Running seismic simulations is an important application of computational seismology to better understand the dynamics and causes of earthquakes.
Our goal is to utilize observational seismological receiver data to find an approximation of possible source coordinates of an earthquake. Due to not knowing every single parameter and numerical errors of the simulation we apply Uncertainty Quantification (UQ) to infer possible locations of the source.
We use the open-source software SeisSol on a high-performance Linux cluster to simulate earthquakes as well as to determine whether the source fits its receiver data in the wrapper program UQ_SeisSol.

In the previous work the backward propagation applying Markov Chain Monte-Carlo (MCMC) methods was already implemented utilizing the parallel Generalized Metropolis-Hastings (GMH) algorithm combined with fused simulations of SeisSol. However, the implemented GMH algorithm rejected every sample proposed by the Metropolis adjusted Langevin algorithm (MALA) and infinite-dimensional MALA (Inf-MALA). For that reason, we conduct a hyperparameter study to find configurations of the MALA and InfMALA proposal to obtain an acceptance rate greater than 0. Furthermore, we research the number of fused simulations which yields the best performance of UQ_SeisSol.