Abstract

Accurate subsurface characterization is critical for various energy and environmental applications, including, but not limited to, oil and gas exploration, geothermal energy, carbon storage, and hydrogen storage. The success of such projects is highly dependent on the ability to estimate the key properties of rocks and fluids with reduced uncertainty combining data types at different scales. This thesis presents a series of hybrid methodologies that integrate physics-based models with data-driven techniques, particularly Machine Learning (ML), to enhance the estimation of petrophysical and elastic properties using seismic surveys, well-logs, time-series measurements of pressure and production/injection rates, and CT scan images of rock samples.

First, a data-driven framework is developed to estimate petrophysical properties directly from pre-stack seismic data. The method employs optimal basis functions derived from well-log information to map seismic data into band-limited petrophysical reflectivities. These reflectivities are then individually inverted through regularized post-stack inversions to recover properties such as porosity, shale content, and water saturation. In addition, a Bayesian extension of the framework combined with Generative Models is proposed to quantify the uncertainty associated with the inverse process and promote geological realism in the solutions. Applications to synthetic and field datasets confirm the method’s ability to recover petrophysical parameters accurately.

Next, building on the Variational Inference paradigm, a gradient-based ensemble method is applied to Full-Waveform Inversion to address the challenges related to the nonlinearity and multi-modality of the problem. The approach incorporates a multiscale optimization scheme and strategies to improve convergence and obtain a more effective exploration of the posterior distribution. Moreover, the methodology is extended to Reservoir History Matching using time-series production and pressure data. Gradients are computed via automatic differentiation within a fully differentiable reservoir simulator, which enables efficient updates to generate an ensemble of plausible models that align with observed data.

Finally, two additional reservoir characterization approaches are proposed at distinct spatial scales to complement these previous methods that operate a the reservoir scale. First, a neural network trained on coupled wellbore–reservoir simulations is employed to correct phase redistribution effects in pressure buildup tests; this enhances the reliability of pressure transient analysis in high water-cut wells. Second, a generative model is developed to produce digital rock samples for pore-scale flow simulations in scenarios where physical cores are unavailable. All of the methodologies presented highlight the importance of integrating physics-based and data-driven approaches to enhance subsurface property estimation and deliver efficient, scalable uncertainty quantification across diverse geological settings and data types.

Committee