Automatic Borehole Comparison Technology Based On Machine Learning

The information of rock and soil layer obtained from borehole data is the basis of studying the distribution of strata.However,complete stratigraphic information cannot be obtained directly.How to simulate stratigraphic structure by limited borehole data is an important topic in geosciences.The traditional method is to use the spatial interpolation based on geostatistics to simulate the continuous stratigraphic distribution.There are many kinds of spatial interpolation methods.The choice of spatial interpolation method according to experience,and the effect is affected by subjective judgment.Machine learning has been widely successful in multiple applications,offers great potential for solving geoscience problems.The combination of geosciences and machine learning is beneficial to the development of both disciplines.This paper introduces two different borehole comparison methods based on machine learning algorithm to simulate formation lithology in the unexplored area.It is difficult to directly use borehole data for machine learning algorithms.Therefore,both methods reconstruct the borehole data so that the data can be used for machine learning modeling.The borehole sequence simulation method reconstructs the borehole data into lithology sequence and stratum thickness sequence.Two recurrent neural network models were built to simulate the lithology sequence and stratum thickness sequence respectively.Finally,the two simulation sequences were combined into borehole data,the formation lithology simulation is completed.According to the tests,the model works best when the network is built on a bidirectional GRU.The prediction is relatively accurate.The point lithology prediction method reconstructs the borehole data into point data and classify lithology of point by machine learning model.Comparing the accuracy of decision tree,random forest and XGBoost classifier,random forest performed the best.But the effect of classifier in different boreholes is not stable.Both methods can be used to simulation boreholes,but how to improve the scalability to adapt to different condition still needs further research.