realign a spectrometer.

The TCCON spectrometers are not permanently aligned, and normal wear of the instrument, or other, more

disruptive events (e.g., earthquakes, relocation, etc.) can degrade instrument optical alignment. In a future

version of GGG, we will be able to model and fit imperfectly aligned spectra. This improvement, currently in

development, uses the formulation provided by

Murty

[1960] and

Kauppinen and Saarinen

[1992], based on the

mathematical description of aberration-free diffraction images by

Linfoot

[1958] among others. In the long term,

(a) X

2

as a function of DIP

(b) Example nonlinear interferogram

Figure 4: The left panel (a) of this figure shows the difference between the retrieved X

with and without a

nonlinearity correction, plotted as a function of the unitless TCCON nonlinearity proxy parameter, “DIP.” These

data were measured in Indianapolis, a site with significant detector nonlinearity. To limit the difference in ∆X

to less than 0.25 ppm, the magnitude of DIP must be limited to less than 5

−

. The right panel (b) shows a

centreburst of an interferogram from Indianapolis in blue and the low-pass filtered version of the same interferogram

in orange. There is a dip in signal near the largest value of the centreburst (located near the interferogram point

labeled 0 to indicate the Zero Path Difference (ZPD) location) indicating that the interferogram suffers from

nonlinearity. An interferogram without nonlinearity would not show a dip near the centreburst.

however, these instruments will always require some mechanical intervention to ensure that they produce high

quality data. This is both an advantage (these instruments can be realigned and their performance improved

when needed) and a disadvantage (these instruments can change alignment over time and have abrupt changes

in performance) of a multi-instrument network such as the TCCON.

Surface pressure measurement errors are diagnosed using a traveling pressure standard instrument (e.g., a

Paroscientific Digiquartz or a Vaisala PTB330), or multiple pressure sensors on site, and comparing the measured

surface pressure to the

a priori

modeled surface pressure over time to identify drifts. Efforts are underway to

make the travel pressure standard method more systematic across the network.

Network and instrument biases need to be treated separately, as they can be conflated or aliased into each

other. For example, an airmass-dependent spectroscopic error that causes a bias between low- and high-latitude

sites could be confused with an instrument alignment difference between the two sites. These two biases should

be resolved differently: the first by adjusting the spectroscopy or applying an airmass dependent correction, and

the second by realigning the spectrometer or modeling the alignment in GGG. Statistical bias corrections, such as

those commonly applied to single satellite instruments [e.g.,

Wunch et al.

, 2011b;

Mendonca et al.

, 2021;

Taylor

et al.

, 2022;

Keely et al.

, 2023], in which large datasets are compared to a proxy for the “truth” and machine

learning or other techniques are used to reduce spurious variability and systematic biases, are less effective for

the TCCON. This is both because of the scarcity of data that can serve as a proxy for the “truth” for a TCCON

bias correction, and because the various instrument biases within the network cannot be resolved with a single

set of statistical corrections.

2.1 Quality control process

The network aims to provide only high quality TCCON data to the public archive (

https://tccondata.org

Of course, the data that are not of high enough quality for the public archive are still valuable to the network

to improve the instrumentation and data processing, so they are stored on an internal, password-protected part

of the archive. With future development of the retrieval code, data that currently do not pass the quality

assurance/quality control (QA/QC) process may also be recoverable and released to the public.

To ensure that the public datasets are of high quality, we have a small team who provides assessments of each

dataset as it is submitted to the archive. There is a standard set of figures created for each new dataset, and

these figures are evaluated by three people within the team who are assigned to each TCCON station, and who

are not involved in the day-to-day operations of that station. Similar to a paper peer review, there is one “editor”

and two “reviewers.” Each reviewer and editor evaluates the datasets on a standard set of criteria, guided by

an online questionnaire, and the editor provides summarized feedback to the station principal investigator (PI).

