Estimating subsurface properties using a semisupervised neural network approach

Abstract

Estimating static rock properties (e.g., density and porosity) from seismic and well logs is one of the essential but challenging tasks in subsurface interpretation and characterization. To compensate for the sparsity of well logs and the limited bandwidth of seismic data, a semisupervised learning workflow is used for efficiently integrating seismic and logs and simultaneously estimating multiple subsurface properties. It consists of two components: (1) unsupervised seismic feature engineering and (2) supervised seismic-well integration, each of which is implemented as a convolutional neural network. Compared to most existing methods, it is advanced in three aspects. First, it allows the use of local 3D seismic patterns for building an optimal nonlinear mapping function with 1D logs, which is more noise robust and significantly improves the lateral consistency of machine prediction throughout the entire seismic survey. Second, it is capable of automatically bridging the gap of vertical resolution between seismic and well logs, which simplifies the workflow of data preparation, such as log upscaling. In addition, it enables Monte Carlo dropout-based epistemic uncertainty analysis. The performance of our solution is evaluated through two examples: relative acoustic impedance and porosity estimation in a synthetic PreSDM data set of 36 pseudowells and sonic and density estimation in the Groningen data set of 375 wells. The good match between the machine predictions and the actual measurements demonstrates the capability of our semisupervised learning method in providing reliable seismic and well integration and delivering robust estimation of subsurface properties, including those of a relatively weak physical link with seismic, such as density and porosity.