870
Optimization of a Grinding Circuit Using Physics-Based and
Machine Learning Models
Utkarsh Sinha, Mohammed Suhail, Vishnu Swarupji Masampally,
Nagaravi Kumar Varma Nadimpalli, Venkataramana Runkana
TCS Research, Tata Research Development and Design Centre, Tata Consultancy Services, Pune, India
ABSTRACT: Mining and mineral processing is an energy intensive industry and optimization of key unit
operations and processes is crucial for achieving profitability. The comminution operation consumes around
40% of the total energy demand of a typical plant. Minimization of power consumption while improving the
throughput and product quality is an important and challenging task. Moreover, the degrading quality of ores is
lowering the yield of minerals per unit mass of ore. As a result, mineral extraction requires smaller particle sizes
and ultrafine grinding may be necessary. The existing industrial plants cannot transition easily to new technologies
for achieving ultrafine grinding due to high capital investment and operational costs. An alternative approach
is to predict grinding circuit key performance indicators (KPIs) accurately and optimize them in real-time.
Existing models consider either product quality or power consumption and not ball mill hold-up (H), which
is a key process variable. A novel framework for model-based optimization of a grinding circuit performance
that combines physics-based or semi-empirical models of the three KPIs – ball mill holdup (H), particle size
distribution and power consumption is presented. The models for each KPI are tested with experimental data
of a lab-scale ball mill. Machine learning models are then developed from simulated data generated using the
physics-based models for use as soft-sensors, required for online optimization. Multi-objective optimization
is carried out to determine the optimal settings to maximize circuit throughput and minimize specific power
consumption while meeting product quality specifications. Although this framework is tested for grinding of
ores, this is applicable for grinding operations in other industries such as cement, pharmaceuticals, chemicals,
and power utilities.
Keywords: Comminution, Mineral Processing, Ball mill, Modeling, Optimization, Soft sensors, Machine learning
INTRODUCTION
Grinding is an important operation in several industries
such as minerals and metals, chemicals, cement, ceramics,
pharmaceuticals, food processing and power utilities. A sig-
nificant advancement in size reduction technology has been
driven by the mineral processing industry. Minerals occur
naturally as ores and exhibit a varying degree of physical
and chemical affinity to their impurities (Zhang et al.,
2020). Grinding is a mechanical operation of size reduction
of crushed ores and is the first step in extraction of miner-
als, followed by flotation, hydrometallurgy, pyrometallurgy,
purification and so on (de Carvalho et al., 2021). Grinding
is the most energy intensive stage in mineral processing,
contributing to around 40% of the total energy demand
Optimization of a Grinding Circuit Using Physics-Based and
Machine Learning Models
Utkarsh Sinha, Mohammed Suhail, Vishnu Swarupji Masampally,
Nagaravi Kumar Varma Nadimpalli, Venkataramana Runkana
TCS Research, Tata Research Development and Design Centre, Tata Consultancy Services, Pune, India
ABSTRACT: Mining and mineral processing is an energy intensive industry and optimization of key unit
operations and processes is crucial for achieving profitability. The comminution operation consumes around
40% of the total energy demand of a typical plant. Minimization of power consumption while improving the
throughput and product quality is an important and challenging task. Moreover, the degrading quality of ores is
lowering the yield of minerals per unit mass of ore. As a result, mineral extraction requires smaller particle sizes
and ultrafine grinding may be necessary. The existing industrial plants cannot transition easily to new technologies
for achieving ultrafine grinding due to high capital investment and operational costs. An alternative approach
is to predict grinding circuit key performance indicators (KPIs) accurately and optimize them in real-time.
Existing models consider either product quality or power consumption and not ball mill hold-up (H), which
is a key process variable. A novel framework for model-based optimization of a grinding circuit performance
that combines physics-based or semi-empirical models of the three KPIs – ball mill holdup (H), particle size
distribution and power consumption is presented. The models for each KPI are tested with experimental data
of a lab-scale ball mill. Machine learning models are then developed from simulated data generated using the
physics-based models for use as soft-sensors, required for online optimization. Multi-objective optimization
is carried out to determine the optimal settings to maximize circuit throughput and minimize specific power
consumption while meeting product quality specifications. Although this framework is tested for grinding of
ores, this is applicable for grinding operations in other industries such as cement, pharmaceuticals, chemicals,
and power utilities.
Keywords: Comminution, Mineral Processing, Ball mill, Modeling, Optimization, Soft sensors, Machine learning
INTRODUCTION
Grinding is an important operation in several industries
such as minerals and metals, chemicals, cement, ceramics,
pharmaceuticals, food processing and power utilities. A sig-
nificant advancement in size reduction technology has been
driven by the mineral processing industry. Minerals occur
naturally as ores and exhibit a varying degree of physical
and chemical affinity to their impurities (Zhang et al.,
2020). Grinding is a mechanical operation of size reduction
of crushed ores and is the first step in extraction of miner-
als, followed by flotation, hydrometallurgy, pyrometallurgy,
purification and so on (de Carvalho et al., 2021). Grinding
is the most energy intensive stage in mineral processing,
contributing to around 40% of the total energy demand