Metals and Mining
Industrial production has objective challenges
Uncertainty and fluctuations
Physical models describe only general dependencies
Expertise of operators is not easily scaled
Process control is limited
  • Reactive in its nature

  • Operation within allowed "ranges"

  • Static rules instead of dynamic adjustments
With AI, operational decisions become "personalized"
  • Automatic predictions and recommendations

  • Generated individually for each iteration
Use Case Examples
Cases previously done by Mechanica AI team
Optimization of ferroalloy use in oxygen-converter process
Steel slab quality prediction
Steel coil quality prediction
Prediction of efficiency decrease of an aluminum electrolytic cell
Prediction of efficiency decrease of an aluminum electrolytic cell
Example case: Optimization of ferroalloy use in oxygen-converter process
Client: large steelmaking company
5%
Average decrease in ferroalloy use
Problem
Solution
Data Used
How much of each ferroalloy should we add for this specific steel batch?
Sample model output
  • Client wants to keep ferroalloy use to a minimum while ensuring that steel complies with specification.

  • The decision on amount of alloys is not straightforward due to process complexity. Currently is made by an operator.
  • Build an ML-based recommender system that suggests optimal amounts of ferroalloys to the operator
Historical data on 200000 smeltings:
  • Logs of converter and ladle stage
  • Results of chemical analyses
  • Equipment data
  • Chemical composition requirements
  • Standards for ferroalloy use
Check out our case study:
Optimization of Ferroalloy Use
Example case: Steel slab quality prediction
Client: large steelmaking company
48%
of slabs prone to defect are identified among the top-10% of slabs ranked by the model
Problem
Solution
Data Used
How likely this specific slab will lead to defect during rolling?
Sample model output
  • Slab quality control has its own costs (cooling and re- heating of slabs)

  • Existing rule-based system has limited accuracy
    • Replace rule-based system with ML-based prediction based on available real-time measurements

    • Score each slab by their probability to have high defect mass

    • Choose the optimal threshold for slab re-routing
    Data on 17000 slabs:
    • Chemical composition
    • Geometrical parameters, weights
    • Casting logs
    • Known defects
    • Product quality requirements
    Example case: Steel coil quality prediction
    Client: large steelmaking company
    41%
    of coils prone to defect are identified among the top-11% of coils ranked by the model
    Problem
    Solution
    Data Used
    How likely this specific coil will have surface defect detected at later stages?
    Sample model output
    • Certain types of defects are not revealed at quality control after hot rolling (based on expert system)

    • This leads to irreversible losses at later stages
    • Label examples of "hidden" defects in past data

    • Build ML model that predicts the probability of such defects for each coil by type (e.g. surface, shape)

    • Use model output to change coil route or add additional processing steps to treat defect


    Data on 100000+ coils:
    • Slab data, chemical composition
    • Hot rolling processing logs
    • Known defects
    • Product quality requirements
    Example case: Prediction of efficiency decrease of an aluminum electrolytic cell
    Client: large aluminum company
    $10m
    Estimated yearly yield growth due to timely prevention of losses
    Problem
    Solution
    Data Used
    Which electrolytic cells are likely to have efficiency decrease in the next week?
    Sample model output
    • Certain electrolytic cells may have lowered daily tapping

    • The efficiency of the cell can be controlled manually, but this requires timely detection of possible problems

    • Existing indicators are insufficient
    • Predict which electrolytic cells will have lowered efficiency in the next days

    • Use this as an input to direct attention of a technician
    3 years of production data:
    • Operating logs
    • Data on raw material and automated feeding parameters
    • Data on the anode used (type, etc.)
    • Tapping data


    Example case: Optimization of the heat treatment process during steel pipe production
    Client: large pipe and tube company
    +5%
    Increase in shop throughput
    Problem
    Solution
    Data Used
    Which parameters (e.g. temperature and speed) we should set for a given batch?
    Sample model output
    • Producer wants to increase the shop throughput (processing speed) without affecting the quality

    • Choice of processing parameters for each batch is not straightforward due to quality fluctuations upstream
    • Predict expected mechanical properties of steel pipes

    • Estimate probability of defects

    • Recommend parameters to increase throughout without affecting quality
    Historical production data:
    • Steel grade, chemical analyses
    • Upstream processing logs
    • Heat treatment logs
    • Quality control results
    Want to learn more?