π Lesson 8
D5
Monte Carlo Dropout for Predictive Uncertainty
Monte Carlo Dropout is a way to estimate how confident an AI model is in its predictions by running the same input many times with parts of the model randomly turned off.
π― Learning Objectives
- β Explain how dropout at inference time approximates Bayesian inference
- β Apply Monte Carlo Dropout to quantify prediction uncertainty in grade estimation models
- β Analyze uncertainty outputs to inform orebody delineation decisions (e.g., drill spacing or stoping design)
- β Design a validation protocol to assess calibration of MCDO-derived uncertainty intervals
π Why This Matters
In orebody delineation, overconfident AI predictions can lead to costly misclassificationsβlike declaring low-grade material as economic ore, or rejecting high-grade zones due to noise. Traditional deep learning models output point estimates without uncertainty, hiding critical risk. Monte Carlo Dropout gives geotechnical engineers actionable confidence intervals for each predicted grade or boundary, enabling risk-aware resource modeling, dynamic grade control thresholds, and regulatory-compliant reserve statements per CIM Definition Standards.
π Core Principles
At training, dropout randomly deactivates neurons (typically with probability p = 0.2β0.5) to prevent co-adaptation and improve generalization. In MCDO, this same dropout is *retained during inference*: multiple forward passes (T = 30β100) are performed on the same input, each with a different random mask. The ensemble of outputs yields a mean prediction (ΞΌ) and standard deviation (Ο), where Ο reflects model uncertainty β high Ο indicates regions where the model lacks sufficient training data (e.g., near geological contacts or sparse assay coverage). Critically, MCDO approximates a Gaussian process posterior under certain assumptions, making it interpretable as epistemic uncertainty β unlike methods that only capture noise.
π Uncertainty Quantification via MCDO
For a trained neural network f(Β·; ΞΈ) with dropout rate p, MCDO computes T stochastic forward passes on input x to yield predictions {Ε·β, β¦, Ε·_T}. From these, we derive predictive mean and standard deviation β key metrics for grade control decision-making.
π‘ Worked Example
Problem: A convolutional neural network predicts copper grade (wt%) from borehole geophysics and adjacent assays. With T = 50 forward passes on a voxel at 420 m RL, outputs are: [0.48, 0.46, 0.51, ..., 0.47] (50 values).
1.
Step 1: Compute arithmetic mean: ΞΌ = (1/T) Ξ£ Ε·β
2.
Step 2: Compute sample standard deviation: Ο = β[Ξ£(Ε·β β ΞΌ)Β² / (Tβ1)]
3.
Step 3: Interpret: ΞΌ = 0.492 wt%, Ο = 0.021 wt% β 95% confidence interval β [0.451, 0.533] wt%
Answer:
The result is ΞΌ = 0.492 wt%, Ο = 0.021 wt%, which falls within the typical uncertainty band of <Β±0.03 wt% for well-calibrated MCDO in porphyry Cu systems.
ποΈ Real-World Application
At Newcrestβs Telfer Mine (Western Australia), MCDO was integrated into their AI-powered grade estimation pipeline (2022β2023 pilot). For a high-variability gold zone near a shear corridor, the model reported mean grade = 2.1 g/t Au with Ο = 0.42 g/t β triggering automatic recommendation for infill drilling before stope design. Conventional deterministic models showed 2.08 g/t with no uncertainty signal; subsequent drilling confirmed true grade ranged 1.6β2.8 g/t, validating the MCDO interval. This reduced misclassified tonnes by 18% in the first quarter of deployment, per internal Newcrest Orebody Modelling Review (2023).
π§ Interactive Calculator
π§ Open AI-Powered Orebody Delineation & Grade Control Calculatorπ Case Connection
π Copper Mine Block Model Refinement Using Neural Kriging
Traditional kriging over-smoothed high-grade chalcocite zones, causing 8.2% reserve underestimation
π Gold Mine Real-Time Grade Control at Development Drift Face
Manual chip sampling caused 24β48 hr delay in stope boundary decisions, leading to 14% dilution
π Limestone Mine Digital Twin for Karst-Related Grade Uncertainty
Solution cavities caused unpredictable grade drops (CaCOβ purity <85%) in otherwise uniform deposits, resulting in 11% p...