Public Abstract

DE-SC0025209: Using apparent relationships derived from machine learning methods to improve the simulation of marine organisms within the Energy Exascale Earth System Model

Award Status: Active

Institution: The Johns Hopkins University, Baltimore, MD
UEI: FTMTDMBR29C7
DUNS: 001910777

Most Recent Award Date: 08/28/2024
Number of Support Periods: 1
PM: Davis, Xujing

Current Budget Period: 09/01/2024 - 02/28/2025
Current Project Period: 09/01/2024 - 08/31/2027
PI: Gnanadesikan, Anand

Supplement Budget Period: N/A

Public Abstract

Phytoplankton are critical players in ocean ecosystems, forming the base of the marine food web. Approximately 10 gigatons of biomass generated by phytoplankton productivity sinks out of the surface ocean each year, resulting in the storage of ~2000 Gt of carbon in the deep ocean, lowering atmospheric carbon dioxide. However, phytoplankton also produces calcium carbonate, which is associated with a loss of carbon to the atmosphere. Representing phytoplankton accurately is thus an important goal of Earth System Models (ESM). This project brings together two groups from John Hopkins University and Los Alamos National Laboratory with extensive experience diagnosing and building Earth System Models and implementing phytoplankton biogeochemistry within them is to use machine learning methods to understand the process-level origins of marine biogeochemistry biases in Earth System Models (ESMs) with the aim to guide focused Energy Exascale Earth System Model (E3SM) model development. A compelling case is made that ML emulators trained to reproduce the spatiotemporal variability of marine BGC target fields are computationally efficient tools for exploring sensitivity to different inputs. The project aims to focus on significant differences that exist between satellite estimates and E3SM model results for phytoplankton carbon biomass and particulate inorganic carbon. A set of well-posed hypotheses will guide a systematic assessment of the relative contributions to model bias of environmental conditions, ecological interactions, BGC model formulation, and observational error. The team has already demonstrated the viability of the emulator approach in several publications, and these ML methods will be expanded to develop new understanding of E3SM bias in representing different phytoplankton functional types and zooplankton. The goal is to inform E3SM model developments needed to improve marine BGC realism with a focus on environmental inputs (such as phosphate) and growth/grazing formulations. The impacts of any changes in BGC model formulation coming out of this work will be assessed in climate change simulations. The proposed work is expected to deliver impactful science, technical advancements (including tools), and ultimately reduced bias in E3SM simulations of marine BGC.

Phytoplankton are critical players in ocean ecosystems, forming the base of the marine food web. Approximately 10 gigatons of biomass generated by phytoplankton productivity sinks out of the surface ocean each year, resulting in the storage of ~2000 Gt of carbon in the deep ocean, lowering atmospheric carbon dioxide. However, phytoplankton also produces calcium carbonate, which is associated with a loss of carbon to the atmosphere. Representing phytoplankton accurately is thus an important goal of Earth System Models (ESM).

This project brings together two groups from John Hopkins University and Los Alamos National Laboratory with extensive experience diagnosing and building Earth System Models and implementing phytoplankton biogeochemistry within them is to use machine learning methods to understand the process-level origins of marine biogeochemistry biases in Earth System Models (ESMs) with the aim to guide focused Energy Exascale Earth System Model (E3SM) model development. A compelling case is made that ML emulators trained to reproduce the spatiotemporal variability of marine BGC target fields are computationally efficient tools for exploring sensitivity to different inputs. The project aims to focus on significant differences that exist between satellite estimates and E3SM model results for phytoplankton carbon biomass and particulate inorganic carbon. A set of well-posed hypotheses will guide a systematic assessment of the relative contributions to model bias of environmental conditions, ecological interactions, BGC model formulation, and observational error. The team has already demonstrated the viability of the emulator approach in several publications, and these ML methods will be expanded to develop new understanding of E3SM bias in representing different phytoplankton functional types and zooplankton. The goal is to inform E3SM model developments needed to improve marine BGC realism with a focus on environmental inputs (such as phosphate) and growth/grazing formulations. The impacts of any changes in BGC model formulation coming out of this work will be assessed in climate change simulations. The proposed work is expected to deliver impactful science, technical advancements (including tools), and ultimately reduced bias in E3SM simulations of marine BGC.