AI for Mineral Processing & Beneficiation

Plant Feed Optimization

A beneficiation plant is designed for a specific feed envelope — grade range, moisture content, particle size distribution, mineralogy. Operate outside that envelope and recovery drops, reagent consumption spikes, and energy costs escalate. The challenge: run-of-mine ore is variable. Grade changes bench to bench, moisture swings with season, and mineralogy shifts across geological domains.

Open data/plant-feed-data.csv in the code panel. It contains hourly feed data — tonnes, Fe%, SiO2%, Al2O3%, moisture%, and particle size distribution (P80, % passing key screen sizes) — along with corresponding plant performance metrics (recovery, concentrate grade, tailings grade).

Grade-Tonnage Balancing

The mine delivers ore to the plant from multiple active faces, each with different grades. The blending challenge:

Objective: Minimize |feed_grade - target_grade|
Subject to:
  Σ tonnes_from_face[i] = plant_capacity
  tonnes_from_face[i] ≤ available_inventory[i]
  moisture_blend ≤ max_moisture (handling limit)
  P80_blend within crusher design range

At a large zinc-lead concentrator (Glencore, Teck, or Boliden style operation), feed grade can vary from 8% to 18% Zn depending on which stopes are active. An AI-based blending optimizer running every 4 hours — taking current stockpile inventories, active face grades from blast hole sampling, and plant metallurgical response curves as inputs — reduced feed grade variability by 45% and improved zinc recovery by 1.8 percentage points. On a 4 MTPA plant producing concentrate at $2,500/tonne, a 1.8% recovery improvement is worth roughly USD 15-18 million annually.

Moisture Control

Moisture is a silent plant killer. Wet-season ore at 8-12% moisture blinds screens, chokes chutes, and reduces crusher throughput. Dry-season ore at 2-3% generates excessive dust. AI-based moisture prediction — using weather data, stockpile residence time, and ore source — allows pre-emptive adjustments to screen aperture, spray bar settings, and feed rate.

Flotation Circuit Optimization

Flotation is where mineralogy, chemistry, and hydrodynamics collide. A flotation circuit may have 20-30 cells in rougher-cleaner-scavenger configuration, each with controllable air rate, froth depth, and reagent addition. The number of possible operating points is astronomical.

Open data/reagent-consumption.json — it contains reagent dosage data (collectors, frothers, depressants, pH modifiers) alongside flotation cell performance (recovery, grade, mass pull) for each circuit stage.

Recovery-Grade Tradeoff

The fundamental flotation tradeoff: higher recovery means lower concentrate grade, and vice versa. The relationship is described by the selectivity curve (recovery vs grade plot). The plant operates at a point on this curve determined by reagent dosages, air rates, froth depths, and cell levels.

AI learns the multi-dimensional response surface from historical data and finds operating points that push the selectivity curve outward — achieving better recovery at the same grade, or better grade at the same recovery.

Real-Time Flotation Adjustment

The key challenge is lag time. Reagent additions take 5-15 minutes to show their full effect in flotation performance. Feed grade changes propagate through the circuit over 30-60 minutes. Classical PID control, tuned for a single operating point, cannot handle these dynamics when feed characteristics change.

Model Predictive Control (MPC) with an AI-learned process model handles this well:

At each control interval (e.g., every 5 minutes):
  1. Measure current state: feed grade, cell froth characteristics (via camera),
     air rates, reagent flows, concentrate and tailings assays (online XRF/XRD)
  2. Predict future plant state over next 60 minutes using trained process model
  3. Optimize control actions (reagent dosages, air rates, froth depths)
     to maximize: recovery × concentrate_grade (or economic objective)
     subject to: concentrate_grade ≥ minimum_spec, reagent_cost ≤ budget
  4. Implement first control action, repeat

At a major base-metal concentrator (Freeport-McMoRan and Antofagasta operate copper flotation circuits of this scale), MPC-based flotation control improved copper recovery by around 2% compared to manual operator control, while reducing reagent consumption by 8%. Comparable gains are reported on zinc and lead circuits.

Process Control and Grinding Optimization

Open data/process-control-parameters.json — it contains grinding circuit data: mill power draw, feed rate, cyclone overflow P80, circulating load, and ball charge level.

Grinding Circuit Energy Optimization

Grinding (comminution) accounts for 40-60% of total processing plant energy consumption. A SAG mill drawing 10 MW costs roughly USD 6-9 million per year in electricity alone (depending on regional power price). Even a 5% reduction in specific energy consumption (kWh/tonne) is significant.

The grinding circuit control challenge: maintain target product size (typically P80 of 75-150 µm depending on downstream flotation requirements) while minimizing energy consumption per tonne processed. The variables:

AI-Enhanced PID Tuning

Traditional PID controllers for mill circuits are tuned at commissioning and rarely re-tuned. But the process dynamics change:

Adaptive PID tuning using reinforcement learning continuously adjusts controller gains to maintain optimal performance as process dynamics shift. The RL agent learns from thousands of hours of operational data and discovers tuning strategies that experienced operators use intuitively but cannot articulate — for example, temporarily reducing feed rate 15 minutes before a known hard ore zone reaches the mill, based on blast pattern and truck dispatch data.

Cascade Control

Grinding circuits benefit from cascade control architectures where an outer loop controls product quality (P80) and an inner loop controls the manipulated variable (feed rate or water addition). AI improves cascade control by:

Global Plant Context

Iron ore pelletizing (North America and Australia): Pelletization plants (Cleveland-Cliffs in the US Great Lakes region, Vale and Champion Iron operations) convert fine iron ore concentrate into pellets for blast furnace and DRI feed. AI-based control of the induration furnace — optimizing temperature profile, bed depth, and grate speed — has improved pellet quality (higher cold crushing strength, lower reduction degradation index) while reducing natural gas consumption by 6%.

Mid-tier and single-asset operators: Beyond the majors, many mid-tier and junior producers run plants with limited instrumentation and lean metallurgical staff. Simplified AI tools — running on a tablet connected to basic sensors (power meter on the mill, flow meter on water) — can deliver meaningful improvements. A random forest model predicting product quality from power draw and water flow, trained on 2 weeks of data with manual sampling, can substitute for continuous online analysis where capital for a full analyzer suite is not justified.

Key Takeaways