MST-Direct at Scale: Multivariate and Conditional Geostatistical Simulation via Sinkhorn Optimal Transport
arXiv:2606.07578v1 Announce Type: new Abstract: This paper extends MST-Direct, a Matching-via-Sinkhorn-Transport approach for multivariate geostatistical simulation, from the original bivariate, unconditional, small-grid formulation to multivariate, conditional, and large-grid settings. We address the three main limitations identified in the original work: (i) scalability beyond a few thousand nodes through a sparse, candidate-restricted Sinkhorn matcher with O(nC) memory complexity; (ii) extension to multiple variables by matching target value tuples onto an independent FFT-MA Gaussian backbo
This paper addresses prior limitations of geostatistical simulation methods by extending MST-Direct to multivariate, conditional, and large-grid settings, making it directly applicable to real-world, large-scale problems.
Improved geostatistical simulation at scale is critical for AI applications requiring spatial-temporal data analysis, impacting fields from resource management to climate modeling and advanced robotics.
The ability to run advanced geostatistical simulations with greater scalability and complexity means more accurate and efficient AI models for spatially correlated data, opening new avenues for scientific discovery and practical application.
- · AI/ML researchers
- · Geospatial data scientists
- · Resource exploration companies
- · Environmental modeling agencies
- · Legacy geostatistical software developers
- · Organizations reliant on less efficient simulation methods
More accurate and scalable geostatistical simulations fuel better predictive AI models in various domains.
Enhanced modeling capabilities lead to more efficient resource allocation, improved environmental predictions, and new approaches in robotics.
The widespread adoption of such techniques could accelerate the development of autonomous systems capable of understanding and interacting with complex spatial environments.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG