Real Time Release in Continuous Solid Dose Manufacturing:Systematic Characterization of Material Properties, and Optimal Design of Sensing and Control Methods

Fault tolerant control for safe plant operation

November 10, 2016

Content

Systematic framework for control design & analysis

Example: continuous feeder & blender system

Summary & conclusions

Introduction

Batch vs. continuous manufacturing

[1] Lee SL, O'Connor TF, Yang X, Cruz CN, Chatterjee S, Madurawe RD, Moore CMV, Yu LX, Woodcock J. Modernizing pharmaceutical manufacturing: from batch to continuous production. Journal of Pharmaceutical Innovation. 2015;10:191-199.

Continuous manufacturing in pharma.

FDA Quality Related Guidance Initiativese.g. PAT Guidance (2004)

...e.g. OPQ Standup and ETT Formation (2015)

Flexible demandPrecise medicine Competiveness

... Smaller production facilities

Minimizing manufacturing cost

Enhancing reliability and flexibility

Monitoring & control drug quality on a continuous basis rather than through post-production, batch-based testing

Continuous manufacturing pilot plant

Potter B36 in Purdue University

Minimum

Demands for efficient process monitoring & control

Monitoring & control

Port

abili

ty

Mo

du

lari

zed

de

sign

Op

erat

ing

con

dit

ion

s

Pro

cess

mo

de

ls

Mat

eria

l pro

per

ties

Pro

cess

de

sign

Un

it o

per

atio

ns

Pro

du

ct d

esi

gn

Model-plant mismatchPAT calibrationProcess disturbanceSystem interactionsProcess nonlinearityPlant start-upPlant shut-downControllabilityStabilityRobustnessSelf-adaptivenessController tuning

Challenges:

Implementation layer

Final quality assurance

Au

ton

om

ou

s

Content Introduction

Systematic framework for control design & analysis

Example: continuous feeder & blender system

Summary & conclusions

C3 Process Monitoring & Control

C4 System Performance & Verification

Process model

C1 Material library

Pilot plant

C2 Sensing & RTR strategy

Material database Process characterisation Feasible formulation space

⁞

Operating cost Process maintenance

⁞

Control variables Control set points

⁞

Process parameters Unit operations

⁞

Project objectives & outlines

Systematic framework for control design & analysis (1)

Layer 2

•Use of dynamic models for prediction•Model Predictive Control, Adaptive Control,

Gain Scheduling, etc.•Integrated across multiple unit operations

Layer 1

•Based on PAT to measure CQAs and closed loop control (PFC)

•PID based more advanced control schemes•Single or multiple unit operations

Layer 0

•Built-in control systems by manufacturers at unit operation level

•Simple PID control•Single unit operation

Hierarchical control architecture

R1 Medium

R0 Low

R2 High

Risk levels & failure mode

Sensor accuracy, precision, sampling time Sensor location Control architecture reconfiguration Control method Controller tuning

⁞

Evaluation & regulatory filling

Good

Excellent

Fail

T2P: Time to productM2P: Magnitude to productMRI: Morari’s resilience index

[1] Morari M. Chem. Eng. Sci., 1983;38:1881-1891.

Continuous improvement

e.g. Good Manufacturing Practices (GMP) guidelines

Manufacturing processes are clearly defined and controlled. All critical processes are validated to ensure consistency and compliance with specifications.

Manufacturing processes are controlled, and any changes to the process are evaluated. Changes that have an impact on the quality of the drug are validated as necessary.

⁞

Formulation optimization Process reconfiguration ⁞

Systematic framework for control design & analysis (2)

SCADA: Supervisory control

and data acquisition

MES: Manufacturing

execution system

Hierarchical three-layer control

L0 & L1 control design

1. Selection of input/output variables

• Input and Output Effectiveness

• Controllability e.g., condition number (CN), relative gain array (RGA)

•Resilience e.g., Morari’s Resilience Index (MRI)

2. Pairing and interaction analysis

• Pairing selection e.g. RGA, relative interaction array (RIA), Niederlinski Index (NI), singular value decomposition (SVD)

• Interaction e.g., Dynamic Interaction Measure (DIM)

3. Close-loop control system evaluation

• Stability e.g., Nyquist’s stability criterion, Characteristic Loci Plots

• Robustness e.g., SVD, µ-based method

• Dynamic simulation e.g., MATLAB, Simulink

Mostly decentralized SISO PI/PID control loops Interactions among parallel control loops are important Interactions among hierarchical control loops

[1] Asmar BN. Control of a two-stage refrigeration system. PhD Dissertation. University of Nottingham, 1999

L2 model-based control / real-time optimization

Model Y = f(x, u)

State estimator

Model update

ProcessOutput YInput U

Objective J

U(k)

Control algorithm

Y(k+i)

U(k+i)

Sensors

Y(k)

Measure

x: system state variablesi = 1,2,..,Nc,...,NpNc : control horizonNp: prediction horizon

f: state-space modeldata-based modelempirical modelFirst-principles model...

Kalman filterExtended Kalman filterUnscented Kalman filterMoving horizon state estimator....

