GPGPU-based explicit finite element computations for applications in biomechanics: the performance of material models, element technologies, and hardware generations. Issue 16 (10th December 2017)

• Record Type: Journal Article
• Authors: Strbac, V.; Pierce, D. M.; Vander Sloten, J.; Famaey, N.
• Is Part Of: Computer methods in biomechanics and biomedical engineering. Volume 20:Issue 16(2017)
• Page Start: 1643
• Page End: 1657
• Publication Date: 2017-12-10
• Subjects: Finite element analysis -- general purpose graphics processing unit -- anisotropic constitutive model -- Gaussian integration; Biomechanics -- Data processing -- Periodicals; Biomedical engineering -- Periodicals; Biomechanics -- Periodicals; Biomedical Engineering -- methods -- Periodicals; Computing Methodologies -- Periodicals; 612.7
• DOI: 10.1080/10255842.2017.1404586
• Languages: English
• ISSNs: 1025-5842

Abstract

Abstract: Finite element (FE) simulations are increasingly valuable in assessing and improving the performance of biomedical devices and procedures. Due to high computational demands such simulations may become difficult or even infeasible, especially when considering nearly incompressible and anisotropic material models prevalent in analyses of soft tissues. Implementations of GPGPU-based explicit FEs predominantly cover isotropic materials, e.g. the neo-Hookean model. To elucidate the computational expense of anisotropic materials, we implement the Gasser–Ogden–Holzapfel dispersed, fiber-reinforced model and compare solution times against the neo-Hookean model. Implementations of GPGPU-based explicit FEs conventionally rely on single-point (under) integration. To elucidate the expense of full and selective-reduced integration (more reliable) we implement both and compare corresponding solution times against those generated using underintegration. To better understand the advancement of hardware, we compare results generated using representative Nvidia GPGPUs from three recent generations: Fermi (C2075), Kepler (K20c), and Maxwell (GTX980). We explore scaling by solving the same boundary value problem (an extension–inflation test on a segment of human aorta) with progressively larger FE meshes. Our results demonstrate substantial improvements in simulation speeds relative to two benchmark FE codes (up to 300 while maintaining accuracy), and thus open many avenues to novel applications in biomechanics and medicine.