ML-based Research


Artificial intelligence, and specifically, machine learning (ML) has recently emerged as a powerful tool to address many of the challenges we face in life. We aim at SWAG to be a leading hub for utilizing innovated ML methods, beyond the conventional supervised learning, to help us use recorded seismic data to illuminate the Earth at all scales and make the proper prediction of its content. From fault detection to salt boundary picking to image resolution enhancements, the quest to teach our computing devices how to perform these tasks accurately, as well as, quantify the accuracy has become our main objective.

Utilizing recent advances in machine learning algorithms and the availability of the modules to apply such algorithms, we have managed to focus on developing solutions on a wide range of seismic exploration applications, with a special focus on waveform inversion. We share some of these applications below.

Page_Component2

Self-supervised blind-spot networks remove the requirement of labelled training data but can only suppress fully random noise. Adaptations have been proposed for targeting noise types with a specific, constant correlation however until now none tackle the noise field as a whole.

A scheme to improve resilience of self-supervised denoising procedures to pseudo-random data by incorporating a segmentation scheme into the learning process.

Developing a DNN method to extrapolate high-wavenumber model components for band-limited FWI results.

An automatic approach for microseismic event location which is feasible for anisotropic media and irregular receiver sampling.

Using PINN to overcome the inability of conventional methods when handle large areas of missing information (gap) in the velocity model.

Capturing and storing the local and global features of seismic data in the pre-training stage.

Denoising field seismic data with neural networks.

An implementation for knowledge transfer between synthetic and field seismic data applications of low-frequency data extrapolation.

A deep learning approach for post-stack trace-by-trace matching to reduce the remaining 4D noise

An ensemble of convolutional neural networks (CNNs) for building velocity models directly from the data.

A target-oriented strategy through seismic redatuming to reduce the computational cost yet achieving high-resolution in the result.