JAPAN'S ALPHA PROJECT The Alpha Project is a coordinated effort by Japanese industry (NEC, Fujitsu, Matsushita, Hitachi, NKK, Kobe, Toshiba, etc.) to develop a modern three-dimensional computer program for computational fluid dynamics. This report describes the background of the project and its current status. We also assess its future potential. INTRODUCTION In 1988 several Japanese companies came together to form the Association for Large Scale Fluid Dynamics Code. One of their research activities has been investigation of new software for fluid dynamics, administratively called the Alpha-Flow Project. Initially the association was composed of 15 companies that together provided about $10 million. About 15% was also loaned to the association by the Japan Key Technology Center, which is set up by the Ministry of International Trade and Industry (MITI) and funded by interest on income from the stock of NTT. (Such nonprofit foundations are part of Japan's "third sector" and play an important role in supporting science.) In addition, there were about 30 additional companies who have paid lesser amounts. The motivation behind the project is that mathematical modeling of fluid flow by computer, computational fluid dynamics (CFD), is a crucial part of large scale engineering simulation for nuclear reactors, aircraft design, wind flow around large structures, etc. Japanese industry makes heavy use of programs that perform simulation, and their utilization is bound to increase. For example, it is estimated that between 1% and 2% of Japanese construction industry sales revenue goes for research and development (R&D); the figure in by David K. Kahaner the United States is less than 0.05%, less than 1/20th as much. There is anecdotal evidence that the Japanese are using computational modeling for more long range projects than corresponding U.S. companies. U.S. firms tend to use simulation packages for immediate projects. One Western scientist told me that after the project is over, companies that he was familiar with often forget how to use the package and, in some cases, even forget that they have it. Most CFD programs are either proprietary and only provided as "executables" by commercial vendors or private such as those used at the Department of Energy laboratories, such as Livermore or Los Alamos. Japanese use of Western simulation programs varies from using them as black boxes, without detailed knowledge of the "inside" of these packages, to significant enhancements that have been made to some programs that were made available to them from national laboratories. CFD poses very severe difficulties in terms of the mathematical model and the details of its implementation. Programs that solve "real" problems usually contain tens of thousands of lines of (mostly) Fortran, written over many years by a heterogeneous collection of physicists, engineers, and computer scientists. Maintenance and documentation of these programs are often spotty; they are the typical "dusty decks." Recently, there has been some effort to modernize CFD programs, but it is rare that a completely new package is developed from first principles. A large modeling problem can consume endless hours of supercomputer time and generate enormous quantities of printed and graphical output. Using these programs can be tricky. The underlying model incorporates various assumptions and approximations; designers and users are always hoping to "add more physics," refine the numerical mesh, improve the numerical methods, etc. It is often difficult to validate a program and scientifically risky to use one without a great deal of expertise and/or consulting assistance. Thus a good program represents a very significant economic asset to its owners. Charge-back costs to users are high in order to recoup the substantial development expenses and to support ongoing research. Just as importantly, end user companies are often naturally nervous about being dependent on software that their engineers do not entirely understand. Most companies would rather have their engineers understand the working of computer programs that provide answers they need to rely on. Further, to get the most out of such complicated programs, and to avoid being misled by incorrect and incomplete answers, it is essential that users have some understanding of how these programs work. It seems perfectly reasonable that Japanese industry should want to develop their own in-house expertise. ALPHA PROJECT Three key people who have been instrumental in starting the Alpha Project are: Mr. Hiroshi Mizuta Special Assistant to the President Program Planning and Management Dept interested in thermal hydraulics but is also chairman of the committee that advises MITI about the nuclear industry. He also wrote a clear overview of the Alpha Project in the Proceedings of the International Conference on Supercomputing in Nuclear Applications (March 1989, pages A2:60-65). Takahashi specializes in large scale computing of nuclear models. The association gets input from end users, academic researchers, as well as its members. Basic research, mostly in industrial laboratories or at universities, helps to decide the fundamental mathematical, physical, and numerical National Space Development Agency techniques. Akiyama told me that there of Japan (NASDA) World Trade Center Bldg Minato-ku, Tokyo 106 Japan Fax: +81 3 3432-3969 Email: mizuta@nsavax.span.nasa.gov Prof. Mamoru Akiyama Bunkyo-ku, Tokyo 113 Japan Prof. Ryoichi Takahashi These three have known each other for many years (Mizuta and Akiyama since high school); the common denominator is their association with nuclear technology. Before Mizuta moved to NASDA he spent a number of years at PNC, a government funded organization looking into fast breeder reactors and the nuclear fuel cycle. Now he is responsible for planning issues related to computing. Akiyama is primarily are 50 to 100 people associated with the project in various ways. A contract software company [Fuji Research Institute Corp. (FRIC)] does the program implementations. Currently, there are between 30 and 40 programmers and computer analysts working on the project at FRIC. Dr. Hideaki Koiki Minato-ku, Tokyo 108 Japan is the technical leader of the FRIC group and coordinated a recent visit of mine to their institute. A biannual meeting gives researchers an opportunity to show their latest results and for the Alpha Project developers to present their progress. A Proceedings is issued in