From: AAAI-86 Proceedings. Copyright ©1986, AAAI (www.aaai.org). All rights reserved. ARTIFICIAL INTELLIGENCE A MECHANICAL AND DESIGN: ENGINEERING VIEW John R. Dixon Department of Mechanical Engineering University of Massachusetts Amherst, MA., 01003 ABSTRACT Most AI research into design has been based on or directed to the electrical circuit domain. This paper presents a mechanical engineer's view. parts and products differs Design of mechanical in several from design of electrical circuits fundamental ways: materials selection, sensitivity to manufacturing issues, coupling of non-modularity, high form and of 3-D function, and especially the role geometry. These differences, and the role of design, are discussed. A analysis in mechanical for design is also presented based on the model basically iterative nature of the design process. A brief summary of the related research at the of Massachusetts into application of University AI to mechanical design is included. contribute important perspectives to mechanical design, but also that studies of mechanical design can enrich AI view of design the generally. goal of this paper is contribute a The mechanical engineer's view (and model) of design to the AI discussion of design. The view of design and the model presented here are based on the author's engineering experience and also on recent research at the University of Massachusetts into the possible application of AI to mechanical design automation (Dixon and Simmons, 1984, 1985; Dixon et al, 1984, 1985; Simmons and Dixon, 1986; Kulkarni et Dym , 1985; 1985; Vaghul et al, 1985; Howe et al, 1986; al, Libardi et al, 1986; Luby et al, 1986). MECHANICAL INTRODUCTION "The proper study of mankind of design." (Simon, 1969) is the science Probably so, but since 1969 the subject has hardly been a major theme of AI research. The Handbook of AI barely discusses it. Now, however, AI has rediscovered design, especially engineering design (Mostow, 1985). There has, of course, been some attention paid to design by AI researchers over the intervening years. See, for example, Sussman, 1977, 1978; McDermott, 1978, 1981, 1982; deKleer and Sussman, 1978; Bennett and Englemore, 1979; Stefik et al, 1982, Brown, H. et al, 1983, Mitchell et al, 1983 ; Steinberg and Mitchell, 1985). The vast majority of this research is directed to or based on the domain of circuit design. Unfortunately, one cannot obtain an adequate general model of design, even engineering design, from the circuit domain alone. Recently some AI researchers have become interested in mechanical engineering design (Brown, D. C., and Chandrasakeran, 1983, 1984, 1985, 1986; Brown, D. C., 1985a, 1985b, 1985c; Brown, D. C. and Breau, 1986; Popplestone, 1984; Mittal, 1985). These initiatives into the application of AI in the mechanical domain illustrate very well not only that AI can 872 / ENGINEERING DESIGN COMPARED WITH CIRCUIT DESIGN The Rutgers Workshop on which Mostow's article is based consisted of mostly electrical engineers and AI researchers. A workshop involving both mechanical engineers and AI researchers has since been held, with quite different results et al, 1985). Though it (Cole may be possible that at some (high) level of abstraction the design processes in the circuit and mechanical domains are similar, design in the domains is so two very different at the practicing level not likely to that we are discover the generalizations until we understand more fully in the two domains design much separately. are some important ways that mechanical There design and circuit design are alike. Both are are bodies of engineering, and thus there heuristics (e.g., "good design practices") as reasoning from basic science well as much principles. Both domains also use analysis degrees of sophistication to methods of various support design. Despite the similarities, there are four issues in mechanical design that make it fundamentally different from circuit design. (1) the wide spectrum of material These are: available to mechanical designers; (2) choices the critical role and often hyper-sensitive concerns on mechanical effect of manufacturing (3) the non-modularity of mechanical designs; and (4) the intimate role of complex 3-D designs; geometry in mechanical design. Perhaps the most significant of these between mechanical design and circuit differences design is in the role of 3-D geometry. In material specification, a addition to a mechanical design & a description of a 3-D Is. Though Mostow's paper does not even geometry. geometry as an issue in developing better mention models of the design process, it is a critical design. Moreover, central issue in mechanical manufacturing, and function are highly geometry, coupled in mechanical design (Rinderle, 1986). the natural language of Geometry is Mechanical designers "think" mechanical design. There is no mathematical geometry. in 3-D between the mental image or concept of connection a design and its 3-D visual representation in a drawing. Whether on paper or on a CAD terminal, designers sketch, erase, and sketch mechanical again. A mechanical design is