September 13, 2022
Domain conversion can be a challenging process as it entails multiple time-intensive, user expertise-dependent steps to ensure accurate results. The traditional method uses a step-by-step approach to predict the final depth with the help of processed inputs and qualitative analysis. However, minor discrepancies could occur in this approach that could result in uncertainty. A data-driven, artificial intelligence (AI) and machine learning (ML) approach can help address these challenges effectively.
In this project, the AI/ML approach has been considered for depth prediction by using the seismic attributes and amplitudes in time and depth domains. We found that this approach produced a faster and more efficient way of executing seismic domain conversion and it can be implemented in near real-time for an exploratory field.
We began by extracting and processing data where the region of interest had been cropped from the overall data. The data included 8 seismic attributes in the time domain along with amplitude data in time and depth domains. Following that, an index matching was done using the amplitude data, and time data was converted to depth data. This index-matching was done based on the minimum distance between the two traces. We adopted the Optimisation Method to make the process faster.
In the next step, we created a dataframe. All seismic attributes including the target variable i.e., the converted depth, which was the output of index matching, were merged as columns to form a single dataframe. Also, we concatenated the actual time column at a given range to build the final dataframe.
The model training stage is a key workflow in this process – it is imperative to ensure that we are using the right machine learning model corresponding to the data.
To train the model, we divided the final data into 75 percentage of train data and 25 percentage of test data, where for training purpose, we fit the model by providing input using scaling seismic attributes (X) to reduce the variation of values and target variable depth(y) of the training data. After trying different types of regression models, the final model selection was done based on higher accuracy, visualizing the predicted depth traces, and comparing actual and predicted patterns of depth with time amplitude on the blind data with different inline and crossline inputs provided by the SME. The R2 score (accuracy) of the final model, which was further saved and used for testing, was about 0.88 percentage.
After the model training, we saved the trained model for further predicting the depth of the blind dataset. As mentioned, the blind data provided by SMEs was used to check the model performance for different inline and crossline and for different ranges of depth and time. A segy volume was created between the given range of depth with a particular frequency using linear interpolation method with the depth data predicted.
Results of the blind test data:
The purpose of this POC was to convert seismic data from time-domain to depth-domain using Dynamic Time Warping or any Index Matching Approach and then predict depth by using all seismic attributes.
Based on the results we achieved, it can be concluded that the machine learning approach is one of the most effective ways which can be considered to meet the objective of the project with up to 88 percentage accuracy. Future exploration in domain conversion would be utilizing deep learning or other complex approaches to observe improvement in the results and deploying the models for real time usage.