Japanese, Proceedings is issued in Japanese, although at the March 1990 meeting Dr. Eric Hollnagel (Computer Resources International, Denmark) gave a very general discussion of the potential usefulness of artificial intelligence (AI) and human computer interaction when coupled with the power of a supercomputer. According to Mizuta, a key aspect of the project is that Japanese scientists want access to the insides of complicated CFD programs in order to verify and understand exactly what they are doing. My own experience in other general mathematical software is that this is not always necessary in really well designed programs as long as the physical and mathematical models have been carefully worked out. The apparent need here probably reflects both the string and thumbtack construction of some CFD programs and the fact that the models are usually not entirely satisfactory. Mizuta claims that this phase of the Alpha Project has focused on the physical model to be solved and the numerical techniques to be utilized. The project has been developing single-phase, three-dimensional solvers, using finite differences only, for a variety of physical situations. Finite elements are not being considered at this time as they are thought to be too slow. Approximately 100,000 to 150,000 lines of Fortran have been written to support the solver part of Alpha, organized as follows. Solver for incompressible flows (A module for Cartesian and cylindrical coordinates and one for boundary fitted coordinates. There is also a module for free surfaces.) Solver for heat transfer in solids amounts of material are mixed with the Solver for incompressible flows with multiple free boundaries Solver for chemically reacting flows (low velocity reaction flow) Solver for compressible flows (for high velocity viscous flow) The 15 companies that support the association will be permitted to have copies of all the source programs; other members will get "executables" only. However, during the development phases, participating researchers will have access to some parts of the source programs, presumably for them to study and test. It is expected that each company will modify the programs either for its own needs or to optimize sections for particular hardware. This phase of the project is focusing on the scientific and engineering aspects in the design of a CFD package, i.e., the model. This is natural given the background of the research team and the engineering expertise of the supporting companies. However, there are other issues that are also very important such as the following. • Modularity and portability. • Hardware independence and optimization for different computers. degree of portability. It is more difficult to also get highly optimized software, especially over a wide spectrum of machine architectures. Adding intelligence to numerical programs is a new, major research area and there is significant work at Purdue, Rutgers, and other places. In the West, especially in the national laboratories, petrochemical companies, and other places chemical companies, and other places where CFD programs are in heavy use, there are major efforts to study the implications of new machines, such as parallel and distributed computers. The Alpha Project has not looked into this aspect. I was surprised to learn that Mizuta, and presumably the other Alpha-Flow Project scientists, was not well connected into the community of Western researchers who have been actively studying most of these topics, e.g., participation in the Expert Systems in • Maintenance framework for large Numerical Computation Conference (huge) source programs. • Documentation. • User interface. • Incorporation of expert systems and AI for advice on problem setup, interpretation, and management of results. • Adaption of the programs to advanced computer architectures, such as parallel computers. Mizuta claims that they want to tackle these issues and lists them as among their most important goals. Nevertheless, he admits that they haven't done much in any of these directions and that the concept of using expert systems, etc., is easier in principle than in practice. Some items in the list above will be easier to accomplish than others. For example, we already know how to design numerical software with interchangeable parts and to plan for a high that is held each year at Purdue. On the other hand, papers in the recent AlphaFlow Symposium concerning the manmachine interface are full of references to current research in the West. At the moment there is also almost no participation from scientists in the West, although Mizuta and others are known in the Western CFD community and have recently returned from a trip to several U.S. laboratories. When I asked about this, Mizuta explained that at the beginning the Japanese didn't feel that they had anything to contribute and wanted to bring their own expertise up to a credible level first. After Mizuta said that the emphasis was on the solvers, I was amazed to learn during my visit to Fuji that the user interface has not been neglected; almost 300,000 lines of C have been written in the man-machine section. These include Control module • Module for input generation • Module for interactive postprocessing • AI module (expert system) • Module for data management I was shown a few brief demonstrations of the system, which looked quite powerful as well as flexible, but had no opportunity to study any parts in detail. The design philosophy is to write the Fortran in the form of portable and modular machine independent subroutines and the man-machine interface in similarly portable and modular form. The latter would be expected to run under X-Windows on a Unix workstation and the Fortran on a remote (super) computer. Between these parts would be a "gateway" module that would contain all the machine specific details of the implementation. There is a complicated flow of data between modules, but this is standardized by use of a standard file format and a conversion description standard, which appear to be well thought out. Mizuta and Akiyama have stated that they want international cooperation on the project. Akiyama wrote that "we solicit cooperation from all quarters." I asked what form this might take, as the only people who are entitled to copies of the source programs are the supporting industries. Initially, there was some ambiguity about this. Mizuta thought that Western scientists might want to participate in order to be able to compare