seldom the solution equations; it is instead to a set of constrained a representation -- a drawing -- of a 3-D object. It may be noted that three of the points of -listed above materials, difference and geometry -- each constitutes manufacturing, of knowledge. There are also a huge body interactions among these topics. important this is knowledge that is not naturally Moreover, represented by equations. shall materials be represented so that How can reason about them? How shall design programs processes and machine tools be manufacturing that design programs can be represented so that consider the properly written designs as well as plan the manufacturability of production process? And especially, how shall the a design be represented so that geometry of reason about that geometry, design programs can function and it to both relating are issues that These manufacturability? circuit mechanical design from distinguish and which must be studied in addition to design, to develop adequate common issues if we are process the engineering design models of generally. THE ROLE OF ANALYSIS IN MECHANICAL Sometimes parts of a design problem can be formulated as optimization problems, and in these optimization methods are instances, sub-problem is a whole, real Seldom, however, useful. very expressable as a mechanical design problem optimization problem. The designer's mathematical lot will never be such an easy one. Hopes persist that design can somehow be done directly by some form of "analysis", that is, The essential iteration. I doubt it. without design nature of design -- at least mechanical -to learn to iterative. Therefore, is construct programs that can design, we need to learn first how to guide' iterative processes intelligently so that acceptable designs can be produced efficiently. MODELLING THE MECHANICAL DESIGN PROCESS We need a model or models of the design to formulate design problems, process in order represent design knowledge, and to acquire and develop design inference engines. Most mechanical that the basic nature of design accept engineers are variations in However, there is iterative. exact nature of how a design problem is to be the and just what the process is in formulated, detail. We view design as a hierarchy of nested of (1) decomposition and processes iterative and specification redecomposition, (2) redesign. and and (3) design respecification, the decomposition aspect of this Figure 1 shows A That is, node tree. simple model as a complex problem to be solved. Nodes designates a D represent a decomposition of the B, C, and problem into sub-problems, and so on. Decomposition continues until sub-problem DESIGN Before discussing a model of engineering design from a mechanical view, it is necessary to in mechanical the role of analysis consider equations and mechanical design, design. In the design process analyses enter mathematical only after a trial design has been developed. Analyses are used to simulate or predict the performance of a prospective design for its intended use. It is a fundamental intellectual error, therefore, to believe that analyses, by themselves, can produce designs. Analysis supports and assists design by providing useful information about how a proposed design will perform, but an analysis does not produce a design since there must already be a design in order for an analysis to be performed. Figure I. Decomposition into Sub-Problems. APPLICATIONS / 873 and complexity is reduced to the point where size the problem can be managed intellectually without Usually this is a problem further decomposition. design small number of relatively with a three to ten or twelve. These perhaps variables, (i.e., designs solved then sub-problems are developed) by a process we call "redesign". We this redesign process in some shall now discuss then return to the decomposition part detail and of the model. model of design is The redesign Redesign. is specified in in Figure 2. A problem shown An initial design of problem parameters. terms generates an initial trial design. procedure the final design. If it is -(Seldom is this initial design procedure is so that is, if the acceptable designs, then it produces good that interest further domain is of no the exist in some Such procedures intellectually. are not concerned with such well domains but we trial design is Next the understood problems.) determine its analyzed to evaluated; that is, performance in terms of performance expected include cost, function, and parameters that may issues. Then a decision is made manufacturability design's acceptability. (Simon the trial as to coined the word "satisficing" for this concept.) If the the task is design is acceptable, complete. If not, the design is redesigned and re-evaluated, and so on iteratively. Often in practice, redesign ultimately fails and the process returns to the initial step with the information that the problem requirements need to be changed, usually relaxed, in some manner. all making and hard. the tradeoffs is excruciatingly sub-problems that need no for In summary, further decomposition to be managed, design is