Rigid transformations for stabilized lower dimensional space to support subsurface uncertainty quantification and interpretation

Subsurface datasets commonly are big data, i.e., they meet big data criteria, such as large data volume, significant feature variety, high sampling velocity, and limited data veracity. Large data volume is enhanced by the large number of necessary features derived from the imposition of various features derived from physical, engineering, and geological inputs, constraints that may invoke the curse of dimensionality. Existing dimensionality reduction (DR) methods are either linear or nonlinear; however, for subsurface datasets, nonlinear dimensionality reduction (NDR) methods are most applicable due to data complexity. Metric-multidimensional scaling (MDS) is a suitable NDR method that retains the data's intrinsic structure and could quantify uncertainty space. However, like other NDR methods, MDS is limited by its inability to achieve a stabilized unique solution of the low dimensional space (LDS) invariant to Euclidean transformations and has no extension for inclusions of out-of-sample points (OOSP). To support subsurface inferential workflows, it is imperative to transform these datasets into meaningful, stable representations of reduced dimensionality that permit OOSP without model recalculation.

We propose using rigid transformations to obtain a unique solution of stabilized Euclidean invariant representation for LDS. First, compute a dissimilarity matrix as the MDS input using a distance metric to obtain the LDS for \(N\)-samples and repeat for multiple realizations. Then, select the base case and perform a rigid transformation on further realizations to obtain rotation and translation matrices that enforce Euclidean transformation invariance under ensemble expectation. The expected stabilized solution identifies anchor positions using a convex hull algorithm compared to the \(N+1\) case from prior matrices to obtain a stabilized representation consisting of the OOSP. Next, the loss function and normalized stress are computed via distances between samples in the high-dimensional space and LDS to quantify and visualize distortion in a 2-D registration problem. To test our proposed workflow, a different sample size experiment is conducted for Euclidean and Manhattan distance metrics as the MDS dissimilarity matrix inputs for a synthetic dataset.

The workflow is also demonstrated using wells from the Duvernay Formation and OOSP with different petrophysical properties typically found in unconventional reservoirs to track and understand its behavior in LDS. The results show that our method is effective for NDR methods to obtain unique, repeatable, stable representations of LDS invariant to Euclidean transformations. In addition, we propose a distortion-based metric, stress ratio (SR), that quantifies and visualizes the uncertainty space for samples in subsurface datasets, which is helpful for model updating and inferential analysis for OOSP. Therefore, we recommend the workflow's integration as an invariant transformation mitigation unit in LDS for unique solutions to ensure repeatability and rational comparison in NDR methods for subsurface energy resource engineering big data inferential workflows, e.g., analog data selection and sensitivity analysis.