Self‐Supervised Pre‐Training and Few‐Shot Finetuning for Gas‐Bearing Prediction

Natural gas remains the only fossil energy with sustained production growth amid global carbon reduction efforts, with seismic methods crucial for its exploration and development. Traditional seismic inversion methods, which detect gas indirectly by modeling and inverting gas‐sensitive parameters, can introduce cumulative errors, whereas conventional deep learning methods struggle with few‐shot generalization due to limited labeled data. This paper presents a deep learning workflow for directly predicting gas saturation from multiple seismic attribute data, involving pre‐training on large‐scale unlabeled data and finetuning with few labeled data to address the few‐shot challenge. Our network includes a Depthwise CNN1D for multi‐attribute feature extraction, an iTransformer for feature fusion, and a predictor for outputting the target. We use the iTransformer's self‐attention mechanism to calculate attribute weights for selection, and through extensive experiments, we developed a windowed multi‐attribute data input method that incorporates neighboring information to ensure lateral consistency. Based on the geological understanding that different attributes computed from the same sample convey correlated features reflecting geological properties, we pre‐train the network on large‐scale unlabeled data using a self‐supervised learning strategy of attribute masking and recovery. This approach encourages the network to learn correlation features, thereby improving its generalization by familiarizing it with the data distribution across the entire study field. We then employ the LoRA finetuning method to adapt the pre‐trained model to gas saturation prediction with meager labeled data while preserving pre‐trained knowledge. We applied this method to field data in the South China Sea, achieving accurate and generalized gas saturation predictions.