Geological Domain Drift Score (GDDS)

GMI-MIN-006MiningInfrastructure-Dependent

Definition

Rate at which the AI geology model prediction accuracy degrades as mining progresses into geological domains underrepresented in the training data.

The Measurement Gap

Orebody models are trained on geological data from the domains that have been drilled, sampled, and characterized. As mining progresses, the active mining zone moves into new geological territory. The standard approach to model validation uses reconciliation: comparing predicted grades to actual grades after extraction. This is inherently backward-looking. By the time reconciliation reveals that the model is performing poorly in a new zone, the mine has already made dig-line decisions, truck dispatch assignments, and processing plans based on inaccurate predictions.

Legacy metrics like ore-to-waste reconciliation ratios, model bias factors, and cost per ton of ore all detect model failure after the fact. Physical Availability, MTBF, and PM Compliance are equipment metrics that have no connection to geological model validity. Even geostatistical cross-validation, the standard model testing technique, tests the model against held-out data from the training domain. It cannot detect that the model is being applied in a domain that looks different from what it learned.

Decision Supported:

Has mining progressed into a geological domain that the model was not trained for, and does the model need retraining or domain-specific recalibration?

Formula

Plain Language: Characterize the geological features of the training dataset (what lithologies were present, in what proportions, with what grade distributions and spatial variability patterns). Then characterize the same features in the zone the model is currently predicting. Compute the Jensen-Shannon divergence between these two distributions. A score of 0 means the current zone looks identical to the training data. A score approaching 1.0 means the current zone is geologically unrecognizable to the model.

Notation:

GDDS = D_JS(P_training || P_current)

Variable Definition Unit
D_JS Jensen-Shannon divergence between two distributions Dimensionless (0 to 1)
P_training Distribution of geological features in the model's training dataset Multivariate distribution
P_current Distribution of geological features in the current prediction zone Multivariate distribution
Feature dimensions Lithology type proportions, alteration intensity, structural domain classifications, grade distribution parameters, variogram parameters Mixed (proportions, grade units, spatial parameters)

A result below 0.15 means the model is operating in familiar geological territory. Above 0.3, the model is predicting in a domain that differs materially from its training data.

Proposed formula, pending validation pilots.

At a Glance

Unit
Score (0 to 1)
Trend Direction
Lower
Reporting Frequency
Quarterly
Category
Predictive Geology Model Integrity
← Operator Impact Attribution Accuracy (OIAA)Grade Prediction Calibration Index (GPCI) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment