AI Grade Prediction: Cut Ore Misclassification by 30%

Ore misclassification—sending rich ore to waste or barren rock to the mill—costs millions of dollars annually. Our AI grade prediction ML models cut misclassification by at least 30% in typical applications. We encountered this at a gold deposit in Kazakhstan: ordinary kriging gave 35% misclassification. After implementing an XGBoost ML model with spatial features, misclassification dropped to 13%, saving $2 million per year. The problem with kriging: it requires stationarity and ignores lithology, geophysics, and geochemistry. ML (gradient boosting, random forest, 3D convolutional networks) builds non-linear relationships and uses all available data.

AI Grade Prediction: How It Reduces Misclassification

Kriging provides a BLUE estimate under stationarity. In real ore bodies, this condition is violated: faults, oxidation zones, vein textures. ML models (gradient boosting, random forest, 3D CNN ore) capture non-linear dependencies and leverage all data. XGBoost mining with spatial features gives RMSE 1.75 times lower than ordinary kriging—a 40% accuracy improvement. ML is 1.75 times better than kriging in RMSE.

drill_data = {
    'x', 'y', 'z',           # coordinates
    'au_g_t',                 # target
    'density',
    'lithology_code',
    'alteration_type',
    'magnetic_susceptibility',
    'ip_chargeability',
    'distance_to_fault'
}

Random Forest with spatial features

from sklearn.ensemble import GradientBoostingRegressor
import numpy as np

def spatial_features(x, y, z, drill_holes):
    distances = np.sqrt((drill_holes['x'] - x)**2 +
                        (drill_holes['y'] - y)**2 +
                        (drill_holes['z'] - z)**2)
    nearest = drill_holes.nsmallest(10, key=lambda _: distances)
    return {
        'mean_grade_r50': drill_holes[distances < 50]['grade'].mean(),
        'max_grade_r100': drill_holes[distances < 100]['grade'].max(),
        'nearest_grade': nearest.iloc[0]['grade'],
        'grade_gradient': (nearest.iloc[0]['grade'] - nearest.iloc[5]['grade']) / distances.nsmallest(5).mean()
    }

XGBoost with geological domains: We build separate models for each lithotype × alteration zone. The result is more accurate than a single global model. 3D CNN ore models on a voxel grid yield up to 25% accuracy gain over kriging on complex deposits.

AI Grade Prediction: Ensuring Forecast Reliability

Standard k-fold CV overestimates due to spatial correlation. We use spatial block CV—blocks split by geographic quadrants, with Leave-One-Block-Out. On a copper deposit in Chile, this spatial validation showed that an ensemble of XGBoost and 3D CNN reduces misclassification from 28% to 11%.

Forecast quality metrics (non-Q&A style): RMSE (grade deviation in g/t or %), Slope of regression (predicted vs. actual, ideal 1.0), E-Type variance (conditional variance of estimate), Reconciliation factor (forecast vs. production over periods).

How is the block model built?

For each 5×5×5 m block we predict Au g/t. The cutoff threshold is optimized via ROC analysis with economic weights:

cutoff_grade = 0.3  # g/t
for block in mining_blocks:
    predicted_grade = model.predict(block.features)
    block.destination = 'mill' if predicted_grade >= cutoff_grade else 'waste'

What's included in the work?

• ML model with documentation and validation metrics.
• Integration of block model ore into existing system (Datamine, Vulcan, Leapfrog).
• Pipeline for automatic grade reconciliation and retraining.
• Training geologists and technologists on model usage.
• Technical support during operation (3–6 months).
• Custom dashboard for real-time monitoring of forecast vs actual grades.

Work process

1. Data audit: collection, cleaning, building a feature store.
2. Feature engineering: spatial and geological features.
3. Training and validation: XGBoost, Random Forest, 3D CNN with spatial block CV.
4. Integration: block model + routing system.
5. Follow-up: reconciliation, model adjustment based on actuals.

Typical mistakes and how to avoid them

• Ignoring spatial correlation → we always use spatial validation.
• Overfitting on geophysical data → we apply regularization and stratification by domains.
• Averaging across all lithotypes → we build separate models for each geological domain.
• Using only a single model → we ensemble multiple algorithms for robustness.

Integration with mining software

Predicted grade estimates are fed into standard block models. Supported formats: Datamine CSV, Vulcan BMTF, Leapfrog CSV/DXF, Micromine CSV. For real-time applications we connect XRF analyzers on the conveyor (OPC UA and Modbus protocols). The reconciliation pipeline compares block forecasts with actual chemical assays after extraction and automatically adjusts model bias monthly. Change logs are kept in MLflow. This approach allows the system to self-calibrate as new lab data arrives without stopping production.

Timeline and cost

Basic model (geostatistics neural networks + XGBoost + block routing) — 5–7 weeks, from $30,000. Full system with 3D geophysics, real-time XRF, and reconciliation pipeline — 3–4 months, from $100,000. Cost is estimated individually after data audit. Annual savings from reduced misclassification typically range from $500,000 to $5,000,000 per deposit.

Our AI grade prediction ML models use spatial validation and block modeling to reduce misclassification of ore, leveraging XGBoost mining and 3D CNN ore techniques. Contact us for a preliminary data audit—we’ll assess the potential to reduce misclassification on your deposit and propose a specific architecture. Our engineers have certified experience in MLOps and geostatistics for over 5 years. We have completed more than 20 projects on predictive ore grade forecasting AI at gold, copper, and iron deposits, achieving production optimization ML savings of over $2 million per project.

When does a time series forecasting model fail in production?

The CFO requests a quarterly sales forecast. An analyst builds SARIMA on three years of data, achieves MAPE 8.3% on the test set, and deploys. Two months later, the metric in production jumps to 23%. The root cause: the model was trained on pre‑COVID data, tested on a stable period, but production hit a promotion and supply chain disruption. Data leakage plus distribution shift—perfect notebook numbers, a broken forecast in reality. We have seen this pattern dozens of times across retail, fintech, and IoT. Our team has delivered more than 50 forecasting projects over 5+ years.

Incorrect cross-validation. Standard train_test_split for time series creates data leakage: the model sees future values during training. The correct approach is TimeSeriesSplit or walk‑forward validation with an expanding window.

Multiple seasonality. Hourly electricity consumption has three seasonalities: daily (24h), weekly (168h), yearly (8760h). SARIMA handles only one. Prophet can handle multiple but scales poorly to thousands of series.

Missing values and anomalies. A missing sensor reading is information (the sensor turned off), not NaN. Linear interpolation destroys this signal. Proper handling depends on the missingness mechanism.

Cold start. A new SKU in a 50,000‑item assortment has no history, yet a forecast is needed. Standard approaches fail; cross‑learning or feature‑based methods are required.

Why is model selection critical for your data?

Prophet (Meta) – a solid start for business data with clear seasonality and holidays. Fast setup, interpretable, built‑in outlier detection. Fails on irregular patterns and does not scale beyond ~10k series without parallelization.

Gradient boosting on features (LightGBM, XGBoost) – often underestimated. Engineer lags (t‑1, t‑7, t‑28), rolling means, day‑of‑week, holidays. The model trains on all series simultaneously, solving cold start via transfer learning. MAPE in retail often beats neural nets with proper feature engineering.

TFT (Temporal Fusion Transformer) – a transformer designed for interpretable forecasting with covariates. Built‑in variable selection, temporal attention, quantile outputs. Available in pytorch‑forecasting. Requires ~10,000+ records per series for stable training.

PatchTST – splits the series into patches (like ViT for images), capturing local patterns better than classic transformers. Excellent for long‑horizon forecasting (96–720 steps ahead).

N‑HiTS, N‑BEATS – attention‑free neural architectures, faster than TFT, competitive accuracy. N‑BEATS won the M4/M5 benchmarks for tasks without covariates.

How do we deploy TFT in production?

A typical pipeline via pytorch‑forecasting:

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="sales",
    group_ids=["store", "sku"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,  # 120 days
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,  # 28 days
    static_categoricals=["store_type", "category"],
    time_varying_known_reals=["price", "promo_flag"],
    time_varying_unknown_reals=["sales"],
    target_normalizer=GroupNormalizer(groups=["store", "sku"], transformation="softplus"),
)

A common mistake: the default target_normalizer (StandardScaler) breaks predictions for series with zero values (no sales on weekends). GroupNormalizer with transformation="softplus" is the correct choice for count data.

Case study: retail demand forecasting

A chain of 120 stores, 8,000 SKUs, 28‑day forecast horizon. The original system: SARIMA per series, MAPE 18.4%, retraining cycle – 6 hours. We replaced it with TFT on PyTorch + pytorch‑forecasting: a single model for all series, MAPE 11.2%, retraining – 40 minutes on an A10G. Feature importance via variable selection revealed that day_before_holiday influences more than the holiday date itself. Annual savings on inference alone exceeded $50,000.

Step‑by‑step configuration

1. Data collection and preparation. Handle missing values (mark NaN, interpolate only for technical failures), aggregate to required frequency, engineer covariates (holidays, promotions, prices).
2. Create TimeSeriesDataSet. Set group_ids (store + SKU), time index, forecast horizon. Choose target_normalizer based on target distribution.
3. Train a baseline. Prophet or LightGBM first – to understand complexity.
4. Train TFT. Use TemporalFusionTransformer with loss=QuantileLoss(), tune learning rate and hidden layer sizes.
5. Validate and interpret. Walk‑forward test, analyze variable selection, build attention heatmaps.

How to properly evaluate forecast quality?

RMSE alone is misleading – it over‑penalizes large values. Our standard set:

• MAPE – interpretable, unstable near zero.
• sMAPE – symmetric, avoids division by small numbers.
• MASE (Mean Absolute Scaled Error) – normalized relative to a naive seasonal forecast, ideal for comparing series of different scales.
• Pinball loss – for probabilistic forecasting, inventory management.

We guarantee a model card with these metrics on the validation set and walk‑forward results on at least 6 months of history.

What deliverables do you receive?

• Documentation of chosen architecture and hyperparameter rationale.
• Reproducible training and inference pipeline (Docker + CI/CD + Airflow/Prefect).
• Committed code with unit tests for key components.
• Team training: retraining, output interpretation, deployment of new versions.
• 3 months of post‑delivery support (consultations, bug fixes, fine‑tuning).

The model is deployed via FastAPI or Triton Inference Server. Retraining is scheduled (e.g., weekly) via Airflow with drift validation and automatic rollback if metrics deteriorate.

Process and timeline

We start with EDA: visualization, ADF test, STL decomposition, analysis of missing values and outliers. This takes 2–3 days but often reveals systemic data issues that block forecasting. Then we build a baseline (naive seasonal, Prophet), engineer features for LightGBM, and select a neural architecture if needed. Walk‑forward validation with a realistic horizon. Deployment via API with automatic retraining scheduled via Airflow or Prefect.

Timeline: MVP forecast on one data type – 3–6 weeks. Hierarchical forecasting system with automation – 2–5 months. Cost is calculated individually based on data volume, number of series, and required accuracy.

Our team consists of certified ML engineers (AWS ML Specialty, GCP Professional ML Engineer) with 5+ years on the market and over 50 completed forecasting projects. Contact us for a free analysis of your data – we will assess the task and provide initial recommendations within 1–2 days. Request a consultation to ensure your forecasts work in production, not just in a notebook.