If everything appears nominal, the data are pushed to the public server, though their release to the public may

be delayed if the PI has chosen to withhold data (allowed for up to 1 year from the time of acquisition). If there

are questions or concerns about the data, the PI is consulted to generate suggestions on how to remedy the

problems, either through changes to the data processing, or mechanical adjustments to the instrument. When

the PI and the quality control team are satisfied that the problem has been understood and resolved, the data

are pushed to the public server. If there is a problem with the data that cannot be resolved (for example, an

instrument alignment issue), the data are flagged and not released to the public archive.

The main criteria are:

1. Pressure sensor error, evaluated using the difference between the surface pressure measurement and the

model surface pressure, and evaluated by examining any long-term trends or abrupt changes.

2. Software errors, identified by searching for incomplete retrievals.

3. Checks that the data removed from the public record as part of the standard, automated filtering process

are indeed substandard. The two dominant filters that remove data from the public record should be the

solar zenith angle of the measurement (which is limited to 82

), and the fractional variation in solar intensity

(which is limited to 5%). Other filters that are triggered should be examined carefully to determine whether

there is a problem with the retrievals or the instrument that can be improved.

4. Timing errors, evaluated as X

luft

variability as a function of solar zenith angle.

5. X

luft

consistency as a function of time.

6. ILS, evaluated through the consistency of HCl scaling factors and root mean squared (RMS) fitting residuals

as a function of time.

7. Metrology laser stability, evaluated through fitted O

frequency shifts.

8. Detector nonlinearity, evaluated using the “DIP” parameter.

9. Solar tracker pointing, evaluated using the solar-gas (“S-G”) stretch parameter.

10. Laser sampling errors or “ghosts”, evaluated using the laser sampling error (LSE) value computed in I2S.

2.2 Biases from the choice of a priori profiles

The scaling factor that ties the TCCON X

gas

value to its WMO trace gas standard scale uses aircraft or AirCore

[

Karion et al.

, 2010] profiles measured concurrently with the TCCON spectra as the

a priori

profiles for the

retrievals. We do this to ensure that the scaling factors are related only to uncertainties in the spectroscopy, and

not caused by biases in the

a priori

profile shape [

Wunch et al.

, 2010;

Laughner et al.

, 2024a]. In a profile scaling

retrieval algorithm such as GFIT, it is only the

shape

of the

a priori

profiles that matters, not the absolute value

of the integrated profile.

To then assess systematic biases caused by using our standard

a priori

profiles [i.e.,

Laughner et al.

, 2023],

we retrieve X

gas

from the spectra that were collected concurrently with the aircraft or AirCore profiles twice:

once using the standard

a priori

profiles, and once using the aircraft or AirCore profiles as the priors. These

situ

profiles do not reach the top of the atmosphere, which, for our purposes is around 70 km, so above the

situ

profile, we attach the GGG2020

a priori

profile. Comparing the scaling factors and their variability from

these two retrievals provides an assessment of the bias that can result from using the standard TCCON priors.

Figure 5 shows the ratio of the retrieved X

gas

to the integrated

in situ

profiles as a function of X

luft

. In

panel (a), X

gas

is retrieved using the aircraft or AirCore profiles as the

a priori

profiles, and in panel (b), X

gas

is retrieved using the standard GGG2020

a priori

profiles. The overall scaling factor is the same well within the

uncertainties for all gases, and the standard deviations of the ratios are comparable, indicating that using the

standard TCCON

a priori

profiles does not induce a significant network-wide bias, nor does it add significant

random noise to the retrievals. These ratios are plotted as a function of X

luft

to provide a sense of how these

ratios change with instrument biases.

2.3 Other bias evaluation techniques

Recently, the development of the EM27/SUN spectrometer has provided another method of identifying site-

to-site biases within the network. The EM27/SUN spectrometer is a Fourier transform spectrometer like the

TCCON spectrometers, but has substantially lower spectral resolution (0.5 cm

−

, or

∼

2 cm MOPD unapodized)

and is smaller and lighter, making it portable [

Gisi et al.