their programs against Alpha's, but that certainly would not be enough to appeal to me. When I asked Akiyama the same question he admitted that he really meant cooperation from additional Japanese companies who may want to contribute financially. As far as other researchers are concerned, he thought that some neutral information could be shared with them, but that this would have to be discussed with the steering committee, as was my request to visit FRIC in order to see a demonstration. Subsequently, though, Mizuta explained that he expected to be able to offer Western researchers (academic, not commercial) copies of source modules. This would be a perfectly fair exchange to obtain cooperation and I hope that this is in fact what will happen. In fact, I hope to convince Mizuta to demonstrate Alpha during this year's Supercomputing 91 meeting, in Albuquerque, New Mexico. I was told several times that this is not really a government project. In fact, one of Alpha's accomplishments has been to bring together distinct Japanese companies to build the software foundation that they can each particularize. Industrial partners get source programs and can then make modifications to suit their own needs. The Alpha group has no responsibility to maintain these modified programs. At the current stage of the project, a good deal of program structure has been developed, combining computational modules with an intelligent front end through the gateway. According to the project schedule, most of the documentation work is being done this year. All the documentation that I saw was in Japanese, which would definitely be an impediment to Western cooperation, but is probably essential to effective use of the software here. It was emphasized to me that the two goals are (1) developing vectorizable programs and (2) making the problem setup easy to use. Using the absolutely best algorithm, or at least fine tuning it, was not considered so important. The project participants are hard at work verifying the models by running them on about 20 major test problems, most of which are completed. These included such things as boundary fitted coordinates. The "boundary fitted coordinate❞ technique (sometimes referred to as "body" instead of "boundary") is quite similar to low order finite element methods. Both methods use nonrectangular meshes. I was shown several slides of flow patterns from Alpha side by side with those from other computations or experiments and the agreement looked excellent. I remarked to Akiyama that these pictures didn't give any information about how successful the project was at achieving either (1) or (2) above, but the demonstrations were persuasive with respect to the second. For example, in dealing with incompressible flow the AI module has about 100 rules. There is also an online manual. My opinion is that we should not judge Alpha only by what it has produced thus far. Although it may have made major strides in educating and bringing together Japanese scientists, from a global research perspective its accomplishments seem modest. A great deal of their development has duplicated, i.e., either copied or adapted, existing published work. For example, the "multiple free boundary" work is based on SOLA-VOF, developed in the United States. A recent visit to the vendor displays at Supercomputing Japan '91 shows many Western engineering analysis software packages with very graphically oriented interfaces. Alpha has only been in existence about 3 years, 75% of its funding schedule. Considering that it started with a clean sheet, this is not much time when compared to the efforts that have occurred developing CFD packages elsewhere. Hence, it is unlikely that the Alpha-Flow Project is yet competitive with the best CFD packages either from the United States or from Europe. However, Mizuta is very clear that the code name Alpha represents the first letter of the alphabet, and that he firmly intends to see Beta, Gamma, even Omega. He claims that he has strong support from Japanese industry and that many new companies want "in." Further, there is interest in using the framework to develop software for molecular computations and new materials design. Akiyama is less interested in such grand goals. He admitted to me that the Beta Project will probably focus only on two-phase flow, which is of less general interest (few of the Japanese auto companies are likely to participate), and hence will need only about half its current funding ($5 million). The Gamma Project will probably involve Monte Carlo or stochastic methods development. For his own research, Akiyama is more interested in building a very advanced simulator (using a massively parallel computer and sophisticated software) for major nuclear accidents that he hopes will never occur. The educational aspect of this project should not be ignored. Mizuta emphasized that one goal was to involve universities--the idea being the establishment of knowledge centers that will continue to produce people (students) with knowledge of computational fluid dynamics. About a half dozen university groups have been funded under the Alpha Project. There has been a trend in the United States (possibly elsewhere) to use commercial/private software without understanding how the software functions or what assumptions and approximations are made in the software. This is a very bad trend; perhaps the Alpha Project is partly a reaction to this situation. The Alpha Project seems to indicate that the Japanese are taking a long range view by trying to increase the number of engineers and scientists who are familiar with large scale numerical simulation techniques. This approach will eventually put them in a much better position to exploit computational modeling for a wide range of applications. It would be useful if more U.S. companies thought the same way. A final note. I asked if there was interest from the Japanese aircraft industry, as CFD is such an important aspect of aircraft design. Akiyama said no, mostly because in Japan that industry is concentrated in a few companies such as Mitsubishi Heavy Industries that already have extensive software of their own and see no need to support research outside their own organization. SUGGESTION Scientists outside Japan need to view the project as a first step and as an educational tool. At the same time Alpha's steering committee needs to clarify the question of access to information. Perhaps Western input could be helpful in forming these ideas. |