Designing Reality-Based VR Interfaces for Geological Uncertainty

TL;DR

Geological data uncertainty yields large ensembles of realizations that are expensive to evaluate with full-flow simulations; the paper presents a VR-based uncertainty analysis system to identify representative models that capture ensemble variability. It integrates a variance-based VOI workflow with similarity via mutual information, 3D MDS projection, and kernel $k$-means clustering to select representative realizations, implemented through reality-based 3D interactions. The contributions include a data model, VR interaction design (including a view-dependent cutaway lens and a body-relative clustering graph), and a task-driven user study with 12 reservoir engineers, culminating in design recommendations. The work demonstrates the potential of immersive interfaces to streamline geological uncertainty analysis while highlighting practical considerations for portability and workflow integration in industry contexts.

Abstract

Inherent uncertainty in geological data acquisition leads to the generation of large ensembles of equiprobable 3D reservoir models. Running computationally costly numerical flow simulations across such a vast solution space is infeasible. A more suitable approach is to carefully select a small number of geological models that reasonably capture the overall variability of the ensemble. Identifying these representative models is a critical task that enables the oil and gas industry to generate cost-effective production forecasts. Our work leverages virtual reality (VR) to provide engineers with a system for conducting geological uncertainty analysis, enabling them to perform inherently spatial tasks using an associative 3D interaction space. We present our VR system through the lens of the reality-based interaction paradigm, designing 3D interfaces that enable familiar physical interactions inspired by real-world analogies-such as gesture-based operations and view-dependent lenses. We also report an evaluation conducted with 12 reservoir engineers from an industry partner. Our findings offer insights into the benefits, pitfalls, and opportunities for refining our system design. We catalog our results into a set of design recommendations intended to guide researchers and developers of immersive interfaces-in reservoir engineering and broader application domains.

Figures (8)

• Figure 1: Illustration of a reservoir model (top) and common well profiles (bottom). Corner-point gridding is used by the oil industry as it can represent reservoir features (e.g., faults, fractures) by specifying eight corners of each grid cell. Each cell can be associated with static and dynamic attributes (e.g., rock type, permeability, and oil saturation).
• Figure 2: Cross-section and $45^\circ$ view on CMG SuiteCMG.
• Figure 3: Geological uncertainty analysis workflow. (1) Color-coded variance model from the input reservoir ensemble. (2) VOI selection. (3) VOI-based clustering produces color-coded clusters and a default representative set. (4) Clustering analysis identifies additional samples.
• Figure 4: Deletion (left) and scaling (right) operations are grounded in universally recognizable gestures, designed to promote rapid learnability with low cognitive load, potentially enabling eyes-free execution.
• Figure 5: To rotate the reservoir model, we compute the geodesic arc on $\mathbb{S}^2$ connecting two controller positions---those at the moments of press $p_0$ and release $p_1 \in \mathbb{R}^3$. These positions are projected onto a virtual sphere centered at the model's origin $c$, resulting in two normalized direction vectors $\vec{v}_0$ and $\vec{v}_1$. The rotation axis $\vec{a}$ is computed as the normalized cross product of these vectors, $\vec{a} = \widehat{\vec{v}_0 \times \vec{v}_1}$, which defines the axis orthogonal to the plane $\mathcal{P}$ spanned by the non-colinear pair $\vec{v}_0$ and $\vec{v}_1$. The angular displacement $\theta$ between the two vectors is calculated using the arccosine of their dot product as $\theta = \arccos(\vec{v}_0 \cdot \vec{v}_1)$. Finally, a rotation quaternion $q(\vec{a}, \theta)$ is constructed from this axis-angle representation and applied to the model's orientation.