Search

In this paper two different geophysical electro-magnetic techniques, long-offset time domain EM and the magnetotelluric method, are jointly inverted using a variation of Bayesian inversion called Hybrid Monte-Carlo. Long-offset time domain EM is a method that records Earth response to a controlled source square wave signal, while the magnetotelluric method records Earth response to naturally occurring electro-magnetic signals. We obtain better understanding of subsurface resistivity structures because the information they provide is complementary and joint inversion reduces the set of permissible inversion models. In most situations where long-offset time domain EM is recorded magnetotelluric data is also collected, since receiver instrumentation is shared. Hybrid Monte Carlo uses the gradient of the forward model to enable larger jumps through the solution space than are permissible using the standard Metropolis-Hastings algorithm, avoiding the pitfall of random walk behavior and giving faster convergence. The method is successfully applied to a survey area near the edge of Wuhan China. We obtain inversion results in a highly urbanized area, where the magnetotelluric method is typically compromised by background electro-magnetic noise. • HMC reduces random walks and gives faster convergence. • The joint-inversion of MT and LOTEM combines complimentary EM responses. • Problems with difficult combinations of conductors and resistor would benefit. • This is the first joint inversion of EM data using HMC. [ABSTRACT FROM AUTHOR]

The reservoir rocks in the volcanic strata of the Jilin oil field are characterized by great complexity and diversity in composition and structure of lithology. To enhance the rate of lithology identification in subsurface is very laborious. However, lithology identification is often ignored in quantitative studies, though it is the basis for reservoir characterization. In this paper, an ensemble learning algorithm named gradient boosting decision tree (GBDT) was used to establish the classification model for the volcanic lithology identification of the Lower Cretaceous Yingcheng Formation in the Songliao Basin, NE China. At the same time, support vector machine (SVM), logistic regression (LR) and decision tree (DT) classification models were also adopted in contrast with the classification accuracy of GBDT model. Subsequently, the optimal key parameters for each model were determined by employing validation curves and GridSearchCv. These results indicate that the GBDT model is superior to the single classifier and can accurately distinguish the lithologic interface of breccia tuff and rhyolite. Moreover, it also has better recognition ability for thin layer. It was concluded that the ensemble learning algorithm GBDT has significantly enhanced the accuracy of lithology identification and can be used as a lithologic identification technology. • The GBDT model is applied well in volcanic lithology identification. • Parameter optimization is conducted to optimize the classifier. • The GBDT model is superior to the single classifier. • The GBDT model and has better recognition ability for thin layer. [ABSTRACT FROM AUTHOR]