Methodology For Intelligent Lithology Identification In The Sandstone-type Uranium-bearing Formations

Nuclear energy,as a non-carbon-based energy source,is taking up an increasing proportion in the national energy structure,and the demand for uranium resources has increased dramatically.With the rapid development of in-situ ground leaching uranium mining technology,sandstone uranium ore has become the mainstay of uranium ore resource market supply.Continuing to explore and discover more sandstone uranium resources is an urgent need to reduce China’s dependence on the international uranium market and ensure national energy security.Since the formation of sandstone uranium ore is closely related to the lithological structure of the ore layer,fine and efficient identification of sandstone and mudstone interstratification structure is an important prerequisite for sandstone uranium ore exploration and prediction.However,due to the expensive cost of drill hole coring and lithology compilation,there is an urgent need to develop efficient lithology identification technology to realize the conversion of lowcost and easily accessible geophysical data to drill hole lithology and even 3D lithology structure with high accuracy.In recent years,the rapid development of machine learning technology has provided a new way for automatic lithology identification of mineral strata,but in the process of practical application,there are still some common problems that have not yet been solved.(1)There is an imbalance in the proportion of different types of lithologies in the formation,for example,the drilling results of the Sifangtai Formation in Daqing area show that the proportion of sandstone is about 58%,which far exceeds the proportion of other lithologies,and the training of machine learning models based on such samples often leads to difficulties in predicting small proportion lithologies.(2)The current automatic lithology identification models based on machine learning algorithms mainly rely on one-dimensional logging data for lithology identification along the well direction;such models can effectively capture the correlation between logging geophysical data and lithology in the vertical direction,but cannot consider the lateral correlation between lithology and geophysical data;while in the process of actual formation deposition,the lateral correlation between lithology and geophysical data is often The lateral correlation between lithology and geophysical data is often higher than the vertical correlation,and further consideration of the lateral correlation between different types of data is expected to significantly improve the accuracy and interpretability of the machine learning lithology identification model.In response to the above problems,this paper develops new machine learning algorithms to improve the accuracy of automatic lithology identification under the conditions of multivariate geophysical data fusion with the research objective of automatic lithology identification of typical uranium resource potential areas in Daqing area.The study achieves the following main insights:(1)The SMOTE algorithm is applied to balance the logging data and the corresponding lithology data,so that the number and proportion of each type of lithology in the training wells are basically the same.On this basis,a SMOTE-LSTM one-dimensional borehole lithology fine characterization model is built,with logging data as input data and lithology data as output data.After adjusting the model structure and hyperparameters and training,the average prediction accuracy and precision distribution of the four test wells can reach over 85% and 87%,and the prediction accuracy of a few types of lithology can be improved by about 20% compared with the LSTM model without the data processed by SMOTE algorithm.To further verify the effectiveness of the SMOTE algorithm,the LSTM model balanced by the SMOTE algorithm showed the highest lithology prediction accuracy and precision in all four test wells when compared with the Random Over Sampler and Random Under Sampler data balancing algorithms.(2)A logical 3D convolutional neural network model is proposed based on the correlation between geophysical data and lithology: the model takes 3D seismic data as the main input,3D lithology as the output,3D acoustic data as the intermediate constraint variable,and 3D convolutional neural network as the model architecture.The model structure and hyperparameters are adjusted and trained for validation,and the prediction accuracy and precision distribution of the model can reach 90.1% and 91.7%.The effective fusion of geophysical data such as logging data and seismic data and the effective identification of 3D lithology were achieved.The research has formed a lithology-logging data sample balance,high-precision identification of logging data,and automatic imaging of 3D lithology spatial structure as an integrated lithology identification algorithm for the ore formation,and developed a calculation program,which has been applied and verified in typical sandstone uranium resource potential areas in Daqing area,and can be extended to other resource potential areas for efficient automatic lithology identification.