A method for predicting the TOC in source rocks using a machine learning-based joint analysis of seismic multi-attributes

Total organic carbon (TOC) is an important geological feature of source rocks, which is typically estimated based on expensive core or logging data. However, as oil and gas exploration move into low exploration areas (deep layers, deep water, or new exploration areas), it is becoming increasing significant to predict the distribution of source rocks in these areas to reduce drilling costs. Traditional seismic inversion methods based on well logging and seismic data are difficult to implement in low exploration areas due to a lack of evaluation parameters and well logging data of hydrocarbon source rocks. Seismic data contains rich information about underground media. There are more diverse ways to extract information from seismic data as seismic attribute analysis and machine learning techniques advance. Seismic attributes can be classified into two categories: attributes calculated by mathematical methods, such as instantaneous amplitude, Energy Half Time (EHT), and instantaneous frequency; and physically meaningful attributes obtained through geophysical calculations, such as Elastic Impedance (EI), & lambda;, Young's modulus and & mu;, etc. In the case of the L Depression in the South China Sea, the EHT, EI, density, and & lambda; are sensitive attributes of source rocks. Using Gradient Boosted Decision Trees (GBDT), eXtreme Gradient Boosting (XGBoost), Random Forests (RF), Hist-Gradient Boosted Decision Trees (HGBDT), and Local Cascade Ensemble (LCE) methods, we predicted the TOC of source rocks. The prediction results show that it is feasible to predict TOC of source rocks based on seismic attributes, and the prediction accuracy rate is above 80%, which can provide good data support for further drilling deployment in areas with low exploration areas.