50 high-dimensional ensemble of simulated pressure responses provides the basis for the planned global sensitivity analysis. The current ensemble of one hundred realisations is intended as a preliminary stage to validate the workflow and confirm computational feasibility; the formal analysis will scale to a larger ensemble (on the order of one thousand realisations) consistent with established practice in subsurface uncertainty quantification studies (Fang et al., 2022). 3.4 Distance-based generalised sensitivity analysis The next stage of the workflow will apply distance-based generalised sensitivity analysis (DGSA; Park et al., 2016) to the simulated ensemble to identify which uncertain parameters most strongly control predicted pore pressure evolution. DGSA partitions parameter samples by clustering the corresponding response variables and computes the distance between conditional and prior cumulative distributions for each parameter. The method accommodates uniform, scenario-based, and high-dimensional response variables that conventional variance-based sensitivity methods handle poorly. Sensitivity values above unity will indicate that the corresponding parameter exerts a statistically significant influence on the response at the 95 percent confidence level. The output will be a ranked list of parameters identifying the dominant controls on predicted pressure evolution, which will both inform priorities for further site characterisation and identify which assumptions in the H1 model are most consequential for the planned falsification test. 3.5 Planned falsification test Once the sensitivity analysis is complete, the simulated pressure ensemble will be compared against observed piezometric data from the Mingomba site, following the falsification logic of the Popper-Bayes protocol introduced in Section 3.1. The reasoning is straightforward: if the observed data could plausibly have been produced by the model under the prior assumptions, the prior is consistent with reality and Bayesian inversion can proceed; if the observed data lies outside the range of responses the model can generate, the prior has been proven incorrect, and the modelling assumptions must be revised before any further inference is meaningful. Operationally, the test will compare the position of the observed data relative to the cloud of simulated responses. Because the simulated ensemble is high-dimensional, dimensionality reduction is required. Principal component analysis will be applied jointly to the simulated and observed data, projecting both onto the directions of greatest variability in the simulated ensemble. In this reduced space, the observed data is represented as a single point and the simulated realisations as a cloud of points; their relative position is then measurable. A statistical outlier test will be applied to determine whether the observed point lies within or outside the simulated cloud, using a distance metric (the robust Mahalanobis distance) that accounts for the shape and spread of the cloud rather than treating distance as a simple Euclidean measure. A distance value above the 95 percent confidence threshold will indicate that the observed data is statistically inconsistent with the simulated ensemble and that the H1 prior is therefore falsified.
RkJQdWJsaXNoZXIy MTM0Mzk2