Explainable AI and Process Diagnostics to Understand State-Dependent Precipitation Forecast Errors

Due to the coupled nature of the earth system, precipitation forecast errors at subseasonal-to-seasonal (S2S; 2 weeks to 2 months) lead times are caused by a combination of errors/biases from the atmosphere, ocean, ice and land across a range of spatial and temporal scales. Unrealistic sensitivity of model convection to its large-scale environment, as well as unrealistic strength of prominent feedbacks (e.g. cloud radiative, wind-evaporation), can lead to the inability to maintain subseasonal tropical convection variability in forecasts. Even if models were able to perfectly simulate tropical fields, errors in the subtropical and midlatitude circulations can further introduce forecasts errors in U.S. precipitation via incorrect teleconnections. This means that identifying the correct combination of model biases that lead to specific precipitation forecast errors is incredibly challenging. Furthermore, these different processes are not always active at any given time (e.g. the Madden-Julian oscillation can be in an inactive state), implying that their associated biases only contribute to forecast errors intermittently. Thus, an additional challenge in understanding model forecast errors at S2S lead times is identifying the intermittent states of the system when these biases are most important. Here, we propose to couple novel explainable artificial intelligence (XAI) techniques with process-oriented diagnostics to identify, understand, and correct via post-processing, state-dependent UFS precipitation forecast errors.
The proposed work is organized into three distinct activities that focus on improving UFS forecasts of North American precipitation at S2S lead times. Activity I involves refining and then implementing an XAI framework to identify state-dependent UFS precipitation errors. Activity II revolves around understanding the tropical-extratropical teleconnection processes relevant to the climate states identified by the XAI method. To do this, we will use process-oriented model diagnostics and test our understanding with simplified model experiments. Activity III aims to leverage what we have learned in Activities I and II to develop XAI-derived post-processing corrections to improve UFS precipitation forecasts for specific initialization states.