A Next Step in Industrial Digitalization
Dan Smith, PhD
Business Director
BTG Group
The Digital Journey – Driving Performance
Efficiency
Operations
autopilot
optimize
Data
Measurement operationalize
explore
Advanced process controls /
visualize
Advanced process optimization
collect, store &
transform
measure,
generate &
Analytics, visualization and operations management
control
Instruments Data historian
Automation & platform
2
A Brief History of Industrial AI – What Didn’t Work
1980 – Expert Systems – “Solving All Problems”
Software that mimics human expertise using a knowledge base and
inference engine to solve domain-specific problems.
Example: Recovery boiler advisor
Adoption: 65% of Fortune 500 companies used expert systems in the 1980s
for diagnosis, configuration, and decision support
Decline in the 2000s:
• Knowledge acquisition issues
• Scalability limits
• Verification challenges
• Overhyped expectations
1990 – Neural Networks – “Solving All Problems”
Neural networks are brain-inspired machine learning models using
interconnected nodes to learn patterns and make decisions from data
Example: Tensile Strength Model
Adoption: Significant pulp and paper market penetration, many
Proof-of-Concept examples
Decline in the 2000s
• "Black Box" Problem
• Ignoring the Downsides
• Real-World Problems
• Hard to Use
What Did Work?
1980 – Model Predictive Control – Solving Problems
Model Predictive Control (MPC) is an advanced control technique that uses a
system model to predict the process response and optimize control actions
Example: Bleach Plant Advanced Control
Adoption: Invented 1982; 5000+ applications in 2000; 1,000,000+ in 2025
Uses: Process industries, automotive, aerospace, power, robotics, building
automation,…
Continuous Growth
• Better Technology
• Energy and Cost Savings
• Increasing process complexity
1980 – Model Predictive Control – Solving Problems
• The process is represented by a
collection of dynamic response
models
• The model matrix can have
hundreds or more sub-models
• Well established implementation
methodology
1980 – Model Predictive Control – Solving Problems
S N Andriyashin et al 2021 J. Phys.: Conf. Ser. 2032 012065
Bleach Plant Savings with MPC
• Chemical use optimized and final brightness variability
significantly reduced
• Lower variability enables target shift and additional
chemical savings
• << 12-month ROI from bleach chemical savings
10
Proven Applications in Pulp & Paper
ODL
W ashing
Bleaching
Cooking
Wood
Chipping Tissue Machines
O2
Raw WL ClO2
Material
(fiber) Pulp
BL
Recaust
Lime Kiln GL
Bark Evaporator
Electricity
End Product
Turbine Generator
Recovery
Clean Energy
Boiler
Power Boiler
Steam
Path to Success – Why MPC works
• Built on base of instrumentation, data platforms, and technical capability
• Well established methodology
• Realistic understanding of required engineering effort
• Projects focused on ROI
Efficiency
Operations
autopilot
optimize
Data
Measurement operationalize
explore Advanced process controls /
visualize Advanced process optimization
collect, store
measure, Analytics, visualization and operations
& transform
generate & management
control Data
historian &
Instruments platform
Automation
12
Making MPC Better with AI
• MPC algorithm “understands” how the process responds based on an
embedded model
but….
• Algorithm does not include a methodology to “learn”
• Modern AI techniques hold promise to add dynamic “learning”
What is Reinforcement Learning?
Learning by Trial and Error
• A type of Machine Learning (ML) where an agent learns to make
decisions in an environment
• The goal is to maximize cumulative reward over time
• Learns through interaction and feedback
• Think of it like training a dog: reward good behavior, discourage
bad behavior
The Reinforcement Learning Loop
Interaction and Learning
Agent selects Action Action is Executed
MPC calculates control actions 2 MPC sends remote setpoints
3
Agent observes State 1
Transition to New State
MPC reads process data 4
Process responds to control actions
6 5
Agent updates Policy Agent receives Reward
AI adjusts MPC tuning MPC objective improves or degrades
MPC - Reinforcement Learning (RL) Agent
Reward: RL can be used to optimize the MPC. The reward signal reflects how well
the MPC is performing according to a higher-level objective. This could include:
• Tracking a desired setpoint
• Minimizing energy consumption
• Maintaining stability
• Optimizing throughput
Challenges
It’s Not Always Easy
• Exploration vs. Exploitation: Finding the
right balance between trying new things and
sticking with what works.
• Reward Shaping: Designing a good reward
function is crucial but difficult. A bad reward
function can lead to unintended
consequences.
• Generalization: Learning to perform well in
unseen situations.
• Sample Efficiency: RL algorithms can
require a lot of data to learn effectively.
Conclusion – Take the “Next” Step on the Digital Journey
• Many “next steps” on the Digital Journey
• Digital technology available today can generate excellent ROI
• R&D underway to significantly extend the journey
Efficiency
Operations
autopilot
optimize
Data
Measurement operationalize
explore Advanced process controls /
visualize Advanced process optimization
collect, store
measure, Analytics, visualization and operations
& transform
generate & management
control Data
historian &
Instruments platform
Automation
18
