Bulking Up with the BOSS: Accurate Emulation of the TAU Bin Model's Liquid Microphysics Using Simple Bulk Schemes

We introduce several versions of the Bayesian Observationally constrained Statistical-physical Scheme (BOSS), a set of bulk schemes that accurately emulate the liquid microphysics in the Tel Aviv University bin model (TAU). BOSS uses Markov chain Monte Carlo to infer the best choice of parameters for a set of simple process rate formulas, and this Bayesian inference also produces estimates of the uncertainty in those parameter values. The different schemes presented differ in the details of the autoconversion formula, and by using either two or three moments to represent cloud drops.

It is common practice to produce process rate formulas for bulk microphysics schemes via regression using bin model outputs. However, an accurate fit in "offline" testing may not translate to accurate emulation of the bin scheme when run in a time-evolving model. We avoid this issue by taking the "observations" used for Bayesian inference from a simple 1-D driver running TAU, and use the same driver to evaluate BOSS. Parameters are thus constrained based on performance in an actual time-evolving model, rather than direct "offline" fits to process rates.

The resulting two-moment schemes are reasonable emulators for both nonprecipitating and drizzling cloud, albeit with some degree of compensating error in the drizzling cases. The schemes with three cloud moments produce an even better fit, with little evidence of compensating errors. These schemes substantially outperform equivalent schemes constrained only by offline process rates, demonstrating the importance of tuning and evaluating bulk microphysics schemes in a time-evolving context.

Finally, we demonstrate a four-moment single-category BOSS scheme, which does not artificially distinguish between cloud-sized and rain-sized particles. This scheme is also an effective emulator, it has the same number of prognostic variables as two-moment two-category schemes, and it avoids the difficulties inherent in parameterizing autoconversion.