FrontFlow/red-HPC is a general-purpose software package for thermofluidic analysis that uses unstructured meshes (like that shown in the figure below) to solve heat-flow equations. Whereas typical (structured) meshes are often cubical, unstructured meshes use pyramids and other polyhedra to yield efficient representations of complicated shapes—with the additional advantage of allowing arbitrary local refinement or coarsening of the mesh granularity, This gives researchers the flexibility to analyze and visualize specific regions of model geometries in great detail while keeping the mesh resolution in other regions as coarse as possible to maximize computational efficiency. Today, FrontFlow/red-HPC is widely used to analyze problems involving large-scale and complex physical phenomena, including automobile aerodynamics, combustion in gas turbines, and wind flows in urban environments.
FrontFlow/red-HPC is a software package for thermofluidic analysis that solves heat-flow equations using unstructured meshes, which allow efficient representation of complicated shapes.
FrontFlow/red-HPC is a general-purpose software package for thermofluidic analysis developed primarily by the RIKEN Center for Computational Science (R-CCS) and researchers at Hokkaido University and Kobe University. It is a supercomputer-optimized version of FrontFlow/red Version 3.1, a thermofluidic analysis package originally developed within the Revolutionary Simulation Software project of Japan’s Ministry of Education, Culture, Sports, Science and Technology and now available to the public from Hokkaido University. FrontFlow/red-HPC is designed to support calculations on systems spanning a broad range of computational capabilities—from standard desktop workstations to next-generation supercomputers—and is currently used by many scientists and industrial engineers specializing in design engineering and product manufacturing. In the past, the most widely used software packages for thermofluidic analysis were almost all developed outside Japan. It was to address this situation—to produce a tool, made entirely in Japan, capable of responding more quickly and more flexibly to requests from corporate users for additional features and modifications—that the original FrontFlow/red package was developed. By optimizing this package for execution on supercomputers, FrontFlow/red HPC enables simulations of enormous sizes that could not be carried out in existing commercial packages.
Kei Ambo of Honda R&D Co., Ltd. used FrontFlow/red-HPC (FFR) and Japan’s K-computer supercomputer to study wind flow around the side mirrors of automobiles—a key source of audible noise arising during vehicle motion—via high-precision simulations.
In the past, R&D initiatives by vehicle manufacturers to optimize wind flows around moving automobiles were typically carried out using wind tunnels. More recently, the use of supercomputer simulations—based on methods of computational fluid dynamics (CFD)—has become increasingly common for studying wind flows, starting at the earliest stages of the development process.
However, wind-flow phenomena—particularly those responsible for audible noise—are notoriously complicated, and the goal of achieving complete reproductions of complex wind-flow patterns has posed severe difficulties for CFD calculations based on commercial software tools. This has motivated the use of wind-tunnel experiments in conjunction with numerical simulations, but wind-tunnel studies require accurate models of test bodies, which are not only costly and time-consuming to build, but also require significant effort to correct and update. Moreover, CFD simulations conducted at low numerical precision may indeed predict motion in entirely erroneous directions, substantially increasing the likelihood that major adjustments will be required at subsequent stages of wind-tunnel testing. Thus, the ability to perform high-precision simulations at early stages of the development process is a key enabler of accelerated schedules and reduced costs.
These were the challenges facing Kei Ambo when he reached out to Professor Makoto Tsubokura, leader of R-CCS’s Complex Phenomena Unified Simulation Research Team and a lead developer of the FFR HPC suite, for ideas and guidance. Ambo believed that wind vortices produced near automobile side mirrors—and their eventual dissipation—were major causes of audible noise. On the advice of Professor Tsubokura, he narrowed the focus of his attention to a single simulation target—wind flow near side mirrors—and took up the challenge of performing high-precision FFR calculations on the K-computer.
“Actually, for simulations targeting only specific localized regions, like the vicinity of side mirrors, we were able to make some progress with just the modest in-house computational resources we had available at Honda Research,” explains Ambo. “However, without high-precision simulations of wind-flow trajectories throughout the entire vehicle region, it’s impossible to identify the origin of the flow patterns responsible for creating vortices near side mirrors, or to trace how those wind flows interfere with the vehicle chassis after they have produced vortices, and ultimately there is simply no hope of developing a correct understanding of the origins of the noise. And our in-house resources were not sufficient to allow us to tackle those problems. Another problem was that the mesh granularity—that is, the spatial resolution—of commercial CFD packages was too coarse for our needs. So instead we decided to run FFR simulations using the K-Computer.”
