Search for a project

Machine learning for atmospheric chemistry modelling.

Prof Mat Evans (YDC), Prof Ally Lewis (YDC), Dr Christopher Keller (NASA GMAO)

Project partner(s): NASA Global Modelling and Assimilation Office (GMAO) - NASA Goddard

Contact email: mat.evans@york.ac.uk

Introduction

Air pollution, climate change and ecosystem degradation are due to the changing composition of our atmosphere. Numerical models of the chemistry, emissions, deposition and transport in the atmosphere help us to understand the science underpinning these problems, predict the future and evaluate mitigation strategies.

These simulations are hugely numerically complex with the chemistry component often the most computationally intense aspect. Over the last 30 years efforts have been made to improve the efficiency of the algorithms used, or reduce the complexity of the chemistry, these efforts have not lead to significant reductions in the computational burden. This computational bottle-neck has hampered atmospheric chemistry’s ability to take full advantage of tool and techniques (ensembles, data assimilation) used for weather forecasting or climate simulations.

 Figure 1. Aerosols simulated by the NASA GEOS-5 system. Further examples from the old NASA GEOS-5 system can be seen here: https://svs.gsfc.nasa.gov/30017.

Project aims and objectives

In a proof of concept study, we have shown that “machine learning” (specifically random forest regression https://en.wikipedia.org/wiki/Random_forest) can be used to represent the chemistry in a chemistry transport model replacing the standard integrator approach. 

Figure 2. Results from the proof of concept activity. Two simulations of ozone using the GEOS-Chem model showing the results for a number of locations around the world. The standard simulation is shown in Black, with a simulation run with the standard representation of chemistry replaced by a machine learning representation shown in Red. The machine learning shows good comparison to the standard approach. 

This new approach to the representation of atmospheric chemistry opens up a range of possibilities to speed up air pollution and climate models and so allow their use at high resolution, in ensemble, to more readily undertake data assimilation, to address new science topics etc.  This project will take this proof of concept further, explore differing machine learning approaches and how to optimize a machine learning approach for differing computational hardware. Other aspect of a chemistry transport model (such as the photolysis calculations, aerosol thermodynamics, aerosol microphysics) will also be investigated to see whether a machine learning methodology would allow improved speed and performance in those areas.

The project will use the open-source GEOS-Chem model of atmospheric chemistry and transport (www.geos-chem.org) which is extensively used by the atmospheric chemistry group in York and by the wider global atmospheric chemistry community.

Computational facilities will be provided by the in-house provision in the atmospheric chemistry group, by computation available from the University of York, and (subject to checks) NASA facilities.

Project Partner

The NASA Goddard’s Global Modelling and Assimilation Office will act as the project partner and will help guide the project to ensure that the scientific developments are focussed on developing techniques which have a practical operational focus. Their GEOS-5 forecasting and assimilation system (https://gmao.gsfc.nasa.gov/forecasts/) uses GEOS-Chem for atmospheric chemistry, thus developments made in York should readily fit into the GEOS-5 system. The student would be expected to spend two periods at NASA Goddard; an initial visit to help scope the nature of the problem and a subsequent longer one to potentially help with implementation of the machine learning into the NASA GEOS-5 forecasting model.

Training and Student Background

The student need to have a good computational background and an interest in environmental science issues. It would suit students with wide range of backgrounds in chemistry, computer science, mathematical, physics, engineering etc.

The student will work in the Wolfson Atmospheric Chemistry Laboratory (WACL), part of the Department of Chemistry, University of York and the National Centre for Atmospheric Science (NCAS).  These laboratories were established in 2013 and comprise a state of the art 800 m2 dedicated research building, the first of its kind in the UK. Supported by a large award from the Wolfson Foundation and a private donor, the Laboratories enable experimental and theoretical studies relating to the science of local and global air pollution, stratospheric ozone depletion and climate change. The Laboratories co-locate around 50 researchers from seven academic groups and from NCAS. The Laboratories are also home to independent research fellows, postdoctoral researchers, PhD students and final year undergraduate research projects.

The University of York and the wider NERC SPHERES DTP provide comprehensive training programmes for PhD students with a range of courses on both hard and soft skills. The student will also have access to the wider resources of NCAS such as the introduction to atmospheric science course and instrumental science course, together with other courses in computations and data analysis. Additional training for students without an atmospheric chemistry background will be provided by auditing the undergraduate atmospheric chemistry courses provided by the University of York. Similarly, training in machine learning will be provided through engaging with a range of online training courses that are available from MOOCs such as EdX and Coursera.

The student will have the opportunity to present their work to the scientific community at national and international meetings and conferences. They will also be encouraged to take part in outreach events organised by both WACL and NCAS in order to disseminate the research beyond the immediate scientific community (e.g. to policymakers and the general public).

Useful References and Links

https://en.wikipedia.org/wiki/Machine_learning

https://en.wikipedia.org/wiki/Random_forest

http://www.geos-chem.org

https://gmao.gsfc.nasa.gov/forecasts/

http://wiki.seas.harvard.edu/geos-chem/index.php/Profiling_GEOS-Chem

http://www.sciencedirect.com/science/article/pii/S1352231010006242

https://www.atmos-chem-phys.net/4/2025/2004/acp-4-2025-2004.pdf

Related undergraduate subjects:

  • Applied mathematics
  • Atmospheric science
  • Chemistry
  • Computer science
  • Computing
  • Earth system science
  • Electrical engineering
  • Mathematics
  • Meteorology
  • Natural sciences
  • Oceanography
  • Physical science
  • Physics
  • Statistics