Key terms and concepts#

This page explains the main modelling, forecasting, uncertainty, and transfer concepts used throughout GeoPrior.

Brier score#: Mean-squared error for probabilistic forecasts of a binary event, such as whether subsidence exceeds a chosen hazard threshold. Lower values indicate better probabilistic skill.
calibration#: Agreement between stated predictive probabilities and observed frequencies. A calibrated forecast says “about 80%” and is correct about 80% of the time in repeated use.
closure consistency#: Property of a solution in which the learned consolidation timescale remains compatible with the timescale implied by the effective closure variables, especially \(K\), \(S_s\), and \(H_d\). In GeoPrior, this idea is central to making physics guidance testable rather than merely decorative.
coverage#: Fraction of observations that fall inside a stated prediction interval. For example, an 80% interval has good coverage when about 80% of held-out observations fall within it.
cross-city transfer#: Evaluation setting in which a model trained on one city is applied to another city, allowing the project to test generalisation under basin-scale domain shift.
data-dominated regime#: Regime in which the inferred timescale departs from the simple prior closure, yet the residual physics errors remain small. In practice, this suggests that the effective dynamics are being driven more by the data than by the prior timescale relation.
dynamic-balance error#: Diagnostic measuring mismatch between the forecast mean-path settlement rate and the reduced consolidation law. Lower values indicate tighter agreement with the adopted relaxation dynamics.
effective compressible thickness#: Censor-aware thickness quantity used by the physics pathway to represent the compressible part of the subsurface. It is built from mapped thickness with explicit handling of right-censored values, so it should be interpreted as an effective modelling input rather than a direct uncensored borehole-thickness estimate.
effective drainage thickness#: Hydro-geomechanical drainage-path length controlling the reduced consolidation timescale. In the current implementation it is not learned as a fully free field; it is derived from externally supplied thickness information through the model configuration.
exceedance probability#: Probability that subsidence exceeds a chosen threshold at a given location and time. This is the basis for threshold-risk maps and Brier-score evaluation.
forecast horizon#: Number of future years predicted jointly from one historical input window.
hotspot#: Location or spatial cluster identified as especially important for action because forecast severity, exceedance likelihood, and sometimes persistence jointly indicate elevated concern.
hotspot stability#: Measure of whether the same high-priority locations remain important under transfer, recalibration, or repeated forecast years. In practice this helps answer the policy question “where should we act first?”
identifiability#: Ability to recover or meaningfully separate effective model quantities from the available data and physics constraints. In GeoPrior, the effective timescale is usually more robustly recoverable than its full decomposition into \(K\), \(S_s\), and \(H_d\).
identifiability regime#: Named configuration profile used to reduce ridge-like non-uniqueness in the closure. Current profiles include base, anchored, closure_locked, and data_relaxed. These profiles adjust how strongly the model is anchored by priors, bounds, or closure controls.
physics-consistent forecast#: Forecast whose trajectories not only fit the data reasonably well, but also remain consistent with the adopted reduced groundwater and consolidation scaffold.
prediction interval#: Interval-valued forecast, such as a central 80% interval, intended to capture predictive uncertainty rather than provide only a single deterministic value.
sharpness#: Typical width of a predictive interval. For the same level of calibration and coverage, sharper forecasts are narrower and therefore more informative.
warm-start adaptation#: Transfer strategy in which a model trained in one city is used to initialise the target-city model and is then fine-tuned on local target data. In the manuscript, this performs much better than direct zero-shot transfer.
zero-shot transfer#: Direct application of a source-city model to a target city without local refitting. This is a stricter test of generalisation and is usually less reliable than warm-start adaptation under strong distribution shift.

How to read these terms#

A useful reading order for new users is:

Start with forecast horizon, prediction interval, calibration, coverage, and sharpness.
Then read closure consistency, dynamic-balance error, effective compressible thickness, and effective drainage thickness.
Finish with cross-city transfer, zero-shot transfer, warm-start adaptation, and hotspot stability.

Key terms and concepts#

How to read these terms#

See also#