A physics-constrained two-stage GAN for reservoir data generation: enhancing predictive accuracy

Abstract Accurate modeling of reservoir properties and dynamics is crucial for efficient development, but the scarcity of high-quality data often limits the accuracy of characterization and prediction. Existing generative models can create static geological models or dynamic production sequences. However, they struggle to produce comprehensive, coupled data that is both high-fidelity and physically consistent. The key challenge is generating data that adheres to complex geological and fluid-flow physics and can reliably augment sparse real-world datasets. To address this, we propose a two-stage Physics-Constrained Conditional GAN (PC-CGAN). The first stage generates geologically realistic static reservoir properties. Conditioned on this static data, the second stage then generates the corresponding dynamic production time series. Our framework embeds differentiable physical constraints directly into the generators. The static generator uses these constraints to ensure geological rules are followed, while the dynamic generator employs an LSTM-based architecture and a hybrid loss function to maintain physical consistency in the time-series output. The experiments show that our framework generates high-fidelity static data and significantly improves downstream production prediction when used for data augmentation. Our model outperforms both traditional generative models and state-of-the-art diffusion models in physical consistency and predictive utility. In conclusion, this research demonstrates that embedding physical principles directly within a generative framework is key to creating high-utility synthetic data for complex engineering systems. This approach provides a practical and robust solution to data scarcity that promises to enhance the accuracy and reliability of real-world reservoir management.