Restructuring deep learning for sustainable subsurface data analysis
Contact person: Anita Torabi
Keywords: supervised machine learning, 1D and 3D deep learning, automatic seismic interpretation, transfer learning, pretrained deep learning models
Research group: Sedimentary Basins
Department of Geosciences
Understanding geological heterogeneities that result from geological structures are important research areas for many applications such as CO2 storage underground; geothermal energy management; and geological hazard studies. The application of Deep Neural Networks (DNN) in geoscience is mainly in the seismic interpretation related tasks (Fig. 1, a light U-Net from Bonke, 2022). The size of data used for training in the interpretation of subsurface data (e.g. seismic data) is not as large as the other disciplines, where DNN algorithms were originally developed. In addition, the architecture of most of networks, and hyperparameters are designed for prediction or classification of problems with less complexity and a lot more training data compared to subsurface data. The objective of the research is to modify the architecture of Deep Neural Networks to fit classification of geological structures. This would be in collaboration between geosciences and mathematics departments at UiO.
Methodological research topics:
- Deep Neural Network selection and restructuring
- Revisiting training parameters, model parameters, optimization, and loss functions
- Geological verification of results
Topics from natural sciences or technology:
- Sustainability
- CO2 storage underground
- Green transition
- Reservoir Engineering
External partners:
- The Norwegian Computing Center (NR)