It is an important iterative redesign. done by intelligent process. Our work with Independent" Redesign. "Domain redesign has progressed to the point where we prototype of a program (called have a working that designs redesign class problems in Dominic) et al, 1986). Dominic has several domains (Howe successfully to date in four distinctly designed tests are in other and different domains, progress. hill-climbing essentially a Dominic is this sense it is similar to algorithm. In It differs, however, in otimization techniques. is guided by heuristic First, it two respects. obtained from a domain expert. domain knowledge its input format is a kind of low level Second, design problem natural to more language formal mathematics of than the formulation optimization methods. a very strong nor a very is neither Dominic method. It lies in-between weak problem solving extremes. It works "generally", but only on these of design problems. It makes use of a sub-class knowledge, but also possesses some general domain of how to go about solving problems in knowledge It may be that such strong/weak class. its can be a practical way to apply AI in methods design. A great deal of mechanical design is done by this iterative redesign process. Despite its importance, however, it appears to me that many AI researchers tend to belittle or at least redesign. There ignore, are perhaps two reasons this. One stems from an association of for with "generate and test" (Lindsay et redesign al., 1980); the other from an association of "debugging of redesign with with process of (Sussman, 1977). almost right" solutions 1 1 GET SPECIFICATIONS and test & a rather poor process Generate for design if what one means is to generate and exhaustively. But this is not at all randomly In redesign, the what is meant by redesign. for and the reasons analysis results used heuristically to guide are unacceptability the changes made in the design. phrase "debugging of almost right The something rather solutions" also suggests design is already trivial. After all, if the "almost" right, the really hard intellectual work already been done, hasn't it? In mechanical has designing a design, usually not. When refrigerator, one does not begin with the design of a tractor. When designing an automobile drive shaft, one does not start with an initial design In all but the most novel and of a fender. unstructured design situations (say, starting out to design the first Polaroid camera), getting an "almost" right solution is relatively easy, whereas getting rid of all the unacceptabilities 874 / ENGINEERING I REDESIGN fails Figure 2. The Redesign Model of Design. the specification changes what next round of should be. In other words, the sub-problems can request a change that will affect the other sub-problems, but they cannot impose it. Only the module in a position to tradeoff competing power to impose or "propagate" requests has the constraints on sub-problems. This is as it must be to retain control of this very complex process. Decomposition. We return now to the that decomposition process leads to the solvable by redesign. A tentative sub-problems a decomposition node is shown in architecture of Figure 3. A problem specification is received from above in the hierarchy. If the problem can be solved by redesign, this is done, and the results returned upward. If not, an intial decomposition is made. Using this decomposition, initial specifications are assigned to the sub-problems created. (This assignment of specifications is a key step; the sub-problem specs are, in fact design variables at this stage.) These problems are then passed to the modules below, which are similar in structure to the one being described. The results returned from the various sub-problems are then integrated and analyzed as a complete system. If the complete system result is acceptable, it is passed upward. If not, new sub-problem specifications are assigned, and the process repeated. If the respecifier must admit defeat, then a new decomposition may be tried. If the re-decomposer must admit defeat, the system reports failure up the line, and asks for some change in the overall problem assignment. Others are also working decompose problems for advanced is the excellent and Chandrasakeran, (Brown GEOMETRY How shall we represent design geometries? Answering this question is key to our future to construct knowledge-based systems that ability can serve to integrate CAD, CAM, and engineering analysis (CAE). Since so much of our knowledge of manufacturing is currently expressed in terms of geometric features, the answer appears to be "In terms of features." But, what, exactly is a "feature", and how shall the desired features representations of designs be obtained? It is to be noted that this model does not allow for passing of information or "constraints" between sub-problems in the hierarchy. Only chaos can result from such communication. Our model is autocratic; information is passed only up or down. What we do include, however, is a mechanism for each sub-problem to pass up a re-specification request. This information