Posters

This poster provides a brief introduction to triangular transport. I introduce the context in which the method is relevant (statistical inference), briefly illustrate the deficiencies of state-of-the-art ensemble-based inference methods like the EnKF and the particle filter. I then briefly explain how triangular maps address this challenge: first, we learn to express an unknown joint distribution by transforming a well-defined standard Gaussian reference; then, we use this map to realize a nonlinear ensemble-based inference operation. Finally, I highlight two of the features that make triangular transport interesting for the geosciences: the ability to exploit conditional independence (which ensures scalability) and the ability to adapt the map to the system's demands (which ensures parsimony).

This poster presents my work on map adaptation for nonlinear triangular ensemble transport maps. In brief, we have developed an efficient map adaptation strategy which automatically identifies parsimonious triangular maps for ensemble transport filtering. This adaptation strategy can also detect conditional independence properties, which permits an efficient form of localization. In our experiment, we observe a number of promising features: for small ensemble sizes, the adaptation algorithm favours simpler linear maps. As ensemble size grows, greater degrees of nonlinearity become viable. This poster includes augmented reality elements which animate some of the static figures. If the videos display as black rectangles, try using a different browser.

This poster complements my talk at the EnKF Workshop 2023 of the same name, and focusses on topics that did not fit into the talk itself. In brief, we have developed an efficient map adaptation strategy which automatically identifies parsimonious triangular maps for ensemble transport filtering. This adaptation strategy can also detect conditional independence properties, which permits an efficient form of localization. However, the degree to which we can exploit these properties depends on variable ordering. We propose a modified version of the reverse Cuthill-McKee algorithm, which detects efficient orderings while respecting variable block structure.

This is online interactive poster was presented at the AGU session "Advances in Data Assimilation, Predictability, and Uncertainty Quantification". It summarizes our research on ensemble transport smoothers (Part 1 and Part 2). In case the interactive poster is removed from the AGU website, I have also provided a recording of the poster's contents here.

This hybrid poster presents intermediate progress of my work on nonlinear smoothing with transport methods. We discuss two different smoothing strategies (joint-analysis and backward smoothers), often applied as Kalman-type algorithms. We then introduce transport methods as a pathway for nonlinear generalization of these smoothers. We demonstrate the efficacy of the resulting algorithm with the Lorenz-63 test case. The voiceover can be found here.

This presentation showed a later stage of my work on the second chapter of my dissertation, in collaboration with Prof. Matteo Camporese from the University of Padova and Prof. Philippe Renard from the University of Neuchâtel. As part of the EU Horizon2020 INSPIRATION ITN, I also presented at this conference. At this stage, our work on the second chapter of my dissertation has progressed sufficient that we could already show prediction results for the optimized parameter fields.

This presentation showed an early stage of my work on the second chapter of my dissertation, in collaboration with Prof. Matteo Camporese from the University of Padova and Prof. Philippe Renard from the University of Neuchâtel. The goal of this research was to come up with a sequential optimization scheme for groundwater models with uncertain subsurface structures, as - for example - generated by multiple-point statistics.

This poster explored early experiments to use particle filters for optimization of groundwater model parameters. However, instead of optimizing the parameters directly, we optimize a set of hyper-parameters used to generate the parameter fields used by the model. We explored what happens when geological features are accidentally misspecified. This was also the first poster for which I experimented with augmented reality elements. This poster won one of the 2018 EGU OSPP awards. 

My first presentation at a major scientific conference, the 2017 AGU Fall Meeting in New Orleans, early into my PhD. This poster shows early experiments with particle filters for parameter inference in hydrogeological models.