Notes: 2D CAD USE CAE TEAMS SOLIDS 3D CAD OWN 2D OWN 3D MAKE MFR OWN CAE USE DFA OWN DFA CAE IN MFR = essentially all engineers have access to 2D CAD and few or no paper drawings are made except as informational output = the company uses some CAE in design = the engineers' workstations or terminals are networked together = the company uses cross-functional design teams = solid modelers are in at least limited use = 3D modelers (solids, wireframe, or surface) are in use = company uses 2D CAD software it wrote = company uses 3D CAD software it wrote = company makes key manufacturing equipment it uses = company uses CAE software it wrote = overlapping tasks design methodology is used = company has a long-range plan for development of advanced CAD and design methods = company uses a formal DFA methodology = company developed the DFA methodology itself artificial intelligence applications to design are in use or being developed = CAE is used in design of processes (molds, press dies) = the company has at least one integrated end-end CAD/CAE/CAM software system Figure 1. Distribution of computer technologies in 13 companies visited. INTEG The interesting entries are the occasional "1"'s that appear below the diagonal line. These indicate a company with limited capabilities that nonetheless is embarking on a technology that, in general, only much more capable companies have undertaken. Two of these are IHI, which has a definite longrange plan to integrate its computer design capabilities, and Mazak, which is world famous for using its own products in its unmanned, nonstop production lines. A separate graph, not reproduced here, shows that there is a strong though imperfect correlation between company size and the number of technologies observed. This is also not surprising and indicates that smaller companies need help if they are to gain enough capability to serve as qualified suppliers to the large ones. CAD CAD is naturally used by all companies visited for ordinary drafting in two dimensions. Obvious 3D applications like layout and interference checking have been mentioned above. However, most companies check interferences by eyeball inspection of 3D wireframe or 2D cross-section drawings. Few use solid models for this purpose. Toyota and Nissan both can simulate how a car interior looks; at Nissan the display is in stereo. The driver's field of view and windshield wiper clear areas (both subject to government regulation) can also be simulated. At Toyota, integrated CAD/CAM is used to design car interiors as well as exteriors, including use of numerical control (NC) machining to cut out full-size clay models of dashboards and shift lever consoles. The most interesting computer applications are those in which the external appearance of a product can be so realistically represented that physical prototypes are not needed. Examples include Toyota's work on cars and Sony's on videocameras, but there are many. Toyota's goes beyond CAM anything commercially available since it contains models of how its paints reflect light under different light and weather conditions in different cities in the world. Toyota has gone to great lengths in its home-grown surface modeling software to guarantee that the designers can easily manipulate the surfaces (always difficult in past methods) and can evaluate them by methods they are familiar with, such as simulating reflections of fluorescent tube lights. ing reflections of fluorescent tube lights. All the obvious applications are represented here, too. The main one is creation of NC cutter paths directly from CAD data. However, some companies complained that their commercial software does not support this well for sculptured surfaces. Sometimes there are errors converting the data from one form to another. In other cases, the surfaces contain unwanted undulations. Toyota and Nissan do not have surface undulation problems but may have some data conversion glitches. Other commercially available applications in wide use are mold flow simulations and some kinds of process planning. The Australian mold package called Moldflow is popular. Most companies seem to use SDRC's solid modeler as the front end for most CAM and CAE applications since SDRC resells a wide range of third-party software of this type and has taken care of the data conversion process. CAE Common applications in this category include FEM for stress, vibration, and heat flow problems, plus extensions thereof for complex turbulent flow studies. Commercial applications in use include NASTRAN, MARC, ADINA, and PAMCRASH (vehicle crash simulation). ADAMS is a kinematic simulator that is almost 20 years old but has recently come into wide use after SDRC attached it to its solid modeler. I was shown interesting simulations of how a washing machine rocks when the load is unbalanced, how a vacuum cleaner would track on its casters, and how a crane would react while swinging a heavy load. Ambitious fluid flow simulations are used on supercomputers to evaluate exterior car designs for drag and to see if manifolds and injectors provide uniform distribution of fuel particles. Curiously, in spite of the progress made reducing the noise of products, no one admitted having CAE for noise evaluation. Structural vibration and rotor dynamics were often used as proxies for machinery noise studies, but fluid noise is not being simulated. One interesting simulation was of active noise suppression of air conditioner noise. Spectral analysis is involved, I think, but not fluid turbulence. Preparation of FEM data has become a major bottleneck for everyone. Checking packed products for interference between parts, collision between moving parts, and access for parts and tools during assembly is becoming impossible to do without computers or costly prototypes, and huge data requirements make it hard to do by computer. Available research results that would speed up the process (octrees, for example) have apparently not been applied. Except for some limited Monte Carlo methods, no one has software for evaluating tolerances or predicting fitup of nonperfect geometries. The IBM Tokyo Research Laboratory plans to start such research, and some is underway at the University of Tokyo. No one agrees as to whether a statistical approach should be taken or a deterministic one. should be taken or a deterministic one. Statistical approaches sacrifice some accuracy in the highest precision studies, but deterministic approaches are threatened by combinatoric explosion. threatened by combinatoric explosion. Several companies perform failure modes and effects analyses (FMEA) on their products and one does so on manufacturing equipment, but there are apparently no computer tools for doing so as part of the design process. (Toyota painfully recorded the causes of robot failures for several years and, of robot failures for several years and, with cooperation from its two main robot vendors, succeeded in raising the mean time between failures from 3,000 hours to 30,000 hours. News of this has spread throughout Japan and all robot manufacturers are raising their products to this standard. Motorola conducted a similar study of its Seiko robots but I do not know any statistics.) Several companies acknowledge interest in design of human interfaces and one has some expert system work underway. Examples include how to position foot pedals and hand grips in position foot pedals and hand grips in cars and crane cabs. RESEARCH NEEDS AND CAD RESEARCH NEEDS AND CAD IMPROVEMENTS IDENTIFIED FROM COMPANY VISITS The following list comprises both what companies specifically asked for plus what I think they would use if it were available, based on what they said, complained about, reacted to, or implied. Several companies have launched improvements to their present capabilities but would not discuss them with me. Conventional CAD In this category are simply enhancements to existing capabilities that may require considerable effort. Better User Interfaces to 3D Design Systems. Designers are trained on 2D systems and have a hard time adjusting to 3D. No really natural user interface to 3D solid modeling via a 2D screen is in use. Even for skilled operators, construction of a complex 3D model takes a long time. The mechanical designers rightly argue that current surface modelers were designed for cosmetic exterior design of cars and cameras and are not suitable for mechanical parts. Exterior surfaces are, in fact, quite simple and the parts contain few features and have no "inside." Mechanical parts have substantial insides that contain a lot of detail, plus many complex features. This problem appears to be a major blockage to further penetration of solid modeling in Japan. Several companies felt that feature-based design might provide an avenue for attracting mechanical designers to 3D. Common Databases and Data Conversion. Everyone knows that current computing is a Babel of different languages, data conventions, and conversion protocols. It is a dirty problem, but it seriously stands in the way of rapid transfer of complex models from one CAD or CAE application to another. This, in turn, stands in the way of further integration of individual "islands of design automation" into complete design systems. Automatic Data Preparation for CAE. The case cited most frequently is preprocessing of FEM data sets. These are becoming more and more complex. Checking for errors and reasonableness can take a long time. One company cited a month as an example. Advanced CAD In this category are capabilities that are not available in any commercial CAD system, but creating them may be a near-term proposition, since one can imagine what means and information could be mustered. Practical Kinds of Feature Based Design. Providing catalog information, routine engineering calculations (such as how to design bearing seats or choose fasteners), and national and international standards (such as standard fit classes) should be relatively easy to implement. Readymade geometry backed up by parameterized models would provide a natural interface. The necessary calculations for bearing preload and life, for example, could also be stored for easy access. It would be a start on changing CAD from a "draftsman's interface" to an "engineer's interface." A little more challenging would be constraint-based rules such as enforcement of safety factors. Since Japanese designers are in most cases university graduates, this kind of CAD might be well received. Commercial U.S. software that offers or offered similar capabilities is that of Cognition and ICAD. Neither company seems to have made an impact in Japan. Geometric Dimensioning and Tolerancing. This was discussed above. Data Archiving and Retrieval of Past Designs. Some of this is being done now. It is unlikely that advanced data retrieval methods are being used, however. To do so would require developing ways of classifying designs, a decidedly nontrivial task that no one here is working on. Only rudimentary library search methods were seen in required size of a compensatory chamfer regular use. could readily be calculated. The opportunities for part consolidation could be identified based on kinds of material and joints between adjacent parts. It will be much more difficult to provide advice on whether a particular assembly action is "easy" or not. At the moment, companies rely on experienced people who usually do not use hard criteria to make their judgments. No one tries to predict whether a particu Cost Feedback to Designers. It has been said that most designers do not know the cost impact of what they design. Presenting such feedback requires cost analyses of processes plus ways of analyzing the design to determine its cost components. Determining process costs would require doing some preliminary process planning. Materials costs would require a straight-lar assembly task design would cause forward database. Vendor costs would require more than a database since estimates of negotiation results, discounts, shifting competition, and currency exchange rates would be needed. Broad-Based CAD That "Experienced Designs." This interesting term was used by a researcher at a company to mean a feature-based data file of previous designs that included proven process plans, statistical quality control results, process times and costs, customer feedback, and so on. That is, the data would represent actual experience, not just plans. This would be of more than historical interest if a way were available to extrapolate the experience in the database as the designer altered the design to suit a new requirement. DFM and DFA Advice to Designers. As mentioned above, companies want more than just design critics in their CAD systems. They want