Automatic Adaptive Subtraction with Unsupervised Machine Learning

The adaptive subtraction is an important step of the multiple attenuation process. Conventional approach involving an adaptive matching filter design and adjustment is often complicated and time consuming. Existing machine learning algorithm based on the supervised structures have shown some potentiality, but facing limitations with the quality of training data, and requires human intervention for good performance and avoiding signal leakage. Being motivated to overcome those limitations of existing supervised algorithms, we introduce an unsupervised ML solution to design and apply optimized nonlinear matching filter for adaptive subtraction. Unlike the supervised workflow, it takes in both the actual seismic response and initial multiple model, automatically estimates both the optimal time-shift and the spatial-varying scaling factors that correct for the amplitude difference between the input models and observed data,, and applies them to transform the input multiple model into an optimal form with similar characteristics of the observed seismic responses. Besides, the algorithm allows introducing different evaluation metrics as additional constraints to tackle the primary leakage, which was not feasible using the conventional approaches. In a demonstrating synthetic dataset, we observed better results by the ML algorithm when comparing to the traditional approach, both qualitatively and quantitatively.