Deep carbonate reservoir characterisation using multi-seismic attributes via machine learning with physical constraints

Abstract Seismic characterisation of deep carbonate reservoirs is considerable interest for reservoir distribution prediction, quality evaluation and structure delineation. However, it challenging to use the traditional methodology predict a deep-buried because highly nonlinear mapping relationship between heterogeneous features seismic responses. We propose machine-learning-based method (random forest) with physical constraints enhance prediction performance from multi-seismic attributes. demonstrate effectiveness this on real data application in Tarim Basin, Western China. first perform feature selection attributes, then four kinds constraint (continuity, boundary, spatial category constraint) transferred domain knowledge are imposed process model building. Using constraints, F1 score type can be significantly improved combination effective gives best performance. also apply proposed strategy 2D type. The results provide reasonable description strong heterogeneity reservoir, offering insights into sweet spot detection development.