Abstract
The global surge in energy demand has positioned subsurface energy storage and usage as critical solutions for sustainable energy transition and carbon mitigation, where accurate reservoir characterization plays a pivotal role. While data-driven approaches like deep learning show growing potential in resolving the spatial heterogeneity of subsurface reservoirs, their practical applications face limitations due to inadequate data diversity and unexplainable spatial pattern learning in black-box neural networks. In this work, we propose a robust knowledge-informed framework to enhance prior knowledge embedding and improve the characterization of spatial heterogeneity in black-box neural networks. To further improve the interpretability of deep-learning-based reservoir modeling, we investigate the role of implicit features and propose a knowledge-informed operation based on the automatic selection of implicit features. We develop the knowledge-informed neural network (KiNN) based on the proposed knowledge-informed framework. We enable KiNN to handle multiple conditioning data, including sparse observations and auxiliary variables. The proposed KiNN utilizes a novel composite loss function that balances data fidelity with geological plausibility. We conduct a series of experiments by using a shale sample of fractures and a field-scale case study. The experimental results confirm that KiNN can effectively reproduce reservoir structures with strong spatial heterogeneity. Additionally, the modular design of KiNN enables adaptation to hydrological simulations and geophysical inversion tasks.