Exploring Global Aerosol Cloud Interactions in Unprecedented Numerical Fidelity by Combining Breakthroughs in Multi-Scale Modeling, GPU Supercomputing & Neural Network Process Emulation
Aerosol-cloud interactions (ACI) continue to represent one of the largest uncertainties in climate model projection due to multi-scale simulation challenges. Adequately representing the range of scales of important processes from the detailed kinetics of microphysical droplet nucleation around condensation nuclei, through their complex effects in modifying the spatial organization dynamics of cloud-forming updrafts, to their ultimate macroscale effects is a formidable computational challenge. As such the ACI dynamics simulated by current versions of the DOE’s most state of the art global climate models – including prototypes that resolve 2-4 km horizontal resolution natively — are unsatisfying. They must rely on assumption-prone parameterizations for unresolved sub-km updraft statistics, which struggle to represent fundamental coherent turbulent spatial structures, as well as on an incomplete representation of the aerosol activation process occurring within these updrafts.
We see computational opportunities to address both shortcomings with exciting new technology emerging in DOE’s modeling portfolio. Our approach is two-pronged. The first novel aspect is to run an advanced GPU-compatible version of the E3SM-Multiscale Modeling Framework (MMF) at coarse exterior (2-degree) and ultra-fine (< 200-m) interior resolution. MMFs are an interesting class of climate model that copes with the demands of ultra-high horizontal resolution by embedding small micro-domains of explicit fine-scale dynamics within each grid cell of a conventional coarse resolution global climate model. GPUs, or “graphical processor units,” are advanced microprocessor chips having interesting computational capabilities especially for highly parallelized calculations such as the embedded cloud-resolving arrays used in MMFs. This simulation approach, also known as “cloud superparameterization” – though not without its own idealizations — allows a global climate model to capture an unprecedented level of realistic detail in the spectrum of sub-cloud eddies that form low clouds, and which mediate ACI, including a satisfying regime dependence of ACI in the response of raining vs. non-raining cloud, with minimal parameterization assumptions.
The second novel aspect of our proposed work is to then exploit the MMF together with modern machine learning methods to build a credible emulator of aerosol activation. Aerosol activation is the process through which small particles such as soot, dust or sea salt mediate the initial formation of cloud droplets that can then grow to become large enough to precipitate; more aerosol tends to distribute liquid water across more droplets, inducing significant changes in the reflectivity of clouds and thus their energetic importance to the climate system. Our improved numerical representation of this nucleation process will be achieved by harvesting our MMF’s high- resolution atmospheric state information to drive millions of computationally expensive “parcel model” calculations – a form of detailed microphysics computation that takes into account non-equilibrium explicit kinetics of transient aerosol activation dynamics that are traditionally neglected. This allows us to build a machine learning “training library” of paired inputs and outputs that we intend to use to develop a neural network parameterization of high-fidelity aerosol activation on the microphysical scale. Outsourcing the costly explicit parcel model calculations to neural networks can dramatically reduce the cost of using them routinely in the global climate model. Thus we will predict aerosol activation in the E3SM-MMF, replacing a traditional Abdul-Razzak scheme that must make considerably more assumptions about aerosol nucleation physics, in a way that can also interact with the MMF’s resolved, sub-km eddies to produce modifications of the essential dynamics at the heart of aerosol-cloud interactions, to understand their sensitivity to avoiding traditionally necessary process approximations.
The model development aspect of this work exploits a newfound ability generated following several years of sustained investment within the DOE Exascale Computing Project to run MMFs at unprecedented throughput and resolution on GPU-based LCFs. Prototypes of the E3SM-MMF have been proven in concept to achieve impressive throughput for resolutions previously unthinkable. This dramatically reduces the historic barrier to admitting sub-km turbulence on planetary scales and is possible on today’s DOE clusters using E3SM-MMF. Meanwhile, neural network emulators, once trained using automated architecture scans and emerging software to deploy forward inferences within climate simulation code dramatically reduce the computational barrier to the inclusion of explicit aerosol activation kinetics.
We will thus construct a customized experimental configuration of the E3SM-MMF climate model having ~1-degree exterior resolution, ~200-m 3D embedded nterior resolution, and modified microphysics including a new high fidelity emulator of aerosol activation.
The main science goal will then be to deploy the modeling framework on a suite of nudged- and free-running simulations to shed new light on multi-scale aerosol-cloud interaction dynamics and their water cycle implications at a substantially higher degree of freedom, and lower degree of process approximation, than has previously been possible.