ISA FPID Symposium Advanced Program Opportunities for PAT and Advanced Process Controls for Optimization in Continuous Manufacturing Paul Brodbeck/Control Associates Inc. Emerson Local Business Partner
Jul 18, 2015
ISA FPID Symposium Advanced Program
Opportunities for PAT and Advanced Process Controls for Optimization in Continuous Manufacturing
Paul Brodbeck/Control Associates Inc. Emerson Local Business Partner
Benefits of Continuous Manufacturing
State of Continuous Manufacturing
Capital Growth Continuous Manufacturing
Advanced Control Opportunities
1. Model Predictive Control - MPC
2. Neural Networks - NN
3. Linear Programming Optimization - LP
4. Kalman Filter - KF
Agenda
Improved Product Quality • Better Quality Control• Meaningful PAT In-Process Control • Real Time Release• Prevents Segregation and Agglomeration• Reduced Off-Spec Batches & Materials
Cost Reduction• Smaller Equipment • No Scale Up• Faster Development• Less API used in Development• Lower Cost of Quality Assurance• Reduced Processing Time,
Continuous Manufacturing
Less than 20% Time
To produce a product order Continuously versus Batch
2 Days CM versus 2 weeks Batch
Less than 20% Wasted Material
During Development
100 kgs CM versus 5000 kgs Batch
Less than 20% Installed Costs
$20 mil CM versus $200 mil Batch
CM Benefits
“At a factory in Puerto Rico, J&J has built a line that could manufacture the HIV/AIDS medicine Prezista starting in 2016 using the new techniques if regulators approve. The main ingredients will still be made elsewhere, but J&J aims to manufacture 70% of its highest-volume products” using the processes within eight years”, Mr. McKenzie said.
“As a result of such benefits, companies will save an estimated 30% or more in operating costs, according to Bernhardt Trout, director of the Novartis-MIT Center for Continuous Manufacturing, which has been developing the new technologies with funding from Novartis.” – WSJ
Hayden Thomas, a Vertex manufacturing official. If the company’s new cystic-fibrosis drug gets approved, the facility would make 100,000 tablets in an hour, rather than the four to six weeks that would be needed to make a batch at a traditional plant.
WSJ Article – Feb. 8, 2015
Novartis-MIT Center for Continuous Manufacturing $65MM
GlaxoSmithKline Continuous Plant in Singapore $29MM
PfizerCollaboration with GEA & G-CON
Patheon – CMO/CDO Eli Lilly “By early next year, Eli Lilly and Company will have installed
and demonstrated four different continuous-processing platforms. Currently, almost all of our potential medicines that are in development have continuous-processing steps in place.” B. Huff - Executive Director of Chemical Development September 2013
CM – Capital Growth
1. Model Predictive Control - MPCRefineries, Robotics, Drones, Aerospace…
2. Linear Programming - LP OptimizationOperations Research, Economics, Scheduling…
3. Neural Networks - NNPattern Recognition, Stock Market, Genetics…
4. Kalman Filter - KFRobots, Aerospace, Missile Guidance…
APC ‘State-of-Art’ Other Industries
1.Model Predictive Control
Feed Forward to Tablet Press
API Concentration Feedback
2. LP Optimization
Feedrate Optimization
3. Neural Net
Soft Sensor
4. Kalman Filter
PAT Signals i.e. BioReactors
8
CM- Opportunities
Model Based Closed-Loop Control
Multivariable Inputs and Outputs
Decouples interacting loops
Good with difficult dynamics
Deadtime compensated automatically
Feed Forward Implicit
Optimal Constraint Control
Robust and Proven in Industry
Predictive – Not waiting for an error.Anticipates the best strategy
Planning Ahead – like Humans
1. Model Predictive Control (MPC)
Model Predictive Control (MPC)
22 2
1 1 1 1 1 1
1 1y u u
n n nP M My set u u
j j j j j j j j
i j i j i j
J w y k i y k i w u k i w u k i u
y: Controlled variableu: Actuator△u: Predicted adjustment
ManipulatedVariable
Controlled Variable Deviations (SP-PV)
Controller Adjustments(Output Change)
Singh, R., Ierapetritou, M., Ramachandran, R. (2013). European Journal of Pharmaceutics and Biopharmaceutics, http://dx.doi.org/10.1016/j.ejpb.2013.02.019.
