Role of AI in seed science Predictive modelling and Beyond.pptx
McKinley, Galen: Physical knowledge to improve and extend machine learning pCO2 reconstructions
1. Physical knowledge to improve
and extend machine learning
pCO2 reconstructions
Galen A. McKinley
Valerie Bennington, Lucas Gloege, Amanda Fay
Columbia University, Earth and Environmental Science
Lamont-Doherty Earth Observatory
ICOS Science Conference
September 13, 2022
2. 2
Uncertainties remain significant for the ocean carbon sink.
And observation-based products are temporally limited.
Friedlingstein et al., 2022
sink
sink
Fluxes since 1850 (in GtCO2/yr)
Ocean flux since 1960 (in GtCO2/yr)
3. 3
Limitations of observation-based reconstructions of pCO2
• Few observations in 1980s and none prior
• Limited explain-ability from machine
learning
• Existing knowledge about CO2 dynamics is
often not incorporated into algorithms
Gloege et al., JAMES 2021
µatm
4. 4
Limitations of observation-based reconstructions of pCO2
• Few observations in 1980s and none prior
• Limited explain-ability from machine
learning
• Existing knowledge about CO2 dynamics is
often not incorporated into algorithms
Gloege et al., JAMES 2021
µatm
6. 6
Method #1: pCO2-Residual
• Sea Surface Temperature has a known direct impact on
ocean pCO2
• Biogeochemistry and physical processes drive the
remaining variability
• By removing the temperature component, we focus the
statistics on biogeochemical-physical impacts on pCO2
Takahashi et al., 1993
pCO2-T
Bennington et al. 2022, in review
Mean pCO2
9. 9
Reconstruct pCO2-Residual with XGBoost
pCO2 – T = pCO2
pCO2 - Residual
1. XGB learns pCO2-Residual as
function of features
2. Combine with
pCO2-T for final
result Input data (“Features”)
• Satellite data
− Sea Surface Temp. (SST)
− Chlorophyll-a (Chl-a)
▪ Monthly climatological
− Mixed layer depth
− Sea Surface Salinity (SSS)
▪ Location and time
− Day of year (DOY)
− Latitude, Longitude (n-
vector)
▪ xCO2
Bennington et al. 2022, in review
15. 15
Climatological misfits are much larger than interannual
Princeton Model, others similar Bennington et al. 2022 GRL
16. 16
Since climatological misfit dominates, how much skill is gained by applying
only this as correction, as opposed to an interannual?
• HPDClimTest applies the 2000-2020 climatology of the model-observation misft
1959 2020
1982
LDEO-HPD = Model pCO2 + Interannual Misfit
HPD: Model pCO2 + Climatological Misfit
Observations
Begin
Model Period
Begins
HPDClimTest = Model pCO2 +
Climatological Misfit
Bennington et al. 2022 GRL
17. 17
Most improvement over original models is climatological
Comparison data (1990-2020) not
used in algorithm training: GLODAP
and LDEO pCO2 (not in SOCAT) Bennington et al. 2022 GRL
18. 18
Most improvement over original models is climatological
Comparison data (1990-2020) not
used in algorithm training: GLODAP
and LDEO pCO2 (not in SOCAT) Bennington et al. 2022 GRL
19. 19
Most improvement over original models is climatological
Comparison data (1990-2020) not
used in algorithm training: GLODAP
and LDEO pCO2 (not in SOCAT) Bennington et al. 2022 GRL
20. 20
Most improvement over original models is climatological
Comparison data (1990-2020) not
used in algorithm training: GLODAP
and LDEO pCO2 (not in SOCAT) Bennington et al. 2022 GRL
21. 21
Apply climatological misfit to extend back to 1959
• HPDClimTest applies the 2000-2020 climatology of the model-observation misfit
• Since it dominates, apply this in the pre-observed period
1959 2020
1982
LDEO-HPD = Model pCO2 + Interannual Misfit
LDEO-HPD = Model pCO2 +
Climatological Misfit
Observations
Begin
Model Period
Begins
HPDClimTest = Model pCO2 +
Climatological Misfit
Bennington et al. 2022 GRL
23. 23
Conclusions
• Physical knowledge can be incorporated into machine learning algorithms, and
leads to improved reconstruction skill
• pCO2-Residual
• Focuses the statistics on the biogeochemical-physical component of pCO2
• LDEO-HPD
• Uses suite of hindcast ocean models as a prior, corrects with SOCAT data
• Climatological correction most impactful; supporting extension back to 1959