Assessing Sources of Precipitation Predictability in E3SM with Explainable Artificial Intelligence
| Active Dates | 9/1/2023-8/31/2026 |
|---|---|
| Program Area | Regional Global Modeling Analysis |
Project Description
Assessing Sources of Precipitation Predictability in
E3SM
with Explainable
Artificial Intelligence
Principal Investigator: Prof. Elizabeth A. Barnes, Colorado State University
Co-Principal Investigators: Dr. Po-Lun Ma, Pacific Northwest National Laboratory
Precipitation anomalies and extremes in the midlatitudes are predictable, in part, by the large-scale climate circulation. This is especially true on subseasonal-to-seasonal (S2S) timescales, where successful forecasts rely heavily on large-scale modes of variability. Unfortunately, climate models (including E3SM) do not capture the relevant processes with fidelity. Furthermore, it is unknown whether there are additional sources of S2S precipitation predictability that have yet to be identified and whether climate models capture these dynamics as well. Finally, external forcing has, and is expected to continue, modifying large-scale modes of climate variability. It is thus very likely that anthropogenic forcing has already modified precipitation predictability and will continue to do so in the coming decades, but it is unclear what magnitude of changes have already occurred.
Here, we propose to leverage advances in explainable machine learning (XAI) to quantify and understand teleconnections associated with large-scale modes of variability which act as sources of precipitation predictability in E3SM at S2S timescales. Specifically, we will (1) use XAI methods to quantify large-scale teleconnections that act as sources of precipitation predictability within E3SM and reanalysis, with explicit consideration of uncertainty; (2) understand the E3SM modeling choices relevant to improve the representation of the sources of predictability, and test our hypotheses with XAI-informed E3SM nudging experiments; and (3) quantify changes in E3SM predictability under different climate change scenarios. A particularly novel aspect of this work is that we will explicitly consider the full predicted conditional distribution of precipitation in order to assess more than just the mean precipitation response. That is, we will be able to explore the predictions of the extremes (i.e. tails of the distribution) as well as how the spread of the distribution shifts when combinations of sources of predictability (e.g. large-scale climate modes) are active.
Principal Investigator: Prof. Elizabeth A. Barnes, Colorado State University
Co-Principal Investigators: Dr. Po-Lun Ma, Pacific Northwest National Laboratory
Precipitation anomalies and extremes in the midlatitudes are predictable, in part, by the large-scale climate circulation. This is especially true on subseasonal-to-seasonal (S2S) timescales, where successful forecasts rely heavily on large-scale modes of variability. Unfortunately, climate models (including E3SM) do not capture the relevant processes with fidelity. Furthermore, it is unknown whether there are additional sources of S2S precipitation predictability that have yet to be identified and whether climate models capture these dynamics as well. Finally, external forcing has, and is expected to continue, modifying large-scale modes of climate variability. It is thus very likely that anthropogenic forcing has already modified precipitation predictability and will continue to do so in the coming decades, but it is unclear what magnitude of changes have already occurred.
Here, we propose to leverage advances in explainable machine learning (XAI) to quantify and understand teleconnections associated with large-scale modes of variability which act as sources of precipitation predictability in E3SM at S2S timescales. Specifically, we will (1) use XAI methods to quantify large-scale teleconnections that act as sources of precipitation predictability within E3SM and reanalysis, with explicit consideration of uncertainty; (2) understand the E3SM modeling choices relevant to improve the representation of the sources of predictability, and test our hypotheses with XAI-informed E3SM nudging experiments; and (3) quantify changes in E3SM predictability under different climate change scenarios. A particularly novel aspect of this work is that we will explicitly consider the full predicted conditional distribution of precipitation in order to assess more than just the mean precipitation response. That is, we will be able to explore the predictions of the extremes (i.e. tails of the distribution) as well as how the spread of the distribution shifts when combinations of sources of predictability (e.g. large-scale climate modes) are active.
Award Recipient(s)
- Colorado State University, Fort Collins (PI: Barnes, Elizabeth)