An intelligent user interface for computational fluid dynamics software

The development of advanced and successful numerical modelling packages for Computational Fluid Dynamics ( CFD ) within recent years has led to their widespread use in engineering and industrial environments. Increasingly this has led to a situation where package users have little or no specialist knowledge of the underlying physical principles upon which CFD packages are based. Generally users also have limited knowledge of the numerical software itself, or how to obtain accurate results efficiently. However this knowledge can often be vital to the correct usage of the software for producing reliable simulation data. This paper describes an on-going research project to incorporate such expert knowledge into CFD software. The special problems encountered are those of interfacing knowledge based components with numerical routines via a blackboard^ architecture. The design of appropriate user interfaces and the provision of dynamic run-time graphical displays and solution monitoring are also problematic areas. The approach adopted here allows the interaction of knowledge sources, numerical solution routines, display tools, and pattern recognition tools. The architecture of the system is described, together with experience of its capabilities and benefits to CFD simulation. Transactions on Information and Communications Technologies vol 1, © 1993 WIT Press, www.witpress.com, ISSN 1743-3517 78 Artificial Intelligence in Engineering INTRODUCTION TO CURRENT RESEARCH The widespread use of Computational Fluid Dynamics by novice users and the extensive demands placed on CFD techniques has led to a situation where traditional batch mode packages are no longer satisfactory. This paper outlines the merits of using a Knowledge Based System ( KBS ) approach to CFD and proposes an alternative architecture for interactive simulations. The design and development of a prototype system is considered with emphasis placed on reliability and overall efficiency rather than speed performance. Prior research into intelligent front ends to numerical simulation packages has indicated the feasibility of KBS techniques for CFD. Much of the existing research treats the numerical component as a batch process. The system proposed by this paper provides interactive control of the CFD numerical process. The current research is based on FLOWES^, a prototype, inference controlled CFD code. The continued development of the FLOWES architecture has indicated problems associated with existing approaches for developing and using CFD codes. This paper details solutions and proposes a new approach to CFD simulations. BACKGROUND TO RESEARCH A KBS is a system that uses human expertise and knowledge, usually in the form of heuristic rules, to reason about specific application areas. The application areas have to be limited because computational demands for inferencing are far higher than those for numerical calculation. Knowledge is elicited from human Experts and then encapsulated in some suitable form within an application specific database, often called a Knowledge Source ( KS ). In the work described here the application area is that of fluid dynamics simulation. The encapsulated knowledge is used to make decisions about the solution strategy adopted for a particular simulation. Much of the previous research has been devoted to KBS support for the specification of a simulation. The techniques utilised have the advantage of ensuring that the simulation is specified correctly, completely and consistently. Such support is limited because there is no dynamic control of the numerical simulation component. FLOWES makes use of a blackboard architecture which provides a flexible and extensible framework within which numerical and knowledge based components can interact cooperatively. There has been limited research into KBS control of a numerical code^l This was restricted to the adjustment of a limited Transactions on Information and Communications Technologies vol 1, © 1993 WIT Press, www.witpress.com, ISSN 1743-3517 Artificial Intelligence in Engineering 79 number of parameters by hard-coded conditional rules. This integration of rules into the source code of an application suffers from a lack of flexibility and has only very limited inference capability. There is little possibility of conflict resolution and the monitoring capabilities are limited. The approach adopted did not provide any interaction for the system user and the numerical component was still, essentially, a batch mode process. PREVIOUS RESEARCH FLOWES^ is a two dimensional heat transfer code that provides KBS support for problem set-up, mesh generation and solver selection. The system also provides dynamic solution control by the integration of a KBS component^ into the numerical solver. There have been a number of problems during the development of FLOWES due to the nature of the control structure required. A blackboard architecture^ was chosen for truly interactive solution control. The severe restrictions of existing CFD codes to external interactive control have necessitated the re-implementation of the numerical component to provide KBS control capabilities. The FLOWES system is currently under test for reliability of the inference techniques and for suitability of the design. The results, to date, are promising and suggest that the integrated blackboard approach is suitable for CFD. FLOWES currently has a number of limitations, namely:• Support for two-dimensional un-structured meshes only. • No solution for fluid flow properties. • No user interface for operations controlled by the KBS component. These limitations have restricted the current research because the prototype system needs flow simulation capability and threedimensional meshes. Other researchers are investigating KBS techniques for CFD. Jambunathan, Lai, Hartle and Button are developing an Intelligent Front End ( IFE ) for PHOENICŜ , a well known commercially available CFD code. This IFE is designed to support problem set-up and specification during an interactive question and answer session at a computer terminal. Expertise is used to support the set-up and to pre-set the parameters and switches used by the PHOENICS system̂ . The output from the IFE is a simulation specification file