Simulating the Growth of Poly-Crystal Snowflakes

Thomas Clune

Overview

Building on the previous success of a mono-crystal snowflake growth model, a team of NASA researchers has invented and implemented a parallel poly-crystal variant of the model to simulate simultaneous crystalline growth along lattices of different orientations. The scattering properties obtained from particles produced by the mono-crystal model have significantly advanced the consistency of quantitative snowfall estimates derived from active (radar) and passive (microwave radiometer) instruments. The poly-crystal model will further advance this capability.

Project Details

This project uses the new poly-crystal snowflake growth model to generate a variety of realistic crystals. The shapes of these simulated crystals serve as inputs to programs that calculate their scattering properties at various wavelengths of electromagnetic radiation. Finally, these scattering properties are used in radiative transfer calculations that simulate the responses of remote sensing instruments, thereby establishing the basis for quantitative snowfall estimates via remote sensing.

Parallelization of the poly-crystal snowflake growth model was unusually challenging because any regular domain decomposition can align with at most one lattice. Irregular boundaries for all the other lattices require nontrivial calculations for exchanging data between domains.

Results and Impact

Based on particles generated by the mono-crystal model and scattering properties obtained from it, we have built an extensive ice particle and scattering property database, OpenSSP, that has been invaluable in significantly improving quantitative snowfall estimates by spaceborne remote sensing. However, a considerable portion of the snow particles in nature is in poly-crystal form, which the mono-crystal model is incapable of generating. Snowfall remote sensing relies on accurately knowing the snowing particles’ scattering properties, which depend strongly on their geometric shapes. A poly-crystal snowflake growth model is thus necessary to make our database more complete and to provide further improvement in snowfall quantitative estimates.

Why HPC Matters

Generating realistic snowflakes requires large amounts of memory and computing time. Ideally, a simulation would use ~1 billion cells for each crystal lattice and compute the water vapor on a grid with ~1 trillion cells. A full simulation would require between 10 million and 100 million iterations and use days of computing time on an exascale platform. With existing resources, we expect to simulate crystals that are ~4x smaller in each dimension.

What’s Next

The next steps are to implement a simple adaptive mesh refinement (AMR) strategy for the diffusion of water vapor. Diffusion is by far the dominant computational cost of the model, and AMR could potentially allow us to reach our targeted resolution with current-generation computing platforms.