Massively parallel adaptive mesh refinement and coarsening for dynamic fracture simulations. (July 2016)

• Record Type: Journal Article
• Authors: Alhadeff, Andrei; Leon, Sofie; Celes, Waldemar; Paulino, Glaucio
• Is Part Of: Engineering with computers. Volume 32:Number 3(2016)
• Page Start: 533
• Page End: 552
• Publication Date: 2016-07
• Subjects: GPU computing -- Finite elements -- Adaptive mesh refinement -- Adaptive mesh coarsening -- Cohesive zone model -- Microbranching; Engineering design -- Data processing -- Periodicals; Computer-aided design -- Periodicals; Conception technique -- Informatique -- Périodiques; Conception assistée par ordinateur -- Périodiques; Electronic journals; 620.00285
• DOI: 10.1007/s00366-015-0431-0
• Languages: English
• ISSNs: 0177-0667

Abstract

Abstract We use the graphical processing unit (GPU) to perform dynamic fracture simulation using adaptively refined and coarsened finite elements and the inter-element cohesive zone model. Due to the limited memory available on the GPU, we created a specialized data structure for efficient representation of the evolving mesh given. To achieve maximum efficiency, we perform finite element calculation on a nodal basis (i.e., by launching one thread per node and collecting contributions from neighboring elements) rather than by launching threads per element, which requires expensive graph coloring schemes to avoid concurrency issues. These developments made possible the parallel adaptive mesh refinement and coarsening schemes to systematically change the topology of the mesh. We investigate aspects of the parallel implementation through microbranching examples, which has been explored experimentally and numerically in the literature. First, we use a reduced-scale version of the experimental specimen to demonstrate the impact of variation in floating point operations on the final fracture pattern. Interestingly, the parallel approach adds some randomness into the finite element simulation on the structured mesh in a similar way as would be expected from a random mesh. Next, we take advantage of the speedup of the implementation over a similar serial implementation to simulate a specimen whose size matches that of the actual experiment. At this scale, we are able to make more direct comparisons to the original experiment and find excellent agreement with those results.