ENVIRONMENT AND ENERGY

Scientists of the RWTH Aachen University have carried out a peta-scale direct numerical simulation (DNS) of a temporally evolving lean premixed methane/air jet flame. The DNS is intented to closely mimic gas turbine combustion and can be regarded as an idealized representation of a premixed flame element inside a jet burner. To realize high resolution of flame and turbulence and to obtain converged statistics, the simulation domain was discretized with almost three billion grid points which together with the chemistry model resulted in nearly 100 billion degrees of freedom.

In the past decades, the world energy consumption increased rapidly and is expected to further grow significantly in the future. As fossil fuel remains the dominant energy source, combustion devices remain the major technology for the conversion of chemically bound energy into electrical, mechanical, or thermal energy. In combustion processes of hydrocarbon-containing fuels, pollutants, such as unburned hydrocarbons (UHC), carbon monoxide (CO), oxides of nitrogen, collectively referred to as NO_x, and particulates in the form of soot, are formed and emitted into the atmosphere.

The severe environmental effects of increased ambient NO_x concentrations have led to stringent emission laws and as a result, strong efforts are made to develop modern low emission combustion systems. A promising tool to accomplish this goal is the design and control of these systems using computational fluid dynamics (CFD), which in turn requires predictive models for NO formation.

Combustion modeling is complex, because of its multi-scale nature and the strong non-linearities of chemical source terms. Modeling depends on data, but experimentally measuring spatially and temporally resolved species concentrations in turbulent flames is extremely challenging. The so-called direct numerical simulation technique (DNS) has been established as a powerful tool in combustion science to complement theory and experiment, and is as a research area becoming ever more important. In DNS, the governing equations describing fluid flow and chemistry are solved without any modeling assumptions, thereby resolving turbulent motions and combustion in time and space ranging from the smallest scales associated with the fastest reactions up to largest scales of the turbulent motions.

Because of the wide ranges of length and time scales in turbulent flows, these simulations can be extremely costly, but the rapid advancement of supercomputing has recently enabled a series of interesting DNS studies with practical relevance. Although DNS are presently not feasible for realistic combustion systems, the richness of DNS even for simpler cases, the level of detail, the availability of all desired quantities at all locations in the flow, and the well-defined and well-known details of boundary conditions encourage the use of DNS data for model development. Moreover, spatially and temporally resolved high quality data for key quantities like reaction rates including species and temperature distributions, high order moments, or correlation functions are challenging or impossible to obtain experimentally. Yet, these are available in DNS and can enable very systematic analyses as well as model development and validation.

Within the present Gauss project awarded with computing time on HPC system SuperMUC of the Leibniz Supercomputing Centre in Garching/Munich, a peta-scale DNS of a temporally evolving lean premixed methane/air jet flame was carried out at the Institute for Combustion Technology at RWTH Aachen University. This DNS is intended to closely mimic gas turbine combustion and can be regarded as an idealized representation of a premixed flame element inside a jet burner. In contrast to many other DNS, this configuration features a very detailed description of oxidation and pollutant formation chemistry and a mean shear driving the turbulence so that it does not rely on artificial turbulence forcing.

In order to realize high resolution of flame and turbulence and to obtain converged statistics the simulation domain is discretized with almost three billion grid points, which together with the chemistry model results in nearly 100 billion degrees of freedom. Following prior successful scaling tests, the DNS was performed on the SuperMUC supercomputer at the LRZ Munich on 65,536 cores and required around 40 million CPU hours in total. An instantaneous snapshot of the simulation shortly after the flames arrive in the shear layer region is shown in Figure 1.