, 2012]. The long-term stability of the EM27/SUN and

(a) Aircraft and AirCore Priors

(b) Standard Priors

Figure 5: This figure shows the ratio between the retrieved X

gas

and the integrated aircraft or AirCore profile for coincident TCCON

measurements and

in situ

profiles. In (a), the retrievals use the

in situ

profiles as

a priori

profiles, and in (b), the retrievals use the standard

GGG2020

a priori

profiles.

its rigorous cross-referencing [

Alberti et al.

, 2022] makes it an attractive option to use as a travel standard for

evaluating site-to-site bias in the TCCON [

Frey et al.

, 2015;

Hedelius et al.

, 2016;

Frey et al.

, 2019]. There have

been several field campaigns testing this technique [

Hedelius et al.

, 2017;

Mostafavi Pak et al.

, 2023;

Herkommer

et al.

, 2024;

Sha et al.

, 2024] and these promising campaigns were able to identify several issues in the TCCON

data that would not have been as easily diagnosed without the travel standard. For example, in

Herkommer et al.

[2024], the authors discovered that some of the beamsplitters within the TCCON spectrometers have reduced

signal in the oxygen spectral region near 8000 cm

−

than others, causing increased noise in the oxygen retrievals.

This apparent reduction in signal is likely caused by the optical flatness of the beamsplitters, which can differ

from instrument to instrument. This issue would have appeared as increased noise in the X

luft

diagnostic in

our quality control process, but the underlying cause would not have been easily discovered without a direct

comparison of the spectra themselves, which was motivated by this field campaign.

3 Differences between GGG2014 and GGG2020

The following sections will describe the differences in the GGG2014 and GGG2020 datasets. The methods we

use to identify these differences are described in

3.1, and the subsequent sections discuss each gas (X

luft

, X

2

4

, X

2

, X

2

, X

HDO

) in turn.

3.1 Methods

To assess differences between the GGG2014 and GGG2020 TCCON datasets, we consider only interferograms

that were processed by both algorithms, and for which the retrievals pass all of the TCCON quality flags. This

is achieved by matching the times the successfully retrieved spectra were recorded to within 20 seconds, to allow

any timing errors

∗∗

[

Sherlock

, 2013] that were assessed and corrected between GGG2014 and GGG2020 to be

included in the analysis. Typically, the time difference between matched spectra is less than 3 seconds. This

matching and filtering process leaves us with data from 31 TCCON stations, and 3.84 million spectra. Because

the GGG2014 results used the X2007 WMO trace gas scale, we use only the X2007

‡‡

version of the GGG2020

retrievals.

In the results that follow, we assess overall differences in the retrieved gases, and differences in the quality of

the retrieved spectral fits. The GGG forward model produces a spectrum that is compared with the measured

∗∗

Timing errors manifest as a systematic variability in X

luft

as a function of solar zenith angle that is caused by recording the

incorrect time for the spectrum. This, in turn causes the ray tracing through the atmosphere to incorrectly model the atmospheric

path for the true time of the measurement. This effect can be caused by an offset or drift in the computer time, or a mis-calculation

of the time change between the start of an interferogram and the time at which zero path difference (ZPD) is crossed. It is the ZPD

time that is most important for the ray tracing calculating.

‡‡

As described in detail in

Hall et al.

[2021], the CO

WMO trace gas standard scale was changed in from the X2007 to the X2019

scale to account for a small error in a coefficient used to calculate CO

from the standard cylinders, and to account for small losses of

during the measurement process. The overall scale difference is 0.045%, which is small, but we have provided both X2007 and

X2019 scalings in our GGG2020 product.

Gas

GGG2014

GGG2020

∆ (GGG2020-GGG2014)

GGG2020 uncertainty

402.31 ppm

402.47 ppm

0.13 ppm (0.03%)

0.47 ppm

1820 ppb

1814 ppb

-5.3 ppb (-0.3%)

3.9 ppb

83.2 ppb

91.0 ppb

6.3 ppb (7%)

1.7 ppb

314.58 ppb

315.99 ppb

1.7 ppb (0.5%)

N/A

60.6 ppt

62.9 ppt

2.5 ppt (4%)

N/A

2065 ppm

2157 ppm

91 ppm (4%)

33 ppm

HDO

1803 ppm

1734 ppm

-67 ppm (-4%)

N/A

Table 1: This table summarizes the network-median values of our retrievals for the GGG2014 and GGG2020

version of the retrieval algorithm. Their differences are calculated as the median value of the distribution of

differences between retrievals from the same spectra for the two different algorithms and might not exactly match

the differences in the network medians. The percent differences are included in brackets in the fourth column.

The fifth column includes the error budget values from Table 5 in

Laughner et al.

[2024a] and indicates the stated

GGG2020 uncertainties. In contrast to GGG2014, the GGG2020 data product for CO does not contain an

in situ

scaling, which explains the large differences for CO.

spectrum. The

a priori

profiles for the target and interfering gases are scaled along with several other fitted

parameters, such as frequency shifts and solar-gas stretches, to minimize the difference between the modeled

and measured spectrum. Once this minimization is achieved, the final scaling factor is recorded. Our metric for

the quality of a spectral fit is the root-mean-squared (RMS) difference between that final modeled spectrum and

the measured spectrum. The RMS value is normalized by the continuum level (CL) of the spectrum to more

directly compare spectral fits between two spectra with significantly different signal levels. This produces the

“RMS/CL” metric we will use in this report to evaluate the spectral fits; a smaller RMS/CL value indicates that

the forward model is better able to model the spectrum.

3.2 Differences

The network median differences are summarized in Table 1, and discussed for each gas individually in the sections

that follow.

3.2.1

luft

Figure 6 shows the difference in the network mean value of X

luft

between GGG2014 and GGG2020. In Fig. 6a,

the distribution of X

luft

values are plotted for each TCCON station, for GGG2014 in the top panel, and GGG2020

in the lower panel. Fig. 6b shows the overall distributions for GGG2014 and GGG2020, and clearly shows the

scale difference between the GGG2014 data, with a network-wide mean value of 0.982, and the GGG2020 data,

with its network-wide mean value of 0.999. The overall spread of X

luft

has reduced in GGG2020 compared with

GGG2014 (Fig. 6b).

In GGG2020, the O

line intensities were scaled to bring the network-mean value of X

luft

much closer to 1

999). The GGG2014 values of X

luft

exhibited a small diurnal variation caused by a temperature bias in the

(a) X

luft

distributions by site and retrieval algorithm

(b) Network-wide distribution of X

luft

distributions by site

(d) Network-wide distribution of ∆X

luft

Figure 6: These figures show the distribution of retrieved X

luft

and the differences between the GGG2020 and

GGG2014 retrieval algorithms. All histograms are normalized such that their maximum value is 1. In figure (a),

luft

distributions are plotted for each site; the top panel shows the results for GGG2014, and the bottom panel

shows the results for GGG2020. The two-letter site identifiers are defined in Tables A1 and A2. In figure (b),

the network-wide distribution of X

luft

is shown for GGG2014 (blue) and GGG2020 (orange). In figure (c), the

distributions of the differences between X

luft

retrieved using GGG2020 and GGG2014 (∆X

luft

) are plotted for each

site (GGG2020–GGG2014). In figure (d), the network-wide distribution of ∆X

luft

is shown.

spectroscopy, which has been reduced by an order of magnitude in GGG2020 [see Fig. 7 of

Laughner et al.

2024a], but has not been eliminated entirely (Fig. 7 of this paper). The small residual temperature dependence

of X

luft

has impacts on the quality control processes we follow, because colder TCCON stations will have X

luft

values that are larger than those from warmer sites, and therefore a static set of X

luft

limits developed using

data from lower latitudes will remove more data from colder locations. This residual temperature dependence is

a current focus of improvements to the spectroscopy (possibly the water broadening) used in the algorithm.