An Efficient Bayesian Approach to Learning Droplet Collision Kernels: Proof of Concept Using “Cloudy,” a New n‐Moment Bulk Microphysics Scheme

Abstract

Historically, microphysics schemes were tuned to data in an ad-hoc way, resulting in parameter values that are not repeatable or explainable • Bayesian inference puts uncertainty quantification and parameter learning on solid mathematical grounds, but is computationally expensive • We present a proof-of-concept of computationally efficient Bayesian learning applied to a new bulk microphysics scheme called "Cloudy"

New developments in PySDM and PySDM-examples v2: collisional breakup, immersion freezing, dry aerosol initialization, and adaptive time-stepping

“…In PySDM-examples, we introduced a set of notebooks reproducing figures from two publications. In Bieli et al (2022), PySDM results from collisional coalescence and breakup were used as a calibration tool for learning microphysical rate parameters. In Jong et al (in review), the physics of and algorithm for superdroplet breakup are described, and the impact of breakup on cloud properties is demonstrated with box and single-column simulations (the latter based on Shipway & Hill (2012)).…”

EnsembleKalmanProcesses.jl: Derivative-free ensemble-based model calibration

“…Within CES, the trained emulators are used to sample this probability distribution, and by design are most accurate where they need to be. CES has been successfully used to quantify parameter uncertainty within the moist convection scheme of a simplified climate model (Dunbar et al, 2021(Dunbar et al, , 2022, within a droplet collision-coalescence scheme for cloud microphyiscs (Bieli et al, 2022), and within boundary layer turbulence schemes for ocean modeling (Hillier, 2022).…”

Spanning the Gap From Bulk to Bin: A Novel Spectral Microphysics Method

“…For comparison with the BF method, we solve each test case numerically using the flux method for spectral bin microphysics with 32 single‐moment bins (Bott, 1998), a two‐ or three‐moment closure method of moments (Bieli et al., 2022), and a Lagrangian particle‐based code called PySDM (v2.5) (Bartman et al., 2022). The bin method used follows the original setup from Bott (1998), spanning a range of 0.633–817 μm radius with mass doubling between bins, and a time step selected to be sufficiently small as to prevent numerical instability (1–100 s depending on the dynamics).…”

“…Bieli et al. (2022) present a more efficient way to learn these parameters within a similar bulk microphysics framework that still relies on closures. More complex yet, Rodríguez Genó and Alfonso (2022) tackle the challenge of inverting multimodal distribution closures using a machine‐learning based method, which could avoid the necessity for cloud‐rain conversion rate parameterizations.…”

“…A different moment-based method, the BOSS scheme proposed by Morrison et al (2019) leaves all process rates and closures as generalized power series whose parameters are learned from data. Bieli et al (2022) present a more efficient way to learn these parameters within a similar bulk microphysics framework that still relies on closures. More complex yet, Rodríguez Genó and Alfonso (2022) tackle the challenge of inverting multimodal distribution closures using a machine-learning based method, which could avoid the necessity for cloud-rain conversion rate parameterizations.…”