OVERVIEW OF ADVENTURE PROJECT - Development of Computational Mechanics System for Large Scale Analysis and Design
Abstract
Most critical issues in the 21st century include understanding of global scale phenomena towards appropriate actions for global environmental problems, risk reduction caused by natural disasters as well as accidents of complex large systems, and the production of novel artifacts for welfare. In order for any of these issues, it is strongly required to predict behaviors of natural things or artifacts accurately, and design, control and operate them appropriately based on the prediction. Various general-purpose computational mehanics systems have been developed in the last three decades to quantitatively evaluate natural phenomena such as deformation of solids, heat transfer, fluid flow and electromagnetics. Nowadays such systems are regarded as infrastructural tools for the present industrialized society. The existing systems, however, can not be used with massively parallel processors with the order of 100-10,000 PEs, which are expected to dominate the high-performance computing market in the 21st century, as they were developed for single-processor computers, which took leadership an age ago. Neither the current systems can be used in heterogeneous parallel and distributed computing environments. Owing to the fact, they can deal with only medium scale problems with millions degrees of freedom (DOFs) at most. In the ADVENTURE project, we will develop an advanced general-purpose computational mechanics sytem which can analyze a model of arbitrary shape with 10 to 100 millions DOFs within 1 hour to 1 day using the world's fastest computer with 30-100 TelaFLOPS in 2002, and which additionally enables optimal and satisfactory design. To efficiently deal with such an ultra-large scale model, massively parallel algorithms are embedded in the pre-, main as well as post-processes. Neural networks, fuzzy set theory, genetic algorithms and virtual reality then enable ideal user-friendliness. In addition, an object-oriented design and programming method is employed such that new theories and algorithms can be easily implemented for a variety of heterogeneous parallel and distributed environments. This allows us to improve the prediction of behaviors of natural things and artifacts by two to three orders and to further design and operate them.