This competition, will require that teams (1-3 people) develop a machine learning seismic inversion algorithm to predict and generate: compressional impedance, shear impedance, and density property volumes.
In reservoir characterization studies, Geoscientists are usually provided with sparse data (ground truths) from wells and areal datasets from seismic. Such 3D information includes post-stack and angle-stacks seismic amplitude data. In a nutshell, from seismic amplitude, attributes such as the elastic response via the various amplitude variation with offset (AVO) equations are estimated with adequate constraints from the well-data. This challenge aims to predict reservoir properties using training sets at well-locations and propagate the model in 3D using the areal information from seismic.
This process is very interpretative and requires a certain level of expertise by the users. In production environments, this is done by Reservoir Characterization Geophysicists or Quantitative Interpretation Geophysicists.
Skills include a basic understanding of wave propagation, forward modeling (synthetic generation), wavelet extraction, rock physics, and using multi-domain data to calibrate the log response to seismic responses. Then, conventionally, an inversion process can be done by using model-based optimization algorithms the extract elastic properties. These elastic properties are in turn used to predict other reservoir properties.
There is a level of uncertainty involved in seismic inversion mainly due to the band-limited nature of the seismic data, noise, well-to-seismic calibration uncertainty, non-stationarity of the wavelet, converted waves and multiples in seismic data, and the lack of wireline logs in the shallow zones of the earth. For this challenge, the non-stationary wavelet assumption, and well-to-seismic calibration uncertainty can be omitted.
Multiple wells and seismic data will be provided for the training dataset. Competitors are encouraged to be creative and use any ML method. However, typically seismic inversion is a good candidate for supervised methods. We also encourage competitors to apply physics constraints such as, but not limited to, having realistic rock-property relationships, obeying wave propagation physics, or honoring acquired field data when results are used for forward modeling a response.
Please visit our GeoML-SIG GitHub repository for example code, solutions, and leaderboards!