Breakout Summary Report

Breakout Description

Kilometer-grid-scale models are widely used at regional scales and recently became available for global weather and climate simulations. For example, DOE’s Simple Cloud-Resolving E3SM Atmosphere Model (SCREAM, referring to the cloud-resolving configurations of the current EAMxx) has successfully released its 40-day global simulation with 3.25-km horizontal grid spacing (version 0, Caldwell et al. 2021), and version 1 development is nearly complete.    

This session aims to discuss how to best use ARM measurements to improve the representations of cloud and convection processes in kilometer-grid-scale models. Novel approaches with a view to future opportunities is a particular focus. Discussions focus on processes that kilometer-grid-scale models struggle to accurately simulate. These processes include planetary boundary-layer turbulence, shallow-to-deep convection transition, and convection aggregation including upscale growth to mesoscale convective systems. This session includes talks and discussions on current model physics biases, frameworks to connect ARM observations with kilometer-scale model outputs, and observational needs for diagnosing and improving physics parameterizations at local, regional, and global scales.

Main Discussion

This session invited nine short talks (around six minutes or up to four slides). After these talks, we discussed for about 45 minutes.

Among all these talks, seven focused on studies of GoAmazon, and two focusing on development of model diagnostics. This year, the breakout emphasized GoAmazon because we wanted to illustrate how one could build up a modeling test/framework around observations.  The first three GoAmazon talks focused on mesoscale convective systems and isolated convection, kilometer-scale modeling biases, and challenges to fully represent the convective processes and their coupling with the land surface, specifically (1) MCS sensitivities to model grid spacing (M Chasteen), (2) mixing effects on clouds induced by horizontal vorticity in km-scale models (SM Hagos), and (3) shallow-to-deep transitioning sensitivity to moisture perturbations (H Barbosa). The following four talks related to the THREAD project showed: 1) the identification of DOE E3SM/SCREAM global model biases on precipitation cluster size underestimation using GoAmazon radar data (Y Zhang); 2) benchmark LES on convection triggering and aggregation built from observed shallow-to-deep convection cases of GoAmazon (Y Tian); 3) in-depth DP-SCREAM (a standalone doubly periodic CRM version of SCREAM, analog to SCM with respective to low-res climate models) sensitivity tests against benchmark LES on possible fixes of model biases (P Bogenschutz); 4) RRM-SCREAM (inner regionally refined domain of 3-km grid spacing around ARM sites while outer domain slowly transits to low-res climate model resolution) centered around the GoAmazon site to facilitate hindcast runs (coupled with land models and large scale circulations) and intensive tests on the model improvements learning from DP-SCREAM (H-Y Ma). The last two talks presented existing diagnostics packages for E3SM with plans for extension to km-scale models (C Tao and S Tang).

During the discussion, the focus was on two main questions:

1. How to best connect ARM observations with models including E3SM/SCREAM, e.g., ARM case library?


2. What are acceptable “fair” comparisons between model output and observations?

Key Findings

The questions listed above triggered enthusiastic discussions among participants. In general, there was interest in bringing ARM data into the model development cycle for diagnosis and improvement.  Below is a summary of the discussion.

There is consensus that an observation-based case library of golden days is valuable for our physical understanding and model performance evaluations. However, as ARM has accumulated long-term measurements, an ensemble of golden days sampling different meteorological conditions would be more beneficial to test model sensitivities and applicability of parameterized processes to different climate regimes. Such a library, consisting of different convective/cloud regimes, would be useful for model calibrations of tuning parameters or processes; thus, it is key to include those observable metrics in addition to forcing data. Such observable metrics should also consider information on the observational uncertainties, which is useful especially for machine learning-based calibration methods. Global model tuning processes often introduces compensating errors and model evaluations often cannot isolate errors and attribute them to each of the parameterized processes.  It is still unclear how process-level tuning and calibration may help with this challenge.  It might be helpful to create a hierarchy of ARM cases with increasing complexity, so that it could help isolate model biases step by step with adding complexity of more processes involved.

This approach is similar to “testbeds'' that have been employed in other agencies such as NOAA for model testing and development. While there was broad support for this, a roadblock is resources to facilitate it. Another concern is how to sufficiently include observationalists, product developers, and process-level expertise in the process of model evaluation and development, particularly when many data sets require further quality control and retrievals require carefully curated situations to provide valid comparisons that rely on knowledge of how retrievals were produced. How to best do this is still an open question, so efforts across EESSD might be needed to facilitate it.

Issues

n/a

Needs

n/a

Decisions

n/a

Future Plans

Our discussion was more focused on the case library, with less emphasis on the topic of “acceptable” fair model-observation comparisons. The discrepancy in model size distribution assumptions in microphysics and radiation was mentioned, especially in the context of application of radar simulators. However, we did not have time to elaborate on this.  Also, we did not have enough time to discuss those observables key to constrain the parameterized processes related to model biases.  This requires both modelers and observational experts to be heavily involved in the discussion.  We feel it is important to have such discussions more frequently at DOE meetings.

Action Items

• Mid-year 2-hour virtual gathering of GoAmazon studies for progresses and discussions


• Continue discussions with ARM infrastructure team on diagnostic developments


• Explore more opportunities to talk about these topics at DOE meetings