Introducing User-Control in Deep Learning Seismic Gather Conditioning for Quantitative Interpretation

Quantitative interpreters require accurate seismic data to estimate subsurface properties. Parameter-rich and time-demanding gather conditioning procedures are often a crucial step to ensuring the reliability of the seismic amplitudes. To that end, we propose a deep learning-based workflow that relieves users from such laborious manual tasks. Moreover, user-influence in challenging scenarios is ensured through the strength parameter in the context of multiple removal and via an iterative approach for increasingly finer adjustments in the context of primary alignment. For the demultiple task, we employ a conditional U-net topology trained on synthetic data. This model takes a seismic gather and a strength parameter as inputs and outputs a seismic gather freed of multiples, whose move-out falls below the threshold set by the parameter. On the other hand, our primary alignment model outputs time shift fields given misaligned seismic gathers. The iterative application of the proposed model allows for greater control over the process, with each successive application achieving a more nuanced effect. The proposed pipeline has yielded promising results across a range of scenarios, demonstrating the generalizability of our models on real datasets. It consistently ensures the preservation of true amplitudes and exhibits versatility in various gather domains.