An agent-based approach to building multidisciplinary problem-solving environments

Tzvetan Todorov Drashansky, Purdue University

Abstract

Electronic prototyping is becoming a part of every scientific inquiry and product design. The rapid prototyping of manufactured artifacts and the rapid solution of problems with numerous interrelated elements requires the fast, accurate simulation of physical processes using knowledge and computational models from multiple disciplines in science and engineering. We argue the necessity of a generally applicable and flexible programming environment for simulating (numerically) multiple-domain, heterogeneous, and possibly very complex multiphysics phenomena that is accessible to users who are not experts in scientific computing. We argue that one can be built with the extensive use of available legacy simulation software. We present a paradigm for simulating such complex models that may involve multiple physical phenomena and complicated geometry. The computational structure is of cooperating agents, and much of the proposed methodology is widely applicable; but the focus in this work is on phenomena modeled by partial differential equations (PDEs). The computational process is to subdivide the physical object into components of simple geometric shapes each modeled by a single problem solving environment (PSE). PSEs are viewed as agents that solve PDEs on each component independently. The interfaces between the components must have physical interface conditions satisfied--mediator agents use relaxation techniques for this. An agent-based architecture of an environment for building systems to implement this paradigm, along with the necessary mathematical and software tools for building it, are described, using PSEs that are encapsulated and transformed into solver agents. This approach is naturally parallel and highly scalable; it is suitable for a wide variety of parallel and distributed high performance computing (HPC) architectures; it is easily applicable in the context of networked and network-centric computing; it allows for the reuse and evolution of existing HPC software and for a convenient abstraction of the solution process and the low cost building of the simulation software for non-expert users. An implementation of the architecture, named SciAgents, is presented which illustrates this multidisciplinary problem solving environments (MPSE) framework. The detailed software architecture and the prototype implementation of SciAgents are discussed in the context of a set of examples.

Degree

Ph.D.

Advisors

Rice, Purdue University.

Subject Area

Computer science|Engineering

COinS