Optimizing Objective Model Calibration Approaches using Single Column Models

TL;DR

This paper introduces a framework using the Single Column Atmosphere Model to evaluate and improve objective calibration methods for atmospheric model parameters, balancing computational efficiency with realistic physical processes.

Contribution

It presents a novel perfect-model experiment setup with synthetic data to assess parameter recovery and calibration approaches in atmospheric modeling.

Findings

Bayesian calibration yields more consistent parameter recovery than point estimates.

Tighter parameter constraints are achieved with a Gaussian Process emulator and MCMC.

Synthetic experiments identify observables most informative for parameter calibration.

Abstract

Sub-grid scale parameterizations in atmospheric models involve numerous uncertain parameters that must be tuned to align simulations with observations. Here, we propose a framework for assessing objective tuning frameworks using the Single Column Atmosphere Model (SCAM), which retains key physical parameterizations of general circulation models (GCMs) while greatly reducing computational cost. We conduct a perfect-model experiment where we run SCAM with a known "true" parameter set to generate synthetic observations that mimic Atmospheric Radiation Measurement (ARM) Intensive Observation Periods. Perturbed parameter ensembles are constructed by varying microphysics, convection, and aerosol parameters, and cloud-radiation fields are evaluated over the Southern Great Plains. We find that point estimates find solutions that greatly reduce model-observation misfit without recovering the true parameter values. In contrast, a Bayesian framework using a Gaussian Process emulator with Markov Chain Monte Carlo sampling yields tighter constraints on some parameters and more consistent recovery across experiments and variables. The perfect model framework allows to assess which observables yield most information, which parameters are recoverable given a certain set of observations, and what is the minimum observational record needed. Although this study focuses on a single location with synthetic observations, such experiments provide a controlled setting to evaluate and identify robust calibration frameworks, which can then be extended to multiple locations and real observations with greater confidence.