Multisensor Geophysical Data Fusion for Real-Time ReservoirMonitoring via AI

Thesis Abstract

The exploitation of subsurface reservoirs increasingly relies on integrated geophysical sensing and data-driven interpretation to enable real-time decision making for enhanced recovery and risk mitigation. This study addresses the persistent gap between multi-sensor measurements (seismic, electromagnetic, gravity, and borehole logs) and actionable reservoir state estimation under dynamic operational conditions. The aim is to develop a multisensor data fusion framework powered by artificial intelligence that delivers continuous, probabilistic reservoir property maps and predictive alerts for transient events. Specific objectives include (1) constructing a unified data-collection schema that harmonizes time-aligned seismic, electromagnetic, gravity, and core/borehole datasets from a 12-month monitoring campaign across a multi-well field; (2) designing and validating a hybrid fusion model that combines physics-informed deep learning with Bayesian data assimilation to yield real-time estimates of porosity, permeability, fluid saturation, and pressure fields; (3) assessing model robustness under varying noise levels, sensor outages, and reservoir-management scenarios; and (4) evaluating operational impact through decision-support metrics such as forecast lead time, false-alarm rate, and economic risk reduction. Methodologically, the research adopts a sequential explanatory mixed-methods design. The population comprises operational reservoirs within a sedimentary basin characterized by heterogeneous lithology and ongoing production optimization. A stratified random sample of 8 wells and 3 monitoring arrays yields a dataset of approximately 2,400 time-stamped geophysical measurements, combined with 1,200 control-volume reservoir state observations derived from production telemetry and well-test data. Data collection instruments include time-lapse 3D seismic surveys, magnetotelluric/electrical resistivity surveys, gravity gradient measurements, borehole sonic and neutron logs, along with conventional production indicators. Data preprocessing encompasses calibration, co-registration to a common voxel grid, and noise-conditioning using wavelet denoising and anomaly-aware normalization. The fusion model integrates a physics-informed convolutional neural network (CNN) for spatial feature extraction with a recurrent neural network (RNN) temporal module, augmented by a Bayesian hierarchical operator to quantify uncertainty. Model specification leverages transfer learning from regional geological priors and enforces consistency with reservoir simulators through an adjoint-based optimization. Validation employs k-fold cross-validation and a held-out 6-month test set, with performance metrics including RMSE, MAE, R-squared for property fields, and Brier score for probabilistic forecasts. Sensitivity analyses examine the impact of sensor failure scenarios and alternative weighting schemes in the fusion kernel. Statistical inference for hypothesis testing employs regression-based significance testing and Bayesian model comparison using Bayes factors. The anticipated findings indicate that the integrated fusion framework will yield high-fidelity, near real-time maps of porosity, permeability, and fluid saturation, with reduced uncertainty intervals compared to single-sensor interpretations. It is expected that the AI-driven approach will outperform baseline physics-only models in forecasting transient reservoir responses (e.g., pressure drawdown and saturation fronts) with improved lead times of 12–36 hours under typical production perturbations. The study also anticipates identifying critical sensor combinations and data gaps whose reinforcement yields disproportionate gains in predictive accuracy, thereby informing intelligent sensor deployment strategies. The contribution to knowledge encompasses (i) a novel, scalable methodology for multisensor geophysical data fusion in real-time reservoir monitoring that explicitly propagates uncertainty; (ii) an integrative theoretical framework combining physics-informed machine learning with Bayesian data assimilation for subsurface applications; and (iii) practical guidelines for implementing AI-assisted reservoir surveillance in industrial settings, including data governance, computation workflows, and decision-support interfaces. Conclusion emphasizes that real-time AI-enabled data fusion significantly enhances reservoir awareness, enabling proactive management and optimized recovery while mitigating operational risk. Recommendations include extending the framework to include microseismic and chemical sensing modules, integrating active learning for continual model improvement, and developing standardized benchmarks and open datasets to facilitate cross-field replication and validation.