Tuning parameters1. Output weights (wy
j) 2. Rate weights ( ) 3. Input weight ( ) 4. Prediction horizon5. Control horizon
u
jw
u
jw
The Objective Function J
minimizes the Error &
Changes for the sum of
all three internal
functions.
MPC BioReactor Perfusion
Multi-Loop PID
Multi-Loop Control
Multi-Loop MPC
BioReactorTemperature, pH, DOGlucose, Nutrient…
Feed Back Closed Loop API %
Feed Forward Density to Tablet Press
2. LP Optimization
Linear Programming
Discrete Optimization
A mathematical/computer optimization technique – Simplex Method
Solve a system of linear equations
Can be used to find the minimum and maximum states of process control
Can be made subject to multiple constraints
Common Algorithm
Simplex Method
Pusher Function
Maximize Flowrate subject to constraints
LP Optimization of Feedrate
Moving Horizon Based Dynamic Real Time Optimization - MHDRTO
Objective: The objective of this work is to integrate the Moving
Horizon Based Dynamic Real Time Optimization (MHDRTO) with a
well controlled continuous tablet manufacturing process.
Economically driven Dynamic Optimization
MHDRTO is Maximizing Profit. The outcome of optimization is the Production Rate. Constraints: Equipment operating limitations/ CQA’s. Control: Tablet weight, Hardness, Level, API composition.
MATLAB/gPROMS vs. COTS Research by Dr. Ravendra Singh/Assistant Research Professor – C-
SOPS NSF Engineering Research Center at Rutgers University
Non-Linear Data Modeling
Combination of Linear Regressions
Build up a series of linear models (regressions) to create a non-linear model
3. Neural Network Algorithm
E-mail Spam
Internet Browser
Recommender systems
Pattern RecognitionBar codersFacial identificationRobotics
PharmaSoft SensorsNon-Linear Control
Neural Networks
Linear Regression
0
2
4
6
8
10
12
14
1 2 3 4 5 6 7 8 9 10 11
Fit
Raw Data
0
100
200
300
400
500
600
700
1 3 5 7 9 11 13 15 17 19 21 23 25
0
100
200
300
400
500
600
700
1 3 5 7 9 11 13 15 17 19 21 23 25
1
2
3
Combine Multiple Linear FitsRegress Against Y Data
Built a Neural Net Model against Process Data and Hard Sensor values to Build Soft Sensor
Cross Validate the Soft Sensor against Hard Sensor
Can be used as a check against the Hard Sensor for alarming
Soft Sensor used when Hard Sensor is offline.
Neural Net – Pharma Soft
Statistically Optimal State Estimator Used to filter noisy data when you have a model
you can compare it against Uses a weighted average of the measurement
model prediction Weighted average is automatically updated every
new measurement to find the optimal value Numerous ApplicationsDe facto Standard RoboticsAerospaceMissile GuidanceEconomics Signal Processing
4. Kalman Filter
Takes a statistical average of:Measured VariableModel
Acts Recursively to continuously predict most probable state.
First used by NASA to predict location of rocketsUncertain GPS SignalPhysical Model error increases with time
Use measurement signal to correct errors with model.
Use model to validate measured values.
Kalman Filter
Kalman Filter Cannonball
Kalman Filter – Logistic GrowthBioReactor Application
dF(x)/dx = f(x)*(1-f(x))
PAT Data Noisy
Often models are available to predict the state
Kalman Filter can use both the find the OPTIMAL value
Statistically KF is the Best Guess
Good for control
Lower lags than typical first-order smoothing functions. Moving Block, Sovitzky-Golay
Kalman Filter in Pharma
Difficult to Introduce New Technology New Industry InitiativesPATContinuous Manufacturing
Now is time to introduce Advanced Controls! MPCKalman FilterNeural NetLP OptimizationMSPC-Multivariate Statistical Process Control Batch and Continuous Analytics
Fuzzy Logic
Opportunity
Thanks!