Project Outline
Exposure to poor air quality has detrimental effects on human health, particularly for high-risk groups (e.g. the very young and old and those with health conditions). Process-based air quality models are increasingly used to alert the public and health care providers of upcoming air pollution episodes. Like weather forecasts, air quality forecasts are provided up to several days ahead with forecast confidence decreasing with increasing forecast lead time. However, these models are subject to substantial bias with forecast skill often inadequate, particularly for the most ‘extreme’ episodes. This is concerning since such events may present the greatest health risk and are likely to become more common in a warmer climate (Gouldsbrough et al., 2022, 2023a).
Emerging data-driven forecasting systems (e.g. with ‘deep learning’) show remarkable promise in forecasting air pollution, with skill to rival or better process models, and are rapidly becoming the new state-of-the-art (e.g. Kleinert et al., 2021; Gouldsbrough et al., 2023b). With sufficient training data, machine learning approaches may provide automated, computationally efficient forecasts and have ‘game changing’ potential to sit alongside or even replace large-scale process-based models. Operational systems that use machine learning to forecast ground-level ozone, a pollutant responsible for >14,000 premature deaths annually in Europe, are in advanced development for use in (e.g.) Central Europe (Leufen et al., 2023). However, despite having an extensive air quality monitoring infrastructure and being data-rich, the UK is behind in the application of AI in this field.
Working across disciplines, this PhD project will explore how machine learning approaches can improve our ability to hindcast and forecast UK air pollution. Depending on the interests of the student, areas of focus might include: (1) development of probabilistic forecasts, (2) forecasting the co-occurrence of pollutant episodes with (e.g.) heatwaves, and/or (3) examining approaches most suitable for forecasting extreme air pollution events. A key aspect of the work will be to prepare data-driven forecasts and to evaluate their skill relative to a range of existing process-based models serving the UK over a range of lead times (i.e. forecasts 24-96 hours ahead).
Benefiting from a multidisciplinary supervisory team, the PhD student will join the vibrant atmospheric science research group (‘AtMOS’) in Lancaster Environment Centre (LEC) and the Environmental Statistics group in the Department of Maths & Stats. They will be trained in state-of-the-art machine learning methods, gain specialist knowledge in their application to air pollution data and will become expert in relevant atmospheric processes (e.g. heatwaves) and context. The student will interact with members of Lancaster’s growing Centre of Excelleance in Environmental Data Science (CEEDS), and will develop transferable skills in environmental data science, including across data acquisition, processing, visualisation and end-product presentation.
Who should apply: The project will involve substantial data curation and analysis and is well suited towards gradutes with a background in Data Science, Maths, Statisitics, Computer Science, Physics or a related quantitaive discpline. An interest in atmospheric science and climate would be a key advantage. Open to home students.
Start date: October 2023
Supervisors: Ryan Hossaini (LEC), Emma Eastoe (Maths/Stats), Andrea Mazzeo (LEC)
External supervisor: Lily Gouldsbrough (UKCEH)
To Apply: Send a CV and cover letter to Dr Ryan Hossaini ([Email Address Removed]) and Dr Emma Eastoe ([Email Address Removed]) by the closing date of 5 pm on Monday 31st July 2023.
References
Gouldsbrough, L., Hossaini, R., Eastoe, E., & Young, P.J.Y. (2022). A temperature-dependent extreme value analysis of UK surface ozone, 1980-2019. Atmos. Env. 273, 118975.
Gouldsbrough, L., Hossaini, R., Eastoe, E., Young, P.J.Y. & Vieno, M. (2023). A machine learning approach to downscale EMEP4UK: analysis of UK ozone variability and trends. Atmos. Chem. Phys., Submitted.
Gouldsbrough, L., Eastoe, E., Hossaini, R., & Young, P.J.Y. (2023). Identifying the drivers of high-level ozone events using machine learning classification. Submitted.
Leufen, L.H., Kleinert, F., & Schultz, M.G. (2023). O3ResNet: A deep learning based forecast system to predict local ground-level daily maximum 8-hour average ozone in rural and suburban environment. Artificial Intelligence for the Earth Systems, in press.
Kleinert, F., Leufen, L.H., Lupascu, A., Butler, T., & Schultz, M.G. (2022). Representing chemical history in ozone time-series preictions – a model experiment study building on the MLAir (v1.5) deep learning framework, Geosci. Model Dev., 15, 8913–8930, https://doi.org/10.5194/gmd-15-8913-2022.