Integrated Machine Learning for Combustion in HPC Applications

Moses, Adam (US Naval Research Lab)

Sharma, Alisha


The massive scales required for physics modeling, such as jet engine simulation and advanced climate modeling, have made them ready targets for the largest HPC systems and most efficient software codes. As HPC has leapt further ahead in compute power thanks to GPUs and multi-core systems, the software code efficiency side has not as been lucky, often facing the same performance bottlenecks within their codes that has kept the continued search for new approaches alive.

To cite one particular such issue, the simulation of chemically reacting flows, a critical component to many physics, science, and engineering problems, remains a massive bottleneck wherever it is used. The timescale difference in modeling the flow portion of a simulation versus the chemistry are often several orders of magnitude different, thus dragging down the model performance as it spends nearly all its time in modeling just the chemistry. In certain combustion simulations the percentage of compute time spent on chemistry can be as much as 80-90% of the total model [1], rendering some advanced combustion modeling computationally prohibitive if not impossible.

This presentation will explore alternative solutions to the traditional chemistry solvers used as part of these models, with an emphasis on how machine-learning based approaches may offer break-through performance gains making the previously prohibitive models now within reach. Techniques such as time-splitting the chemical modeling from convection and diffusion will be discussed, as well as a range of machine learning model types and their various levels of effectiveness, both in performance and accuracy. This will all be discussed in the realm of HPC computing and the specific integration of some of these approaches as part of a major HPC-hosted NRL combustion code.

[1] Lu, T., and Law, C. K., "Toward Accommodating Realistic Fuel Chemistry in Large-Scale Computations," Progress in Energy and Combustion Science, Vol. 35, No. 2, 2009, pp. 192–215. doi: