Toward Combining Multi Seismic Attributes, A Novel Machine Learning Approach

Abstract

AbstractThe objective of this study is to develop and examine a machine learning model that combines multiple seismic attributes in a way that preserves all the features of the combined attributes with, hopefully, better resolution and efficiency than the conventional seismic attributes integration method. We used the Auto- Encoder algorithm, which is a neural network method. It consists of two identical networks known as encoder and decoder. Auto-Encoder was initially developed to compress data in the encoder part, and then the same structure is used to decompress data in the decoder part. Our interest is in the output of the encoder which is the reduced version of the input.For fault and fracture detection, we examined and compared the results of Octree as a conventional multi- seismic quantization method and Auto-Encoder as a new machine learning technique. The study used five 3D volumes, one for each of the five attributes Raw Data, Skeleton, Detect, Negative Curvature 1 and Negative Curvature 3. Each 3D volume is divided into 100 slices. One slice per attribute (i.e. a set of five slices) comprise an input dataset to the encoder. The process is iterated 100 times to obtain 100 slices of encoded attributes.The result of the Auto-Encoder solution is compared with that of Octree. The Auto-Encoder accurately preserved the three faults from the input attribute images whereas Octree preserved only one fault. Additionally, the details of the fractures in the Curvature attributes are also preserved with more clarity in the Auto-Encoder solution than Octree. The machine learning technique (Auto-Encoder) preserves all the features (i.e. faults and fractures) from the combined attributes and shows better resolution than Octree. It can also combine unlimited number of attributes whereas Octree is limited to eight attributes only.