🎓 Lesson 17 D5

Geological Reasonableness Testing Workflow

Geological reasonableness testing checks whether an AI-generated orebody model makes sense given the known geology, rock types, and structural features of the site.

🎯 Learning Objectives

  • Explain how geological constraints (e.g., fault throws, bedding dip) limit permissible orebody shape deviations in AI models
  • Analyze AI-generated grade shells for violations of geological continuity using cross-section and plan-view validation techniques
  • Apply rule-based logic (e.g., 'no mineralization across impermeable shale barriers') to flag geologically implausible AI predictions
  • Design a geological reasonableness checklist aligned with local deposit type and regulatory reporting standards (e.g., JORC, NI 43-101)

📖 Why This Matters

AI models can generate statistically precise orebody models—but without geological guardrails, they may produce 'beautiful nonsense': grade shells that cut across faults, high-grade zones floating in barren host rock, or mineralized lenses violating known stratigraphy. In 2023, a major Australian copper project delayed feasibility by 5 months due to undetected geological implausibilities in AI-delineated resources—costing >$12M in rework. Geological reasonableness testing is your first line of defense against costly misinterpretation and non-compliant resource statements.

📘 Core Principles

Geological reasonableness rests on three pillars: (1) Structural fidelity—models must honor observed fault geometries, fold axes, and deformation timing; (2) Lithostratigraphic consistency—mineralization must occur within permissible host units (e.g., no porphyry Cu-Au in unaltered granite); and (3) Spatial plausibility—grade gradients, domain transitions, and contact relationships must reflect documented paragenesis and hydrothermal flow paths. Unlike statistical validation, this is *domain-constrained* reasoning: it asks 'Could this exist in nature?' before 'How well does it fit the data?'

📐 Geological Plausibility Index (GPI)

The GPI quantifies deviation from expected geological behavior using weighted rule violations. It supports rapid screening of AI models prior to full geological review.

Geological Plausibility Index (GPI)

GPI = Σ(w_i × v_i)

Quantitative score representing total weighted geological rule violations in an AI-generated model

Variables:
SymbolNameUnitDescription
w_i Weight of rule i unitless Assigned severity weight (1–5) reflecting confidence and consequence of rule violation
v_i Count of violations of rule i count Number of times rule i is breached in the model volume
Typical Ranges:
Porphyry Cu-Au deposit: 0–5
Orogenic gold system: 0–8

💡 Worked Example

Problem: An AI model predicts 0.8% Cu mineralization intersecting a mapped thrust fault with 120m vertical throw. The model places mineralization on both sides of the fault without offset compensation. Geological rules assign +3 points for each unadjusted fault intersection, +2 for grade discontinuity >0.3% across lithological contacts, and +1 for dip deviation >15° from regional bedding. Observed violations: 1 fault intersection, 2 litho-contact jumps, 0 dip errors.
1. Step 1: Identify violation counts per rule category: Fault = 1 × 3 = 3 pts; Litho-contact = 2 × 2 = 4 pts; Dip = 0 × 1 = 0 pts
2. Step 2: Sum weighted violations: GPI = 3 + 4 + 0 = 7
3. Step 3: Compare to threshold: GPI ≤ 3 = 'geologically reasonable'; 4–6 = 'review required'; ≥7 = 'reject and reprocess'
Answer: The result is GPI = 7, which falls within the 'reject and reprocess' range of ≥7.

🏗️ Real-World Application

At the Tasiast Gold Mine (Mauritania), a 2022 AI-grade control pilot used geological reasonableness testing to prevent overestimation of near-surface oxide material. Drill data showed a strong oxidation front dipping 8° NE—consistent with regional weathering. Early AI models generated steep, vertical oxidation boundaries. Applying a dip-consistency rule (max allowed deviation = ±3° from measured 8°) flagged 23% of blocks as implausible. Manual review confirmed all flagged blocks were erroneous extrapolations into unweathered basement. Revised models reduced grade variance by 31% and improved reconciliation accuracy from 89% to 96%.

📋 Case Connection

📋 Copper Mine Block Model Refinement Using Neural Kriging

Traditional kriging over-smoothed high-grade chalcocite zones, causing 8.2% reserve underestimation

📚 References