PathTrAce-GwT: Coupling Pathline Tracers to Multi-Step Neural Networks for the Fast Assessment of Parametric Uncertainties in Groundwater Transport Models

• Applicant: Dr.-Ing. Martin Ortwin Binder
• Subject Area: Hydrogeology, Hydrology, Limnology, Urban Water Management, Water Chemistry, Integrated Water Resources Management
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 562568682

The project PathTrAce-GwT investigates the link between analytical errors of observation data and the associated parametric uncertainties in flow and transport models. Revealing the quantitative characteristics of aquifer systems is a central task in hydrogeological projects. Observations obtained at monitoring wells are valuable sources of information. Sophisticated computer-aided methods exist which help to inversely extract parameters from it. Unfortunately, uncertain data can also introduce gaps into the knowledge base. Hence, it is critical to also extract parametric uncertainties. Since existing methods of post-calibration uncertainty analysis are either computationally costly or not robust enough for a general use, more efficient and intelligent methods are needed. PathTrAce-GwT aims at closing this methodological gap by developing and testing a novel method for efficiently performing the parametric uncertainty analysis. It is based on the simultaneous, coordinated usage of two selected surrogate model approaches. On the one hand, complex 2-D / 3-D model data are intentionally reduced to pathline surrogates in form of, among others, breakthrough curves (BTC). On the other hand, the typically process-based, time-consuming procedure of parametric information extraction is replaced by neural networks with multiple steps (component 1: (Inverse) Physics-Induced Neural Networks (PINN), 2: Bayesian Neural Networks (BNN), and eventually, 3: Convolutional Neural Networks (CNN)). The model training is based on result data of previous runs during the initial model calibration. Coupled together, an efficient and intelligent novel approach will be formed which can learn the model characteristics in order include them into the analysis. The work program comprises of six interacting work packages (WP). In WP A, the pathline extraction algorithm for obtaining arrays of 1-D BTCs and CPs will be developed. In the double WP B, neural networks will be tested regarding their suitability for (WP B1) reconstructing transport parameters based on a feature and shape analysis (via PINN) as well as (WP B2) extracting their respective uncertainties (via BNN). The WP C focuses on providing benchmark datasets. The coupling of the surrogates will be done in the final double WP D. Here, WP D-1 is focusing on the development of a CNN structure for transforming the pathline surrogates into a data format usable for the PINN and BNN analysis steps. In the parallel WP D-2, a validation to experimental data will take place. Eventually, PathTrAce-GwT will not only develop a workflow, but will also assess the robustness of surrogate modelling, and specifically neural network-based modelling, by a comparison to classical process models.

• DFG Programme: Research Grants
• International Connection: Australia, Switzerland
• Co-Investigator: Dr. Catalin Stefan
• Cooperation Partners: Professor Henning Prommer, Ph.D.; Dr. Oliver Schilling; Professor Mario Schirmer, Ph.D.