Artificial intelligence methods for predicting hydrocarbon deposits from three-dimensional seismic images

Abstract

The application of artificial intelligence to hydrocarbon reservoir prediction based on 3D seismic images has shown significant potential, offering the possibility of improving the efficiency, accuracy and reliability of exploration operations. A review of the state-of-the-art reveals a growing body of research utilizing various AI models, especially deep learning ar-chitectures such as CNNs, for tasks such as reservoir property prediction, lithology classifi-cation, and structural interpretation. However, several limitations and challenges remain, including issues with data scarcity and quality, the need to improve the generalizability and interpretability of AI models, and the effective management of complex geological structures. To overcome these limitations, future research should focus on several promising areas. The development and application of more complex AI models, such as hybrid architectures that combine the strengths of CNNs, RNNs, and Transformers, as well as the study of graph neural networks to exploit structural relationships in seismic data, deserve further investiga-tion. Physically-informed neural networks that integrate geophysical principles into the train-ing process have significant potential to improve the accuracy and physical plausibility of predictions, especially with limited data. Advances in data augmentation techniques, including the use of generative adversarial networks to generate realistic synthetic seismic data, and the development of effective transfer learning strategies are crucial to address the problem of limited labeled datasets. In addition, research into multiple example learning methods can allow AI models to learn from very small amounts of data, which is especially relevant in areas of new area exploration. Improving the interpretability and explainability of AI models is key to their widespread adoption in the industry. The application and further development of XAI methods such as LIME and SHAP, along with the analysis of attention mechanisms in deep learning models, can provide valuable insights into the decision-making processes of these complex systems. The design of internally interpretable model architectures and the development of advanced visualization techniques will also help build trust and facilitate the integration of AI into geo-scientists' workflows. The creation of larger, more diverse, and well-labeled reference 3D seismic datasets is important to ensure fair and reproducible comparisons of different AI models and method-ologies. Future research should also focus on the integration of multimodal data sources such as seismic data, well logs, geological maps, and remote sensing data using AI techniques to provide a more complete understanding of subsurface hydrocarbon potential. Finally, quanti-fying the uncertainty in AI-based hydrocarbon forecasting will be crucial to provide more re-liable risk assessments and support informed decision-making in exploration and develop-ment.