Use deep transfer learning for efficient time-series updating of subsurface flow surrogate model

In subsurface flow , history matching is a complex high-dimensional sampling problem demanding numerous computationally intensive numerical simulations, which result in high costs. To mitigate this, various deep learning-based surrogate-assisted techniques are employed to approximate the original numerical simulation processes for more efficient computation. Despite the widespread application of deep learning-based surrogate models , significant limitations remain, particularly the need for the frequent reconstruction of surrogate models when new observed data become available. This process requires additional high-fidelity simulation data, further increasing computational burden of history matching. To tackle this problem, we introduce a transfer learning-based framework within an image-to-sequence surrogate model to enable efficient updates as new data become available. This surrogate model consists of a geological feature extraction module and a sequence data regression module. To facilitate rapid learning of new tasks while retaining previously learned knowledge, we employ a decoupled learning strategy where the backbone structure remains fixed, and only the sequence data regression module is fine-tuned. The surrogate model is subsequently used for posterior sampling of the geological model using the Randomized Maximum Likelihood (RML) method. The proposed workflow is evaluated on both two-dimensional and three-dimensional oil-water flow models. Experiments indicate that our surrogate model achieves a Root Mean Square Error (RMSE) below 0.08 and a Coefficient of Determination (R 2 ) above 0.97. Moreover, by leveraging transfer learning , the model attains comparable predictive accuracy with only 40 % of the original training data. The results demonstrate the surrogate model's capability to reduce computational costs while preserving accuracy.