is then a part of the information used to determine We have been addressing these questions about features in our research. So far, the definition of a feature is simply "any geometric form or entity whose presence or dimensions in a domain are germain to manufacturing evaluation or planning, or to automation of functional analyses". We have experimented with several different types of features in several domains COMMUNICATIONS I INITIAL , DECOMP II c INITIAL *SPECS fails INTERFACE 1 I I c l e OK I I REDECOMP on design models that solution. The most and useful work by 1986). < ACCEPT? fails I OK I COMMUNICATIONS TO SIMILAR Figure 3. INTERFACE SUBPROBLEM I MODULES A Typical Decomposition Module. APPLICATIONS / 875 molding, and casting), and (extrusion, injection working prototype research have constructed "design-with-features" that provide a programs (Dixon, 1985; Vaghul, environment for designers Libardi, 1986; Luby, 1986). These programs 1985; of the design, create a features representations this representation to draw the part in and use extrusion In the solid form. wire frame or the features representation is then used program, automatic finite element analysis to develop an deflections in a loaded and for stresses injection molding and casting the extrusion. In of the geometry is representation the programs, for on-line evaluation of the used as a basis the in-progress designs. manufacturability of like a bit function a These programs looking and who is expert manufactururing the designer the shoulder of over commenting REFERENCES while he or she designs. The programs are simple; it remains to be seen whether adding complexity will create insurmountable difficulties. and Breau, R, Brown, D. C. (19861, "Types of Design Problem-Solving", in Routine Constraints International Conference on Application of First AI to Engineering Problems, Southampton, England, April. also been working on the Others have been ways (Pratt, 1984; concept in various features 1984). Also, rather Popplestone, 1978; Latombe, than designing with features as we are doing, attempting to extract or infer features are some surfaces and lines, the points, from created by existing CAD systems representation (Henderson, 1984). SUMMARY AND CONCLUSION (1) that this paper are: themes of The from different very design is mechanical also be studied in must and electrical design, general models of the design order to obtain the differences involve the (2) that process; two domains with involvement in the degree of processes, and selection, manufacturing materials (3) that good analysis is especially geometry; for good design, but design is not and important cannot be done by analysis alone; (4) that the design process is iterative; (5) that design can inside redesign iterative modelled as be iterative inside respecification iterative and respecification but redesign decomposition, the most important; (6) that learning how to are iterative processes is intellectually guide methods can be strong/weak (7) that important; developed for design problem solving; (8) that learning how to represent the geometry of designs issue for applying AI to mechanical is a key and (9) that most likely the way to design; in terms of geometry is design represent features. If this paper has expanded the reader's model of design by providing useful insight into the the mechanical design process, it has nature of served its purpose well. ACKNOWLEDGEMENT the University of Massachusetts Research at of AI to mechanical design application into the in part by grants from sponsored automation is General Electric. 876 / ENGINEERING Bennett, J. S. and Englemore, , R. S. (1979). for Consultant "SACON: A knowledge-Based Sixth IJCAI, Palo Alto. Structural Analysis", "Capturing Mechanical (1985a) Brown, D. C., Proceedings ASME Computers in Design Knowledge", Engineering Conference, Boston, MA., August. "Failure (1985b) Brown, D. C. Design Expert System" CAD Journal, Design, Vol 17, No 9, November. D. C. (1985c) "Capturing Brown, Computers Knowledge", ASME Conference, Boston, MA. August. Handling Computer in a Aided Mechanical Design Engineering in and Chandrasakeran, Brown, D. C. Expert Systems Approach to Design", Proceedings IEEE Applications, Gaithersburg, MD. B., (1983) "An for Mechanical Trends and (1984) Chandrasakeran, B. Brown, D. C. and "Expert Systems for a Class Of Mechanical Design Working WG5.2 IFIP Activity, Proceedings on Knowledge Engineering in Computer Conference Aided Design, Budapest, Hungary, September. C. and Chandrasakeran, Brown, D. Design Problem Selection in Warwick, England, April. B. (1985) "Plan AIS85, Solving" Chandrasakeran, B. and Brown, D. C. for a Mechanical Control "Knowledge and Expert System", IEEE Computer, July. (1986) Design (1983) and Foyster, G. Tong, C., Brown, H., An Exploratory Environment for Circuit "Palladia: Design", Computer, Vol 16, No 12, December. Stall, H. W., Parunak, V. (1985) Cole, J. H., Report No. in Machine, Intellipence Machine 85-20, Industrial Technology Inst., Ann Arbor, MI "Expert and Simmons, M. K. (1984) Dixon, J. R. Systems for Design: Standard