Sebastian Brandstäter (Computer-Based Simulation, UniBwM)
This project will enable variance-based global sensitivity analysis based on Sobol’s method for large-scale finite-element models within the research software QUEENS.
QUEENS is a Python framework for multi-query analyses like uncertainty quantification and Bayesian inverse problems of large-scale simulation models. Its applications range from automating routine tasks in numerical method development, such as convergence studies, to enabling global sensitivity analysis of multi-physics systems like the cardiovascular system.
The project is motivated by our research in biomedical engineering, where we develop and analyse complex multi-physics simulation models of the human body that should aid medical decision-making. Examples include predicting disease progression, such as the evolution of an aneurysm over the next fifteen years, and planning optimal treatment by simulating different options, such as comparing gastric bypass, sleeve gastrectomy, and gastric band as bariatric surgery options. The high stakes involved in medical decision-making imply the necessity to consider any uncertainties involved in the model predictions upon which the decisions should be based.
We employ variance-based global sensitivity analysis to address the predictive uncertainty caused by uncertain model parameters. A global sensitivity analysis requires evaluating finite-element models thousands of times with varied parameters. Global sensitivity analysis of large-scale FEM models benefits from using high-performance computing (HPC) in two ways: Firstly, by exploiting embarrassingly parallel algorithms maximising simultaneous execution of numerous (>>100) simulations. Secondly, it allows individual simulations to use finely resolved meshes and produce more accurate solutions by spanning multiple compute nodes on an HPC cluster.
In this project, we enhance QUEENS’ performance by leveraging the aforementioned HPC benefits for global sensitivity analysis. Namely, we will increase the maximum number of simultaneous simulations on an HPC cluster and enable the use of multiple compute nodes for individual simulations. These performance engineering measures will significantly reduce the runtime of global sensitivity analyses of large-scale finite-element models in QUEENS while at the same time enabling the analysis of ever more sophisticated models. In this way, this project will ultimately bring advanced sensitivity analysis methods closer to clinical practice.