Ambo’s choice of FFR as a simulation tool was motivated by multiple factors. “First, we knew the package was highly reliable and had an impressive track record of successful deployment on the K-Computer,” he explains. “Of course, it was also crucially important that FFR could perform large-scale computations at the high resolutions we needed, and it was very reassuring to receive early-stage advice from Professor Tsubokura—who, after all, developed FFR HPC himself—on how to set appropriate mesh resolutions for our simulations.”
Of the many conveniences offered by FFR, the factor that proved decisive for Ambo was the freedom to adjust the fineness (spatial resolution) of the computational mesh precisely as appropriate for the computational objective in question. By using a relatively coarse mesh for most of the automobile—but switching to finer-grained, higher-resolution meshing in the region of interest near the side mirror—Ambo was able to obtain results with high computational efficiency.
For comparison, analyzing the entire automobile at the high resolution needed for the side-mirror regions would entail meshes containing as many as several billion elements, exceeding the capabilities of commercial software tools. In practice, FFR simulations executed on the K-computer allowed Ambo to achieve high-precision visualization not only of wind flows throughout the entire region of automobiles, but also of the production and dissipation of wind-flow vortices near side mirrors.
Results of FFR simulations performed on the K-Computer. The computations produce highly accurate visualizations of both (a) wind flow throughout the entire spatial region of the automobile, and (b) the fine-grained details of wind flow in the vicinity of the vehicle’s side mirrors. The simulation results agree with the results of wind-tunnel experiments.
(Figure provided by Kei Ambo, Practical Applications of Wall-Resolved LES for Mirror Noise Analysis, PPT 3/15)
The use of finer meshing in side-mirror regions allows high-precision analysis of wind flows near side mirrors.
(Figure provided by Kei Ambo, Practical Applications of Wall-Resolved LES for Mirror Noise Analysis, PPT 11/15)
Ambo’s computations were accompanied by wind-tunnel experiments in which pressure sensors and other instruments were attached to the surfaces of vehicle bodies and side mirrors to make detailed measurements of wind flow separations.. The results of these wind-tunnel experiments agreed with computational results produced by FFR, confirming the extremely high reliability of FFR simulations. Moreover, whereas wind-tunnel experiments only lend insight into local pressure variations, Ambo’s simulations produced high-precision visualizations of wind-flow patterns not only near side mirrors but throughout the entire vicinity of a vehicle, yielding a deeper understanding of the mechanisms responsible for audible noise generation. Going forward, Ambo and his colleagues plan to use these simulation results as a basis for further optimization of automobile shapes to reduce audible noise and improve air resistance.
“It’s not easy for any one individual company to harness the full computational power of the K-computer,” notes Ambo, reflecting on the successful trajectory of his research project. “So our success in this project was very much tied to our ability to move forward in tandem with Professor Tsubokura and other RIKEN researchers and university professors, with whom we discussed every aspect of the process—from basic FFR usage to advanced computational techniques. From the response we’ve received, I think people recognize that we’ve taken a big first step into the future of automobile development.”
At present, Ambo’s research group participates in the Consortium to Enable Next-Generation Automobiles, an HPC-centered organization led primarily by R-CCS that is currently working to identify appropriate projects for the Fugaku supercomputer, the successor of the K-computer.
“In today’s world, supercomputers have become essential tools for product design in a wide range of fields, including automobiles,” notes Ambo. “We’re looking forward to using the Fugaku—and forging strong ties between industry and universities as well as research institutions—to stimulate the development of products to pave the way to a brighter future.”
As Kei Ambo’s story illustrates, FrontFlow/red-HPC not only enables efficient massively-parallel supercomputer studies of thermofluidics and other constantly changing physical phenomena, but also facilitates the addition of new analytical models and new methods for users to explore and test, whether for purposes of basic research or product design. R-CCS looks forward to offering extensive technical support for this tool—not only for the K-computer, but also for its successor, the Fugaku supercomputer—to users spanning a wide range of industries, supporting and empowering all sectors of Japanese manufacturing.
Interviewed: November 2019
