Post-doc: Inversion of greenhouse gas emissions at urban scale : new algorithms

Climate change and air pollution are two of the biggest challenges that the world is facing in the 21st century. Anthropogenic air pollutants and greenhouse gases share common sources but specific actions for reducing one can often lead to increase the other. Therefore, an optimal reduction of both air pollution and climate change impact can only be achieved through coordinated monitoring and mitigation strategies. In this context, the French-funded ANR ARGONAUT project (PollutAnts and gReenhouse Gases emissiOns moNitoring from spAce at high ResolUTion) will apply atmospheric inversions to estimate the French anthropogenic emissions of the main gaseous pollutants (nitrogen oxides - NOx - carbon monoxide - CO - and non-methane volatile organic compounds - NMVOCs) and of carbon dioxide (CO2). This inversion will rely on the last generation of satellite imaging within the European Copernicus program: Sentinel-5P/TROPOMI (Borsdorff et al., 2019; Lama et al., 2019) and CO2M (Kulhmann et al., 2019).

The case of local, typically urban, scale inversion raises many challenges. The inversion is based on a metric dedicated to the comparison of satellite images and concentration fields from simulation. Most of the metrics currently used in inverse problems are based on local (i.e. point-wise) comparisons. Such indicators are easy to compute but they are facing severe limitations in the comparison of plumes of pollutants. They are subject to the so-called double penalty effect, the missing data (for example due to cloud coverage) are completely ignored, and there is no theoretical treatment of representation error (Farchi et al., 2016).

Several strategies already exist in the statistical literature to perform verification of non-local nature, for example the Wasserstein distance, based on the optimal transport (Villani, 2009). In more than one dimension, solving the optimal transport is non-trivial and computationally expensive. Recently, new computing strategies have emerged, such as regularised optimal transport (Cuturi, 2013; Ferradans et al., 2013), which make the problem tractable. Furthermore, the link between optimal transport and machine learning is more and more put forward (Peyré and Cututi, 2019). In this context, tools from the machine/deep learning community could be used to solve the optimal transport problem, or even to learn the Wasserstein distance.

Working environment

The post-doctoral fellow will join the data assimilation team of CEREA, joint laboratory École des Ponts ParisTech and EDF R&D, located in Champs-sur-Marne (20 min from Paris with RER A). The work will be supervised by Marc Bocquet and Alban Farchi (data assimilation / inverse modelling). Within CEREA, this work will benefit from the help of Yelva Roustan (atmospheric chemistry) and Youngseob Kim (research engineer / technical support). Furthermore, the post-doc will be in continuous contact with the other partners of the ARGONAUT project (LISA, LSCE, and INERIS) which are all located in the Paris area, and especially Grégoire Broquet at LSCE.

Scientific objectives

His or her work will be a pillar of work package 5 of the ANR ARGONAUT project. Three tasks have been identified:

• Select a database of test cases for the study, based on satellite images of NO2, CO, and HCHO in the Paris/Benelux region, and corresponding simulations obtained from the CHIMERE chemistry and transport model (Fortems-Cheiney et al., 2019). Possibly build a large database of synthetic test cases using a simpler chemistry and transport model (Mallet et al., 2007).

• Define, select, and implement a metric dedicated to comparing satellite images and species concentration fields from simulation. As explained, the metric should be able to perform verification of non-local nature, while being computationally tractable for an inversion system. The Wasserstein distance is a good candidate for such a metric. Taking inspiration from the recent developments related to optimal transport, the goal is to develop innovative tools to efficiently compute the Wasserstein distance in the test cases identified in the first task. For this purpose, the relationship between optimal transport and machine learning should be further studied to determine (i) if machine learning tools can be used to compute the Wasserstein distance and (ii) if the Wasserstein distance can be learnt, bypassed or improved through machine learning algorithms.

• Investigate to what extent it is possible to include the relevant metrics in an inverse modeling system.

References

• A. Farchi, M. Bocquet, Y. Roustan, A. Mathieu, and A. Quérel. Using the Wasserstein distance to compare fields of pollutants: application to the radionuclide atmospheric dispersion of the Fukushima-Daiichi accident. Tellus B, 68:31682, 2016.
• Varon, D. J., Jacob, D. J., McKeever, J., Jervis, D., Durak, B. O. A., Xia, Y., and Huang, Y.: Quantifying methane point sources from fine-scale satellite observations of atmospheric methane plumes, Atmos. Meas. Tech., 11, 5673–5686, https://doi.org/10.5194/amt-11-5673-2018 , 2018.
• Borsdorff, T., aan de Brugh, J., Pandey, S., Hasekamp, O., Aben, I., Houweling, S., and Landgraf, J.: Carbon monoxide air pollution on sub-city scales and along arterial roads detected by the Tropospheric Monitoring Instrument, Atmos. Chem. Phys., 19, 3579–3588, https://doi.org/10.5194/acp-19-3579-2019 , 2019.
• Lama, S., Houweling, S., Boersma, K. F., Aben, I., van der Gon, H. A. C. D., Krol, M. C., Dolman, A. J., Borsdorff, T., and Lorente, A.: Quantifying burning efficiency in Megacities using NO2 / CO ratio from the Tropospheric Monitoring Instrument (TROPOMI), Atmos. Chem. Phys. Discuss., https://doi.org/10.5194/acp-2019-1112 , in review, 2019.
• Kuhlmann, G., Broquet, G., Marshall, J., Clément, V., Löscher, A., Meijer, Y., and Brunner, D.: Detectability of CO2 emission plumes of cities and power plants with the Copernicus Anthropogenic CO2 Monitoring (CO2M) mission, Atmos. Meas. Tech., 12, 6695–6719, https://doi.org/10.5194/amt-12-6695-2019 , 2019.
• Villani, C, Optimal Transport. Vol. 338. Grundlehren der mathematischen Wissenschaften. Berlin, Heidelberg: Springer Berlin Heidelberg, 2009.
• Ferradans S., Papadakis N., Rabin J., Peyré G., Aujol JF. (2013) Regularized Discrete Optimal Transport. In: Kuijper A., Bredies K., Pock T., Bischof H. (eds) Scale Space and Variational Methods in Computer Vision. SSVM 2013. Lecture Notes in Computer Science, vol 7893. Springer, Berlin, Heidelberg
• Cuturi, M.: Sinkhorn Distances: Lightspeed Computation of Optimal Transport, NIPS’13, 2013.
• Peyré, G., and Cuturi, M.: Computational Optimal Transport, Foundations and Trends in Machine Learning, vol. 11 (5-6), 2019 .
• Fortems-Cheiney, A., Pison, I., Dufour, G., Broquet, G., Berchet, A., Potier, E., Coman, A., Siour, G., and Costantino, L.: Variational regional inverse modeling of reactive species emissions with PYVAR-CHIMERE, Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-186 , in review, 2019.
• V. Mallet, D. Quélo, B. Sportisse, M. Ahmed de Biasi, É. Debry, I. Korsakissok, L. Wu, Y. Roustan, K. Sartelet, M. Tombette and H. Foudhil: Technical Note: The air quality modeling system Polyphemus, Atmos. Chem. Phys., 7 (20), 2007 .

The candidate is expected to have a PhD in data assimilation or in any related discipline. Strong python, or any other prototyping and software development skills are required. Oral and written communication skills, as well as the ability to publish in top-ranking journals, are mandatory.