An adaptive data sampling strategy for in-situ stress characterization

Summary

Abstract

Deep geological repositories (DGRs) for nuclear waste require reliable in situ stress characterization to ensure safe construction, operation, as well as post-closure safety. However, determining optimal data requirements and sampling strategies remains challenging due to geological heterogeneity, measurement costs, and varying uncertainty tolerances across repository design phases. This study presents a novel adaptive framework that integrates change point detection (e.g., due to regional stress field variations, geological complex, etc.), Bayesian hierarchical modeling, and machine learning (ML) validation to optimize stress characterization strategies. Using an overcoring database from the Forsmark site in Sweden, we developed a multi-tiered methodology that: identifies statistically distinct segments, characterizes stress distributions within each segment, implements a decision tree framework to optimize selection of characterization methods, and quantifies data requirements. Our analysis reveals significant depth-dependent stress conditions, with optimal measurement zones at different depths and distinct statistical signatures suggesting segment-specific approaches are justified. The framework demonstrates that sample sizes vary dramatically across geological domains, from the minimum for simple unimodal distributions to the maximum measured records for complex multimodal stress fields requiring mixture models. Spatial coverage analysis indicates that achieving complete segment representation is more critical than increasing measurement density within segments. Implementation of ensemble Bayesian and ML characterization methods achieved 78% prediction accuracy with robust uncertainty quantification, while segment-based Bayesian hierarchical models effectively borrowed statistical strength across depth domains. The adaptive framework provides explicit decision criteria for method selection, quantifies data adequacy through statistical power analysis, and identifies optimal measurement campaigns for different repository design phases, from preliminary site characterization (minimum viable datasets) to detailed design (high-confidence predictions). This methodology addresses the fundamental question "how much data are enough?" by providing quantitative, site-specific guidance for stress characterization programs, ultimately enabling more efficient resource allocation and improved confidence for performance assessments in a nuclear waste repository site.