Data-Efficient and Uncertainty-Aware Inverse Modelling for Biomechanical Simulation

Kurzfassung

Human body posture and muscle tension play a crucial role in determining injury severity during vehicle crashes, motivating the use of detailed biomechanical simulations. Building on a finite element model that simulates the motion of a human arm under constant muscle stimulations, this thesis focuses on the mathematical problem of inverse estimation: Determining the muscle stimulation required by the simulation to achieve a desired posture. This inverse problem is high-dimensional, strongly redundant, and nonlinear, making analytical solutions infeasible. To address this, different machine learning algorithms are applied and compared regarding their ability to approximate and invert the system’s behavior. A further key component is the development of an iterative data generation strategy that enhances model accuracy while reducing simulation calls. The final approach combines Gaussian process regression with a tailored optimization approach and a goal-directed, iterative data generation process that adaptively refines the training data. This enables a data-efficient and mathematically grounded estimation of muscle stimulations, where the inverse estimation achieves a mean absolute error of 0.58 when evaluated at 5◦ target increments.