Improving Reliability of Seismic Stratigraphy Prediction: Integration of Uncertainty Quantification in Attention Mechanism Neural Network

Abstract Despite technological advancement, subsurface studies continue to encounter uncertainties caused by structural complexities and data noise, which can result in inaccurate seismic interpretation and drilling locations. Although machine learning holds great potential by enabling the simultaneous analysis of large datasets, its effectiveness is often compromised by data noise and ambiguity, which can degrade the accuracy of the algorithms. Hence, this research incorporates uncertainty quantification into attention mechanism neural network to produce more reliable outcomes in seismic interpretation and stratigraphic mapping. The methodology is also benchmarked against other uncertainty quantification methods such as dropout and randomized ensemble techniques, followed by an evaluation using the Brier score.