XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3 857
Simulation Environment Results
In a customized simulation environment replicating
industrial-scale plants, the Soft Actor-Critic (SAC) algo-
rithm showcased the most promising results, particularly
in terms of flexibility. One notable example from our trials
involved a dramatic shift in target product size from 35 µm
to 25 µm. The SAC algorithm responded swiftly, stabiliz-
ing new setpoints within 10 minutes, and aligning closely
with expectations for real-world industrial applications.
Within another 10 minutes, adjustments in feed rate and
mill fan speed were optimized. Notably, the recirculating
load increased from 55% to 120% during this change, con-
sistent with theoretical expectations and practical observa-
tions for achieving a finer product size. Moreover, the feed
rate to the mill was carefully modulated to keep the reject
rate below the circuit’s maximum capacity of 120 tph.
This efficient modulation of operational parameters by
the SAC algorithm emphasizes its capability to swiftly adapt
to new observations, leading to two significant benefits in
the operation of grinding plants: improved product quality
and enhanced energy efficiency. These benefits stem directly
from the reinforcement agent’s reward function, which, as
outlined in the methodology chapter, is designed to con-
stantly optimize these objectives. By efficiently adjusting
to changes in operational conditions, the SAC algorithm
ensures that product quality remains consistently high,
regardless of variations in the grinding process. This adapt-
ability is crucial for meeting stringent quality standards and
responding to dynamic market demands.
Additionally, the focus on maximizing energy efficiency
reflects a key aspect of sustainable mineral processing. The
algorithm’s ability to operate close to the circuit’s limits,
while maintaining optimal performance, demonstrates a
significant advancement in reducing energy consumption.
This not only aligns with the overarching goal of sustain-
able operations but also offers substantial cost savings and
environmental benefits.
These results, visualized in Figure 6, confirm the effi-
cacy of the reinforcement learning approach in enhanc-
ing grinding circuit operations. The SAC algorithm sets
a benchmark for future implementations and research in
the field, offering a data-driven pathway to higher product
quality and energy efficiency in mineral processing.
CONCLUSION AND FUTURE WORK
The research journey detailed in this research has been a
transformative exploration into the autonomous control
of dry grinding circuits in mineral processing, grounded in
data-driven machine learning techniques. The research has
successfully investigated complex technological paradigms
Figure 5. Dynamic Reward Curve and Regulation Quality for SAC Algorithm
Simulation Environment Results
In a customized simulation environment replicating
industrial-scale plants, the Soft Actor-Critic (SAC) algo-
rithm showcased the most promising results, particularly
in terms of flexibility. One notable example from our trials
involved a dramatic shift in target product size from 35 µm
to 25 µm. The SAC algorithm responded swiftly, stabiliz-
ing new setpoints within 10 minutes, and aligning closely
with expectations for real-world industrial applications.
Within another 10 minutes, adjustments in feed rate and
mill fan speed were optimized. Notably, the recirculating
load increased from 55% to 120% during this change, con-
sistent with theoretical expectations and practical observa-
tions for achieving a finer product size. Moreover, the feed
rate to the mill was carefully modulated to keep the reject
rate below the circuit’s maximum capacity of 120 tph.
This efficient modulation of operational parameters by
the SAC algorithm emphasizes its capability to swiftly adapt
to new observations, leading to two significant benefits in
the operation of grinding plants: improved product quality
and enhanced energy efficiency. These benefits stem directly
from the reinforcement agent’s reward function, which, as
outlined in the methodology chapter, is designed to con-
stantly optimize these objectives. By efficiently adjusting
to changes in operational conditions, the SAC algorithm
ensures that product quality remains consistently high,
regardless of variations in the grinding process. This adapt-
ability is crucial for meeting stringent quality standards and
responding to dynamic market demands.
Additionally, the focus on maximizing energy efficiency
reflects a key aspect of sustainable mineral processing. The
algorithm’s ability to operate close to the circuit’s limits,
while maintaining optimal performance, demonstrates a
significant advancement in reducing energy consumption.
This not only aligns with the overarching goal of sustain-
able operations but also offers substantial cost savings and
environmental benefits.
These results, visualized in Figure 6, confirm the effi-
cacy of the reinforcement learning approach in enhanc-
ing grinding circuit operations. The SAC algorithm sets
a benchmark for future implementations and research in
the field, offering a data-driven pathway to higher product
quality and energy efficiency in mineral processing.
CONCLUSION AND FUTURE WORK
The research journey detailed in this research has been a
transformative exploration into the autonomous control
of dry grinding circuits in mineral processing, grounded in
data-driven machine learning techniques. The research has
successfully investigated complex technological paradigms
Figure 5. Dynamic Reward Curve and Regulation Quality for SAC Algorithm