These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

High Performance Computing (HPC) has been recently applied to different scientific and engineering fields and even financial sector. With increasing available and demanding computational resources, (e.g., CPUs, GPUs, memory, storage space, network), more HPC calculations for faster turn-around time are expected in supporting large Finite Element (FE) analyses, design optimization, design space exploration, etc. Simply porting FE solvers from desktops to HPCs and brute forcing executing them at a HPC environment will not fully utilize the HPC'™s capability. Software and applications would need to be adapted or refactored if needed to effectively use the available hundreds, thousands, and more processors in a HPC system. This paper presents two advanced numerical techniques combined with HPC to solve structural acoustic analyses accurately and efficiently, especially for large FE models. The first numerical technique is 'œFinite Element Tearing and Interconnecting (FETI)' which is an iterative solver and is a massively parallel code developed by researchers from Stanford university. It can effectively scale up across hundreds and thousands computing nodes. The other technique is 'œAdaptive Krylov subspace and Galerkin Projection (AKGP)' which applies user-defined tolerance to speed-up frequency-sweep of a FE model. Instead of calculating frequency response functions (FRF) at every frequency point, the AKGP only needs to calculate a small subset of the original frequency range and use AKGP with the defined tolerance to calculate FRF for the rest of frequency points. FRF is commonly performed in structural acoustics analyses to predict structural responses and identify dynamic characteristics of the underlining structures. Several examples will be presented to demonstrate the benefits of combining HPC with the two advanced techniques in conducting structural acoustic predictions. The accuracy of the advanced techniques is firstly validated by comparing with Abaqus prediction. The efficiency of combining the advanced techniques with HPC will be demonstrated by a large FE model with more than 60M Degree of Freedom (DOF). The model is exercised at the DoD HPC system and utilizes more than 100 nodes to speed up the overall turn around time.