that can then be used by the CFD code to run that problem. The CFD code is still treated as a numerical "black-box" since the user has little control of the computations once the processing has started. Transactions on Information and Communications Technologies vol 1, © 1993 WIT Press, www.witpress.com, ISSN 1743-3517 80 Artificial Intelligence in Engineering Finn, Hurley and Sagawa are developing AI-DEQSOL̂ . This is an Artificial Intelligence ( AI) system for the numerical simulation of engineering problems that are defined by partial differential equations. The system comprehensively supports the specification of the mathematical model and the boundary conditions using an IFE. However the numerical solver is again used as a batch process. The output from the IFE is generated simulation code which is then processed by the DEQSOL simulation system. AIMS OF CURRENT RESEARCH The previous research has demonstrated the potential benefits of KBS techniques but little research has been conducted into comparison between KBS support for CFD and fully interactive CFD. There are currently no CFD packages that support dynamic solution control by either a KBS or the application user. The current project will have an intuitive Graphical User Interface ( GUI ). The KBS can then be switched off so that meaningful investigations into the relative benefits of a KBS, can be conducted. Knowledge will be acquired for two example CFD application areas, namely external vehicle aerodynamics and fire simulation modelling. The CFD numerical component will have to support threedimensional meshes, body fitted coordinates and solve turbulent and elliptic flowŝ . These requirements will affect the overall system design criteria. Much of the existing CFD research is restricted to use within specific codes. It is intended that an application framework for KBS techniques will be developed during the current research. The data and control architectures will be sufficiently flexible to allow for the integration of other research work. The limitations of many existing CFD codes is, in part, due to the techniques employed in their development. The current research is intended to highlight, to CFD developers, the benefits of using sound software engineering principleŝ . Future research and code maintenance would be greatly facilitated by the use of a software design methodology. CFD codes would be easier to re-use and modify if supported by comprehensive and accurate documentation. These techniques are rarely used by most CFD code developers. THE PROBLEMS OF CFD RESEARCH There are problems associated with the usage of many of the CFD Transactions on Information and Communications Technologies vol 1, © 1993 WIT Press, www.witpress.com, ISSN 1743-3517 Artificial Intelligence in Engineering 81 codes currently available. Many of these difficulties stem from the complexity of the underlying code'"""' and the vast scope of fluid flow type simulations. The numerical approximation techniques used for CFD are vital to obtaining reliable solutions and users must currently be aware of the limitations and restrictions of the various techniques available. CFD techniques have been used successfully for over 20 years but still lack the reliability and robustness of some other disciplines ( e.g. CAD and structural Finite Element Analysis applications ). The use of CFD codes is still only viable for CFD experts. These experts need to be well versed in both the physical principles upon which the codes are based and the quirks and implementation details of specific codes. These joint requirements mean that there are few users who can specify and run a simulation successfully in one attempt. The complexity of the relationships between physical phenomena and the numerical modelŝ "' mean that there are many parameters to set for even relatively simple simulations. CFD simulations also have the problem that the configuration required for an accurate solution depends on the results of the simulation but these are obviously unknown at the start of the simulation. To overcome this problem CFD users frequently run quick coarse-grid simulations to get a "look and feel" of the solution. A new simulation can then be run with a more appropriate configuration. Unfortunately the sensitive nature of the CFD discretization techniques can mean that important features are missed in coarse-grid runs and hence the configuration could still be invalid. The numerical processing of CFD is generally a batch mode process which follows the traditional cycle of specifying the simulation, running a batch mode solver and finally interpreting the results. The extended duration of many simulations ( e.g. 64 hours for a typical three dimensional fire simulation on a workstation ) is such that this batch mode of processing can be extremely inefficient because:• The results can be found to have diverged early in the simulation. # There can be localised or wholly erroneous results. * Important features can go undetected ( e.g. Plumes above fires ). # The results can be unrealistic or physically impossible because of instabilities for a given configuration. Many batch mode CFD codes provide only minimal information about the intermediate state of the solution and even where potential Transactions on Information and Communications Technologies vol 1, © 1993 WIT Press, www.witpress.com, ISSN 1743-3517 82 Artificial Intelligence in Engineering problems are detected there is generally no way to correct the situation without restarting the simulation with a new configuration. The poor monitoring of the intermediate solution status and the lack of run-time interaction means that problems can only be detected after the simulation. This is clearly unsatisfactory for novices who need help and guidance to get realistic results. CFD is also used widely in design and safety applications where solution reliability and the accuracy of results are of paramount importance. The fact that CFD techniques are now widely available and are used more extensively can only exacerbate this problem. STRUCTURE OF A KBS FOR CFD The proposed structure for a CFD application that supports dynamic interaction and knowledge based reasoning is simplified to the following five components:Knowledge about CFD. (Heuristics) Knowledge Based Component 1 Numerical Component