Speaker: Marcus van Lier-Walqui (GISS/CU)
Topic: Model tuning approaches to quantify and reduce parametric and structural biases associated with clouds

Perturbed-parameter ensembles (PPE) are leading methods for improving GCM performance. These methods are attractive because they do not resort to "black-box" approaches that replace physical model components with machine learning substitutes. Instead, PPEs retain a connection to known physics, allowing for "bottom up" (detailed process-based) advances as well as "top-down" tuning based on emergent constraints of the climate system. We demonstrate how this methodology has been applied to the NASA ModelE using neural-network emulators together with a unique treatment of satellite observational biases. This PPE approach has resulted in unambiguous improvement in ModelE, while providing multiple plausible parameter sets. However, only a portion of GCM bias is attributable to misspecified parameter values -- other sources of bias are structural, i.e., owing to the hard-coded mathematic formulae and approximations of the various GCM parameterization schemes. We present a method for relaxing this structural rigidity in microphysics, while retaining a physical basis, called the Bayesian Observationally-constrained Statistical-physical Scheme. We demonstrate that this method can smoothly span between disparate physical formulations, essentially turning a multi-physics ensemble into a perturbed-parameter ensemble.