Accurate characterisation of in situ stress plays a vital role in supporting underground excavation assessment in mining engineering, but data-driven stress estimates can become heavily biased by target-derived information and spatially unrepresentative validation practices. To address these vulnerabilities, a research team comprising Yi Li, Zhanshan Shi, Gang Li, and Beifang Wang evaluated a physics-consistent workflow designed to rigorously separate blind prediction, conditional stress completion, and consistency screening across complex geological environments.
The comprehensive investigation evaluated empirical data from 725 hydraulic-fracturing records assigned across 33 coalfield-scale database groups and 36 representative positions. To examine how well predictive algorithms generalise across space, the authors held out complete coalfield and coordinate groups while applying repeated validation, cluster bootstrap uncertainty estimation, feature ablations, and limited nested parameter selection to test model robustness under realistic spatial constraints.
The evaluation demonstrated that the introduction of target-derived inputs causes severe information leakage, leading to artificially elevated performance metrics. Specifically, target-derived inputs increased the coalfield-held-out determination coefficient for a LightGBM model from 0.383 to 0.920 when assessing maximum horizontal stress. This shift represents an absolute increase of 0.537, illustrating how unmonitored feature inclusion can fundamentally distort perceived machine learning reliability.
Baseline comparisons revealed that algorithm complexity did not guarantee superior predictive skill across the stress data. A full-feature Ridge regression model attained determination coefficients of only 0.440 and 0.425 for the two principal horizontal stresses. In contrast, simpler models relying exclusively on depth features reached determination coefficients of 0.526 and 0.521, outperforming the full-feature linear configurations.
The authors reported that relative model rankings depended heavily on group weighting schemes, validation partitions, and feature ablations. Although blind geographic transfer remained limited, the workflow demonstrated that conditional stress completion remained consistently accurate under grouped constraints, successfully retaining grouped determination coefficients ranging from 0.867 to 0.904 across the examined datasets.
Statistical uncertainty analysis exposed notable trade-offs between interval coverage and the precision of estimated stress bounds. Applying group-block calibration increased record-level interval coverage to between 97.9% and 99.3%. However, achieving this high degree of empirical coverage required substantially widening the prediction intervals, indicating that nominal uncertainty measures can mask broad error ranges.
Spatial transfer experiments conducted within specific mining basins highlighted clear regional discrepancies in model portability. In the Lu’an coalfield, conditional completion achieved determination coefficients near 0.93 under both mine-group and representative-coordinate hold-out evaluations, whereas blind transfer to unseen locations remained severely limited across testing setups.
Furthermore, combining local records with data from other coalfields improved hidden-point stress predictions in the Lu’an and Xishan coalfields, yet it failed to yield consistent improvements in Jincheng. The findings ultimately support establishing explicit information boundaries and conducting group-specific evaluations during stress-database appraisal, while advising caution regarding claims of geographic transferability and engineering substitution.
Subject of Research: Evaluation of information leakage, spatial generalisation, and uncertainty in data-driven in situ stress estimation for coal mining
Article Title: Information leakage, spatial generalisation and uncertainty in data-driven in situ stress assessment for coal mines
Article References: Li, Y., Shi, Z., Li, G., & Wang, B. (2026). Information leakage, spatial generalisation and uncertainty in data-driven in situ stress assessment for coal mines. Geomechanics and Geophysics for Geo-Energy and Geo-Resources. https://doi.org/10.1007/s40948-026-01275-8
Image Credits: AI Generated
DOI: 10.1007/s40948-026-01275-8
Keywords: in situ stress, coal mines, hydraulic fracturing, information leakage, spatial generalisation, data-driven estimation, LightGBM, Ridge regression, maximum horizontal stress, conditional stress completion, consistency screening, cluster bootstrap uncertainty, group-block calibration, blind prediction, Lu'an coalfield, Xishan coalfield, Jincheng coalfield, geomechanics
News Source: Blake Davidson. (November 16, 2026). Target-Derived Inputs Artificially Boost Data-Driven Stress Predictions in Coal Mines. Scienmag.