corrective advice. Providing this will likely require deep knowledge to be represented, although near-term implementations of some valuable kinds of feedback and advice could be easier. For example, a tolerance stackup analysis could be followed by advice on which elements in the chain contribute the most to the final error. The fatigue or carpal tunnel syndrome, or how hard it would be to retrieve a dropped part. Some companies use simulation to predict robot cycle time but none feed this information back to the designer in the hope of finding a design that will yield a shorter cycle. Ways to Use Partial Information. The essence of the overlapping tasks method is to launch designs based on partial information and assume values for information that is delayed. Companies want ways to categorize this information according to how important it is, when it is needed, and how the impact varies depending on how much the delayed information, once it arrives, deviates from what was assumed. Among the possible difficulties are wasting time in extra design iterations or creating grounds for product liability if incorrect assumptions are not eliminated before the design is released. Past data, experienced designs, sophisticated change notification methods, and standardized designs will likely be utilized to solve this problem. How to Automate in the Face of Diversity and Design Change. Only Nippondenso appears to have given deep thought to this problem. Most companies use people where more flexibility is needed than current automation can provide. Most researchers try to make smarter automation. However, Nippondenso has merely applied a form of sophisticated planning to such designs as the alternators and radiators. Nippondenso has extended the range of types or sizes it can handle in one automated system, but it is still totally restricted to the factors it planned and designed for. One or two dimensions can be varied within a fixed range, for example. No major product configuration change can be accommodated without the same equipment redesigns that any other company would face. This problem is one of many I could cite for which even existing research efforts would be insufficient. Despite the apparent difficulty, however, companies badly need this problem solved. MAIN THRUSTS IN UNIVERSITY RESEARCH I saw a great deal of very innovative university research in design during my stay. Some projects were motivated by discussions with industry while others were clearly the brain-children of the researchers. Topics covered below are: can represent various "aspects" of a design, such as the kinematic, thermal, design, such as the kinematic, thermal, or structural portions of the behavior of something. These meta-models are being constructed using qualitative physics, which provides symbolic representations of what are normally modeled by equations or logical constructions. Facts about nature (if a body with positive velocity experiences positive force, the velocity will increase) and about logical state changes in a system (if the wire melts, the coil will stop conducting electricity) can be expressed. These are stored in a library. A designer can construct a model of a physical system by describing geometry roughly and placing library objects in relation to each other. The computer augments this basic model with a number of side effects (the engineer describes the coil but the computer describes the heating effect that might lead to melting). When the model is complete, the computer can determine that the motor will turn continuously between some discrete angular states if it starts in the • Knowledge Representation - quali- right state. tative reasoning Two applications of this idea other than analyzing designs are underway. One is "self-maintenance machines" and the other is simulation of designers' actions while designing. The self-maintenance machine currently under study is a photocopier with sensors for copy density and other quality issues. The computer has a network model of causes and effects input by the user that tells what happens to each visible variable (copy density) as each internal variable (lamp brightness, lamp voltage) varies up or down. From this the computer can calculate a failure modes and effects analysis for certain failures. When a failure is observed, the computer reasons backwards to a set of possible causes and reasons forwards to determine a set of possible remedies. The remedy with the fewest side effects is chosen. Simulation of designers' actions is less well developed. It employs several logical techniques to follow a protocol recorded from a real designer and can imitate his reasoning from a first concept to the discovery that the concept will not work, to trying a second concept, and so on. However, this system has no physical knowledge and apparently only simulates the logic. Future work will connect this work with the meta-models more directly. Qualitative physics of the kind used in the self-maintenance machine has been pursued at length by chemical engineers for at least 5 years. The approach is rather good at imitating how people think and can store a great deal of partly structured information. Unlike equation models, this method can deal easily with logical state changes. However, it takes a lot of work to describe even a simple system, and the user contributes most of the real knowledge about how things relate to each other. Therefore, it remains to be seen how or when this approach will be able to do better than people. The promise is in the automatic generation of the side effects, giving the ability to tell the designer something he overlooked or might not have expected. Kyoto University. At the laboratory of Prof. Norio Okino, work is going on to create a hierarchical representation of physical things in the world of manufacturing. The approach is quite object oriented and consists of replicating an object called a modelon at every level. A modelon has a common memory and a set of processes describing its behavior connected to the memory. It also contains subprocesses that are modelons as well with the same general structure. In software, each modelon is a Unix process. Modelons operate independently of each other and seek to answer requests for action from higher levels while sending requests laterally and to lower levels. There is not much structure to |