Last week, the Research Group on Statistical Model–Data Integration celebrated a double success: Philipp Luca Reiser and Manuel Álvarez Chaves both defended their doctoral theses at the University of Stuttgart within the SimTech environment. Their supervisor, Dr. Anneli Guthke, emphasized how much both projects contributed to the group’s work at the interface of simulation, statistics, and data: “Philipp and Manuel pushed forward methods that make model–data integration more reliable and more transparent. I’m thrilled to see them complete this important milestone and wish them all the best for the next phase of their careers.”
In his dissertation, Bayesian Surrogate Modeling for Uncertainty-Aware Inference and Data Integration, Philipp tackled a key bottleneck in simulation-based research: high-fidelity simulators are often too expensive to run at scale. His work advances Bayesian surrogate modeling in a way that does not merely replace the simulator with a fast approximation, but also quantifies and propagates surrogate-induced uncertainty so that downstream Bayesian inference remains well-calibrated. A central outcome is his framework UA-SABI (Uncertainty-Aware Surrogate-based Amortized Bayesian Inference), which links surrogate modeling with amortized Bayesian inference while explicitly tracking uncertainty throughout the workflow. He demonstrated the approach in multiple application settings, including epidemiology, subsurface CO₂ storage, and biogeochemical processes.
Manuel’s dissertation, Evaluation and Enhancement of Data-Driven Hydrological Models Using Information Theory, focused on how modern data-driven and hybrid hydrological models can be assessed and improved beyond standard error metrics. By bringing information-theoretic tools into hydrological model evaluation, he provided a sharper view of what models learn, where they fail, and how uncertainty should be represented. Among his contributions are systematic insights into the practical estimation of quantities such as entropy and mutual information, an entropy-based method for diagnosing hybrid models that combine conceptual hydrology with LSTM components, and a variational LSTM (vLSTM) approach for probabilistic streamflow prediction that learns predictive uncertainty directly from the
Both projects were carried out in close exchange with cooperation partners, and the group warmly acknowledges the valuable input and scientific discussions with PD Dr.-Ing. Uwe Ehret, Prof. Hoshin V. Gupta, Daniel Klotz (IT: U Linz), and Prof. Paul Bürkner (TU Dortmund). The themes addressed in these theses – uncertainty-aware inference, rigorous model–data integration, and a deeper understanding of hybrid modeling – will continue to shape the research agenda in Dr. Guthke’s group.