MST-Direct: Matching via Sinkhorn Transport for Multivariate Geostatistical Simulation with Complex Non-Linear Dependencies
New algorithm tackles non-linear dependencies in geostatistics that traditional Gaussian methods fail to capture.
A new research paper introduces MST-Direct (Matching via Sinkhorn Transport), a novel algorithm designed to solve a critical problem in geostatistics: accurately simulating complex, non-linear dependencies between multiple geological variables. Authored by Tchalies Bachmann Schmitz, the method moves beyond traditional techniques like the Gaussian Copula and LU Decomposition, which often fail because they assume simple linear correlation structures. Real-world geological data, however, exhibits intricate patterns like bimodal distributions, step functions, and heteroscedastic relationships that these older models cannot faithfully reproduce.
MST-Direct's innovation lies in its application of Optimal Transport theory, specifically using the efficient Sinkhorn algorithm. Instead of modeling pairwise linear dependencies, it treats all variables simultaneously as a single, high-dimensional vector. This allows for 'relational matching' across the entire joint probability space, directly aligning complex source and target distributions while crucially maintaining the underlying spatial correlation structure of the data. The approach promises significantly more realistic simulations for fields like mineral exploration, reservoir modeling, and environmental science, where understanding these nuanced, multi-variable interactions is essential for accurate prediction and decision-making.
- Proposes MST-Direct, an algorithm using Optimal Transport & Sinkhorn to match complex multivariate distributions.
- Solves limitations of Gaussian Copula/LU methods by handling non-linear dependencies like bimodal & heteroscedastic data.
- Processes all variables as one multidimensional vector, preserving spatial correlations for more accurate geostatistical simulation.
Why It Matters
Enables more accurate modeling of real-world geological systems for resource exploration, environmental risk assessment, and reservoir forecasting.