Reservoir Characterization and Monitoring for Subsurface Storage using Deep Learning Tools

The PhD thesis discusses the integration of data-driven approaches such as Deep Learning (DL) with physics-driven methods including Full-Waveform Inversion (FWI) and rock-physics modelling to advance geophysical monitoring for Carbon Capture and Storage (CCS) and Underground Hydrogen Storage (UHS). Motivated by the urgent need to reduce atmospheric carbon-dioxide (CO2) and to enable safe subsurface storage technologies, this work emphasizes the advantages of DL in geophysics in improving interpretation accuracy, reducing subjectivity, and enabling cost-effective time-lapse surveys. We developed a data-driven workflow to map time-lapse seismic gathers directly to CO2 saturation. A U-Net architecture enriched with multi-scale features via the Continuous Wavelet Transform (CWT) improves sensitivity to subtle and spatially variable plume signatures, providing more accurate fluid saturation estimates. A cost-effective monitoring strategy is then introduced based on pre-trained convolutional encoders arranged in a Siamese architecture to extract baseline versus monitor features. UMAP projects the high dimensional embeddings to a lower dimensional space, enhancing time lapse evolution tracking and supporting rapid anomaly screening and detection. This framework highlights baseline-monitor differences and reduces the need for dense data acquisition and manual screening. This approach is at first demonstrated on synthetic datasets, and then validated on a time-lapse Distributed Acoustic Sensing (DAS) seismic dataset from a controlled CO2 injection at the Svelvik CO2 Field Lab, Norway. As part of the PhD program, I led the SBEM (Svelvik Borehole Electromagnetic Monitoring) project in collaboration with SINTEF, during which both electromagnetic (GPR and ERT) and seismic (cross-well and VSP) data were acquired using hydrophones and Distributed Acoustic Sensing (DAS). This campaign provided the experimental dataset used for validating the proposed monitoring framework. The results demonstrated robust injected fluid detection at lower acquisition density and provided a real-world testbed for evaluating plume evolution. Furthermore, the thesis introduces a rock-physics-parametrized FWI that performs inversion directly for a petrophysical parameter, such as saturation, thereby mitigating the cross-talk typical of multi-parameter seismic inversions. This formulation, implemented through automatic differentiation, is applied to an underground gas storage scenario, showing improved recovery of saturation-related variations. Furthermore, frequency-dependent effects arising from patchy gas saturation are analysed, with emphasis on wave-induced fluid flow and resulting dispersion/attenuation in the seismic band. The results demonstrate that intermediate gas saturations can maximize signal attenuation and must be incorporated into the inversion schemes to enhance feasibility studies for monitoring purposes. In summary, this PhD thesis highlights how DL and physics-driven methods can be integrated to handle complex time-lapse seismic datasets, extracting crucial subsurface information for CCS and UHS monitoring, thereby supporting safer and more scalable geophysical reservoir assessment and monitoring.

The PhD thesis discusses the integration of data-driven approaches such as Deep Learning (DL) with physics-driven methods including Full-Waveform Inversion (FWI) and rock-physics modelling to advance geophysical monitoring for Carbon Capture and Storage (CCS) and Underground Hydrogen Storage (UHS). Motivated by the urgent need to reduce atmospheric carbon-dioxide (CO2) and to enable safe subsurface storage technologies, this work emphasizes the advantages of DL in geophysics in improving interpretation accuracy, reducing subjectivity, and enabling cost-effective time-lapse surveys. We developed a data-driven workflow to map time-lapse seismic gathers directly to CO2 saturation. A U-Net architecture enriched with multi-scale features via the Continuous Wavelet Transform (CWT) improves sensitivity to subtle and spatially variable plume signatures, providing more accurate fluid saturation estimates. A cost-effective monitoring strategy is then introduced based on pre-trained convolutional encoders arranged in a Siamese architecture to extract baseline versus monitor features. UMAP projects the high dimensional embeddings to a lower dimensional space, enhancing time lapse evolution tracking and supporting rapid anomaly screening and detection. This framework highlights baseline-monitor differences and reduces the need for dense data acquisition and manual screening. This approach is at first demonstrated on synthetic datasets, and then validated on a time-lapse Distributed Acoustic Sensing (DAS) seismic dataset from a controlled CO2 injection at the Svelvik CO2 Field Lab, Norway. As part of the PhD program, I led the SBEM (Svelvik Borehole Electromagnetic Monitoring) project in collaboration with SINTEF, during which both electromagnetic (GPR and ERT) and seismic (cross-well and VSP) data were acquired using hydrophones and Distributed Acoustic Sensing (DAS). This campaign provided the experimental dataset used for validating the proposed monitoring framework. The results demonstrated robust injected fluid detection at lower acquisition density and provided a real-world testbed for evaluating plume evolution. Furthermore, the thesis introduces a rock-physics-parametrized FWI that performs inversion directly for a petrophysical parameter, such as saturation, thereby mitigating the cross-talk typical of multi-parameter seismic inversions. This formulation, implemented through automatic differentiation, is applied to an underground gas storage scenario, showing improved recovery of saturation-related variations. Furthermore, frequency-dependent effects arising from patchy gas saturation are analysed, with emphasis on wave-induced fluid flow and resulting dispersion/attenuation in the seismic band. The results demonstrate that intermediate gas saturations can maximize signal attenuation and must be incorporated into the inversion schemes to enhance feasibility studies for monitoring purposes. In summary, this PhD thesis highlights how DL and physics-driven methods can be integrated to handle complex time-lapse seismic datasets, extracting crucial subsurface information for CCS and UHS monitoring, thereby supporting safer and more scalable geophysical reservoir assessment and monitoring.