Research
Learning the Earth with Artificial Intelligence and Physics (LEAP), an NSF Science Technology Center (STC), will reduce errors in near-future climate projection by replacing many subcomponents of Earth system models with novel machine learning (ML) algorithms. To do so, novel ML will be developed that is able to better extrapolate by including domain and causal knowledge while aggressively leveraging the wealth of recent datasets. The Center will focus on implementation with the open-source Community Earth System Model (CESM). Specifically, the Center will focus upon: 1) reducing the existing model structural errors, related to the lack of comprehension of the process at play (for example, clouds, and microphysics); 2) optimally estimating the model parameters using a Bayesian approach; and 3) developing new observational products which will be used to evaluate the CESM skill.
Through a deep collaboration between climate data scientists, and with engagement from the public and private sectors, national research labs, and multiple partner stakeholders, LEAP will hybridize ML by integrating physical knowledge to develop a next-generation CESM to improve climate prediction 10-40 years into the future, therefore providing to policymakers a “leap” in risk quantification and mitigation knowledge. This improved and more detailed information on future climate will be communicated in ways that are more digestible and relevant to stakeholders. LEAP will support new bidirectional knowledge transfer with the public and private sectors to develop tailored and relevant climate-related information or stakeholders so that they can better adapt to climate change and invest into risk-limited sectors.
LEAP will merge physical modeling with ML across a “Knowledge-Data Continuum.” The benefits to both communities will be usable as a template for other climate modeling centers. Climate scientists and modelers struggle to fully integrate the wealth of existing datasets into their models, while ML algorithms have been good at replacing and interpolating but less targeted at extrapolating. By combining both approaches, LEAP triggers a significant advancement for data science applied to physical problems. LEAP incorporates physics into ML algorithms for better generalization and extrapolation, while optimally using the wealth of data available to climate science, in order to better predict the future.
