Using machine learning to improve predictions of ocean carbon storage by marine life

Dr Adrian Martin, Dr Stephanie Henson, Dr B.B.Cael, NOC; Professor Zudi Lu, S3RI; Dr Sian Henley, University of Edinburgh
Rationale: 

Carbon dioxide (CO2) dissolved in seawater is used by marine phytoplankton to grow, supporting the ocean’s foodweb and resulting in a downward flux of organic matter (the 'biological carbon pump' (BCP)) that sequesters carbon in the ocean and maintains atmospheric CO2 concentrations roughly 1/3 lower than in its absence. However, we lack an understanding of the relative importance of the BCP’s drivers and how they will respond to climate change, and, consequently, a robust way of predicting change in the BCP itself. Such change may decrease the ocean's role as a carbon reservoir, impacting national commitments to "net-zero". The recent IPCC Report (https://www.ipcc.ch/report/ar6/wg1/, Ch.5, section 5.4.4.2) asserted “high confidence that feedbacks to climate will arise from alterations to the magnitude and efficiency of the BCP”, but that the attributed drivers differ significantly. A deeper understanding is needed of the BCP in our predictive models where there is no consensus on how it is represented. Machine learning (ML) techniques provide powerful tools to infer the underlying dominant causal influences on the BCP and how they shift into the future across a range of earth system models, allowing more robust assessment of global change in the BCP.

Methodology: 

The project will address the variability in climate model predictions for the BCP. Although biogeochemical interactions within all models providing input to CMIP6 (key contributors to IPCC AR6 -  https://esgf-node.llnl.gov/projects/cmip6/) are prescribed by explicit equations, the cumulative effect of their non-linear interactions, exacerbated by the ocean circulation transporting organisms and nutrients, makes it challenging to identify the emergent dominant drivers. Contrasting predictions arise from two sources. First, it may be due to differences in the dominant drivers both physical (e.g. temperature) and biogeochemical (e.g. nutrient availability) between models. ML techniques (e.g.  random forest, neural nets) will be applied to a range of CMIP6 models to determine the dominant influences on key metrics in each model for the strength of the BCP (e.g. carbon flux past 1000m). Initially focus will be on the ‘best’ (high mitigation) and ‘worst’ (little mitigation) climate scenarios. Second, variation in predictions for the drivers themselves will play a role. To quantify this, the ML model extracted for each of the CMIP6 models will be driven with driver output from the other models. Using this two-step approach, the student will substantially improve our understanding of the currently large uncertainties regarding predicted change in the globally important BCP.

 

Location: 
NOC/UoS
Training: 

The INSPIRE DTP programme provides comprehensive personal and professional development training alongside extensive opportunities for students to expand their multi-disciplinary outlook through interactions with a wide network of academic, research and industrial/policy partners. The student will be registered at the University of Southampton and hosted at NOC with co-supervision coming from the University of Southampton and University of Edinburgh.. Applicants are welcomed from any numerate discipline so specific training will vary with individual but may include: use of machine learning techniques, ocean biogeochemistry, analysis of complex earth system models. To provide an appreciation of the difficulties and issues associated with collecting observations that can be used to inform and test models, the student would be given the opportunity to participate in a scientific cruise. Sea-going is not compulsory though. 

Eligibility & Funding Details: 

Please see https://inspire-dtp.ac.uk/how-apply for details.

 

Background Reading: 

·         Kwiatkowski et al. (2020) Twenty-first century ocean warming, acidification, deoxygenation, and upper-ocean nutrient and primary production decline from CMIP6 model projections. Biogeosciences, 10.5194/bg-17-3439-2020, 2020.

·         Sonnewald et al. (2021) Bridging observations, theory and numerical simulation of the ocean using machine learning. Env. Res. Lett., 10.1088/1748-9326/ac0eb0

           Wanninkhof et al. (2021) Integrated ocean carbon research: a summary of ocean carbon research, and vision of coordinated ocean carbon research and observations for the next decade. IOC Tech. Ser., 158, 10.25607/h0gj-pq41