Seismic Processing and Machine Learning: What is my Data Space Distribution?

Machine Learning (ML) is now a reality in seismic processing, offering improved results, shorter processing turnaround times, and easier applications. Supervised ML tools are trained on pairs of input and desired data, resulting in more predictable and understandable ML models behaviour. ML models, even if not generalists, can be operationally applied to unseen real datasets using reversible data transformations.

The preparation of the training dataset is essential because standard supervised ML models struggle with extrapolation and generalization. To mitigate overfitting and improve generalization, we use data augmentation techniques and synthetic data, adjust network hyperparameters, and introduce custom loss functions to mimic physics.

Once the ML model is trained, the training set distribution is fixed, but seismic data may not match it. Direct transformation processes can be used to adjust the distribution of the application dataset prior to the ML tool application. To keep original seismic characteristics, these transformations must be reversible after the ML model application. This is illustrated with an industrial 2D ML multiple removal tool applied on the Penobscot dataset. The ML model was trained with pseudo-synthetic data, and the application of reversible geometrical transformations on the real dataset shows an improved multiple removal.