Deep Learning for Seismic Monitoring of Porous Water Reservoirs

Cotutelle doctoral thesis completed under a joint agreement between the University of Eastern Finland (Department of Technical Physics, Kuopio) and KTH Royal Institute of Technology (Department of Engineering Mechanics, Stockholm). Reviewers (pre-examiners): Professor Inga Berre (University of Bergen, Department of Mathematics) and Associate Professor Andreas Hauptmann (University of Oulu, Department of Mathematical Sciences). Study supported by the Research Council of Finland, including the Finnish Centre of Excellence of Inverse Modeling and Imaging, and the Flagship of Advanced Mathematics for Sensing, Imaging and Modelling.

Field mentioning the UEF-side ISBN/series (978-952-61-6079-5 print, 978-952-61-6080-1 pdf, UEF Dissertations in Science, Forestry and Technology N:o 153) so the cross-reference between the two institutional editions is documented in the record

Research funders: Flagship of Advanced Mathematics for Sensing Imaging and Modelling (grant 358944); Research Council of Finland (project 321761)

QC 260724

Abstract

Groundwater is a vital resource that supports ecosystems, agriculture, and human societies. Effective aquifer characterization and monitoring is crucial for sustainable groundwater management, particularly in the context of growing pressure from population growth, climate change, and increasing water demand. This thesis explores the integration of physics-based modeling and data-driven techniques to enhance the monitoring of shallow aquifers using seismic data. Building on Biot's poroviscoelastic theory, this work simulates wave propagation in porous media and develops a neural network-based inversion framework to determine aquifer parameters, including the water table level, porosity, and water volume. The forward model employs discontinuous Galerkin method to solve a coupled poroviscoelastic–viscoelastic model with complex three-dimensional geometries, thus enabling the generation of high-fidelity synthetic seismic datasets. The characterization process begins with a data preparation step, and a deconvolution-based model is used to normalize source effects. This model is introduced in order to remove source-wavelet effects and, as a result, ensure that the neural network analyzes only subsurface responses. Fully connected real-valued and complex-valued neural networks are employed to identify the mappings between seismic waveforms and aquifer parameters. The performance of the model is tested against both synthetic and real seismic data. The model is validated with real data through controlled field experiments at the Laukaa test site in Finland. Transfer learning techniques are implemented to bridge the synthetic-to-real domain gap, thus enhancing the generalizability of the model under limited labeled field data. Furthermore, SHapley Additive exPlanations (SHAP) are used to identify the most informative seismic receivers that have the potential to optimize sensor layouts. The results suggest that combining advanced numerical wave simulations with machine learning can enable efficient groundwater monitoring. The proposed framework also contributes to geophysical inversion tools that integrate physical modeling with deep learning for the assessment of sustainable groundwater resources.

Link to DiVA