Machine learning in seismic exploration

Seismic exploration is a method for the geophysical characterization of the subsurface. Typically, in marine environments, this process begins with the acquisition of seismic data, which are then digitally processed to generate a three-dimensional image of the Earth's subsurface. Analogous to how a CT scan is used in medical imaging, where a doctor interprets the data and makes a diagnosis, geoscientists interpret seismic images to characterize the subsurface. In both fields, there is a significant push towards automating these manual tasks to improve both accuracy and efficiency. Recently, using machine learning (ML) techniques has become a standard approach to improve these processes. Our work on ML for seismic interpretation is twofold: first, we design the algorithms to perform the interpretation task, and then we develop its efficient software implementation.

Developing our own methods includes adopting the latest research from academia, particularly tuning the methods to be robust for the large variance of real-world data. We use ML algorithms from computer vision, such as convolutional neural networks and vision transformers, which we blend with traditional image processing algorithms. By implementing these methods into commercial-grade applications, we enable our geoscientist users to train their models either from scratch or based on our pretrained 'seismic-foundation model'. This foundational model serves as a versatile backbone supporting various decoders and segmentation heads tailored to specific tasks. For instance, one type of decoder may specialize in fault interpretation, while another might be dedicated to seismic horizon interpretation.

The second aspect of our work concentrates on making robust and efficient software solutions. Since seismic datasets typically span from 10GB to 1TB, high-performance computing is critical for any practical application. This necessity leads us to develop low-latency data pipelines and utilize the advanced capabilities of contemporary machine learning frameworks for multi-GPU and mixed-precision training and inference.

Our deployment strategy, which encompasses both cloud-based and on-premises solutions, presents unique architectural challenges. Particularly in the cloud, we employ Kubernetes in conjunction with Argo to develop workflows that can scale elastically, addressing computational demands.