Andrew Shao, PhD

Andrew Shao, PhD

HPC&AI researcher, oceanographer, part-time farmer

Projects

Neural networks for accelerating numerics

Numerical methods are often limited by a CFL condition that limits the ability to take long timesteps. In collaboration with scientists at GFDL, we are investigating ways of using neural networks to predict ‘unroll’ a timeloop for a finite volume model and predict fluxes over multiple timesteps at once. In this way, we are mimicking a multi-dimensional, high CFL algorithm while still maintaining properties of conservation. The following image shows preliminary results using MOM6

Online, neural network inference for sub grid-scale physics

Neural networks trained on high-resolution data can ‘learn’ the relationships between the mean flow state variables and the turbulent terms. With scientists at NCAR and M2Lines, we have demonstrated the first implementations of using these neural networks online in climate-scale, ocean simulations. The image below is from a simulation performed on Frontier which we subseqently scaled to ~20,000 CPUs and ~5,000 AMD GPUs.

AI techniques applied to computational fluid dynamics

OpenFOAM, a widely-used, open-source CFD solver, has a wide user base. With the Data-driven Modeling Special Interest group, we explored a variety of cases where we could inject AI (and related) techniques to classic CFD problems. To give others in the community a starting point, we developed three cases: 1) training a CNN online and then using it to predict the motion of the mesh in the case of a moving airfoil (see the image below, red indicates that the neural network results in better mesh quality than the default algorithm) 2) implementing an online, distributed, PCA algorithm to calculate turbulent eigenmodes for a turbulent cylinder in a box 3) using Bayesian optimization to tune turbulence parameters for the Pitz and Daily combustion case.