V-Belt Drive Design Design-Evaluate-Redesign an Example of the as ASME Computers in Proceedings Architecture", Engineering Conference, Boston, MA. August. and Simmons, M. K. (1985) "Expert Dixon, J. R. for Mechanical Design: A Program of Systems 85-DET-78, Design Research", ASME Paper No. Ohio, Cincinnati, Engineering Conference, September. Simmons, M. K., and Cohen, P. R. Dixon, J. R., "An Architecture for Applying Artificial (1984) Intelligence to Design", Proceedings IEEE Design Automation Conference, Albuquerque, NM, June. Dixon, J. R., Libardi, E. C., Luby, S. C., Vaghul, M. V., and Simmons, M. K. (1985) "Expert Systems for Mechanical Examples of Design: Symbolic Representations of Design Geometries", in Applications of Knowledge-Based Systems to Engineering Analysis and Design, ASME Publication No. AD-lo, New York. de Kleer, "Propagation Synthesis", 485. J. of MIT and Sussman, G. J. (1978) Constraints Applied to Circuit Artificial Intelligence Memeo Dym, C. L. (1985) Applications of Knowledge-Based Systems to Ennineerinp and Design, Publication No. AD-lo, American Society of Mechanical Engineers, New York. Henderson, M. R. "Extraction of Feature (1984) Information From Three Dimensional CAD Data", Ph. D. Thesis, Purdue University, Lafayette, Indiana. Howe, A., Simmons, M Independent Design", Conference Intelligence Southampton, Dixon, J. R., Cohen, P. R., and K. A Domain (1986) "Dominic: Program for Mechanical Engineering Proceedings International First Application of Artificial on to Engineering Problems, England, April. Kulkarni, V. M., Dixon, J. R., Simmons, M. K., and Sunderland, J. E. (1985) "Expert Systems for Design: The Design of Heat Fins as an Example of Conflicting Sub-goals and the Use of Dependencies", Proceedings ASME Computers in Engineering Conference, Boston, MA., August. (1976) Intelligence in Latombe, J. "Artificial Computer-Aided Design: The TROPIC System" TR 125, Stanford Research Institute, February. Libardi, E. C., Dixon, J. R., and Simmons, M. K. "Designing With Features: Extrusions As AN (1986) Design Paper No. 86-DE-4 ASME Example" Engineering Conference, Chicago, March. Buchanan, B. G., Fiegenbaum, E. A., Lindsay, R., (1980) DENDRAL, McGraw-Hill, New Lederberg, J. York. Dixon, J. R., and Simmons, M. K. Luby, S. L., Creating and "Designing With Features: (1986) Using a Features Data Base for Evaluation of Proceedings ASME Castings", Manufacturing of Chicago, in Engineering Conference, Computers July. (1978) "Circuit Design as Problem McDermott, D. Artificial Intelligence and Pattern Solving" Recognition in Computer Aided Design, J. Latombe Publishing Company, North-Holland (Ed), Amsterdam. Knowledge and the "Domain McDermott, J. (1981) Process", Eighteenth Design Automation Design ACM/IEEE, Nashville, Tennessee, July. Conference, McDermott, J. (1982) of Computer Systems, 19, No 1, September. "Rl: A Rule-Based Configurer Artificial Intellipence, Vol Mittal, S, Morjaria, M., and Dym, C. L. (1985) An Expert System for the Design of Paper "PRIDE: in Applications of Knowledge Handling Systems", Based Systems to Engineering Analvsis and Design, Publication no. AD-lo, American Society of Mechanical Engineers, New York. L., Kedar-Cabelli, Mitchell, T. M., Steinberg, Shulman, J., and Weinrich, T., S Kelly, V., "An Intelligent Aid for Circuit Redesign" (1983) Proceedings Third NCAI, Washington, D. C. Mostow, J. (1985)s "Towards Better Models Design Process" AI Magazine, Vol 6, No 1. of the "The Application of (1984) Popplestone, R. J. Design Systems", Intelligence to Artificial International Symposium on Proceedings First Design and Synthesis (ISDS, Tokyo, Japan. ffSolid Modelling and the Pratt, M. J. (1984) Design and Manufacture", IEEE, Interface I Between CG and A. "Function, Form, J. R. (1986) Rinderle, Decomposistion Fabrication Relations and Strategies in Design" Proceedings ASME Computers in Engineering Conference, Chicago, July. Simmons, M. K. and Dixon, J. R. (1986) "Reasoning About Quantitative Methods in Engineering Design", in Coupling Svmbolic and Numerical Computing Svsdtems in Expert Systems, J. S. Kowalik (Ed), North-Holland, Amsterdam. Simon, H. A. The Sciences Cambridge, MA. (1969) "The Science of Design" in of the Artificial, MIT Press, Stefik, M., Bobrow, D., Brown, H., Conway, L. and Tong, C., (1982) "The Partitionong of Concerns in Digital System Design", Proceedings of the Conference on Advanced Research in VLSI, MIT, Cambridge, MA. Steinberg, L. I. and Mitchell, T. M. (1985). "Redesign System: A Knowledge-Based Approach to VLSI CAD", IEEE Design and Test of Computers, Vol 2, Number 1, February. Sussman, G. (1977), "Electrical for Artificial Intelligence IJCAI. Design: A Problem Research", Fifth Sussman, G. J., (1978) "Slices: At the Boundary Between Analysis and Synthesis" Artificial Intelligence and Pattern Recognition in Computer Aided Design, J. Latombe (Ed), North-Holland, Amsterdam. Vaghul, M. V., Dixon, J. R., and Simmons, M. K. (1985) "Expert Systems in a CAD Environment: Injection Molding as an Example", Proceedings ASME Computers in Engineering Conference, Boston, APPLICATIONS / 877