Past FutureCurrent

Past FutureCurrent

kK-1 K+2K+1 k+NPK+NC

L2 MPC real-time implementation strategy

kK-1 K+2K+1 k+NPK+NC

Y(k) Y(k+1)

u(k | t=k-1)

U(k: k+Np-1 | t=k-1)

U(k+1: k+Np| t=k)u(k+1 | t=k)

U(k+1: k+Np | t=k)MPC

Computation time ≤ sampling time

Risk analysis and control evaluation

Content Introduction

Systematic framework for control design & analysis

Example: continuous feeder & blender system

Summary

Loss-in-weightFeeders

Roller Compactor

Continuous Blender

LubeMill

Continuous Blender

Tablet Press

Continuous dry granulation process

Feeder & blender control system

Unit operation Process output (y)Process input

(u)Control Layer Controller type

API feeder API flowrate Screw rotation speed L0 PID

Exp. feeder Exp. flowrate Screw rotation speed L0 PID

Blender

API composition API flowrate L1/2 PID, Ratio, MPC

Powder flowrate Excipient flowrate L1/2 PID, MPC

API mixing RSD Rotation speed L1/2 PID, MPC

Rotation speed Motor current L0 PID

System identification: state-space models

𝑑𝑋

𝑑𝑡= 𝐴𝑐𝑋 𝑡 + 𝐵𝑐𝑈(𝑡)

Y 𝑡 = 𝐶𝑐𝑋 𝑡 + 𝐷𝑐𝑈 𝑡 + 𝑒(𝑡)

Continuous-time identified state-space model

decoupled control designe.g., condition number

Niederlinski indexrelative gain arrayMorari’s resilience index...

L1 control design

𝐺 = 𝐶−1 𝑆𝐼 − 𝐴 𝐵 + 𝐷

𝑋 𝑘 + 1 = 𝐴𝑑𝑋 𝑘 + 𝐵𝑑𝑈(𝑘)

Y 𝑘 = 𝐶𝑑𝑋 𝑘 + 𝐷𝑑𝑈 𝑘 + 𝑒(𝑘)

Discrete-time identified state-space model

Linear MPC designSampling time: 4sKalman filter designState estimation...

L2 control design

min𝑑𝑈

𝐽 = 𝑌𝑠𝑝 − 𝑌2+ 𝑤𝑢 ∆𝑈 2

𝑒 𝑘 ∶= 𝑤𝑒𝑒 𝑘 − 1 + 1 − 𝑤𝑒 𝑒(𝑘)

System identification: training data

*Sampling time of 4 second for all measurements, only 4 measurements shown here

Best fit

System identification: validation data

*Sampling time of 4 sec for all measurements

L1 control design & analysis: a check list

Step one

Step two

Step three

Control architectureSP: Set pointMV: Measured variableFT: Flowrate transducerCC: Composition controlFC: Flowrate controlRSD: Relative standard deviation

L1 control for feeder & blender system

FT

Continuous blender

/NIRMW

API feeder

FT

Excipient feeder

FT

MV:API mean comp & RSD.

SP:API comp.

L1-CC-2: PID

SP: Blend flowrate

L1-FC-1: PID

L0-FC-1: PID L0-FC-2: PID

MV:Exc. flowrate

L1-RC-1: Ratio

L0-MC-1: N./A.

SP: API RSD L1-RSD-1: PID

RC: Ratio controlMC: Motor controlL0: Level 0 equipment controlL1: Level 1 supervisory controlMW/NIR: Microwave/near infrared

L0/2 control for feeder & blender system

FT

Continuous blender

/NIRMW

API feeder

FT

Excipient feeder

FTL0-FC-1: PID L0-FC-2: PID

SP: Blender RPM

L0-MC-1: N./A.

MV:API mean comp. & RSD

MV: Blend flowrate

L2 MPC

Control architectureSP: Set pointMV: Measured variableFT: Flowrate transducerCC: Composition controlFC: Flowrate controlRSD: Relative standard deviation

MPC: Model predictive controlMC: Motor controlL0: Level 0 equipment controlL1: Level 1 supervisory controlMW/NIR: Microwave/near infrared

SP:API comp.

SP: powderflowrate

SP: RSD

Feeder & blender simulation in Simulink/MATLAB

Sampling time:FT: 1.0 sNIR: 2.0 sL0/1: real timeL2 MPC: 4.0 s

Risk Analysis for feeder & blender: R0 & R1

R0 risk: mass flow disturbances due to feeder reloadingR1 risk: calibration errors in the feeders

L0

L1

L2

Risk Analysis for feeder & blender: R2

R2 risk: material flowability changed due to humility while a step change on the powder flow rate was also executed.

L1

L2

Risk Analysis summary

* Blender D2P: Duration to reject, s; M2P: Magnitude to product, %

Control

D2R (s) M2P (%)

Remark

R0 R1 R2 R0 R1 R2

L0 176+ Fail Fail 0.65+ Fail Fail Fail for R1 & R2

L1 97+ 238+ 494* 0.38+ 1.52+ 0.48* Long D2R for R2

L2 81+ 183+ 259* 0.65+ 2.30+ 0.46* High M2P for R1

Performance evaluation based on the API composition.+ Indicators based on API mean composition; * Indicators based on API mixing RSD.

Risk Analysis for feeder & blender: R2 PAT gross error

Layer 0 control: no feedback control based on the PAT tools.

Drift

Drift

The PAT tool of NIR for API composition measurement drifted from 40 to 100 seconds and suffered the gross error from 100 to 240 seconds before it was corrected.

Layer 1 control: The PAT tool with gross error detoured the API mean composition and its mixing uniformity from their set points.

Drift

Drift

Layer 2 control: The Kalman filter with data reconciliation can reject the measurement gross error when large model-plant mismatches detected.

Drift

Drift

𝐱 𝑘|𝑘 = 𝐱 𝑘|𝑘 − 1 +𝐌 𝐲 𝑘 − 𝐲 𝑘|𝑘 − 1 𝐑

The Kalman filter with data reconciliation

𝐑 = 𝑑𝑖𝑎𝑔 exp(−𝑒𝑖𝑐𝜎𝑖

2

Summary & conclusions

The primitive systematic framework for control design and analysis based on SIMULINK platform has been developed.

L0 control loops response fastest to measurable R0 disturbances but fail to response to the unmeasurable disturbance R1.

L1/2 control loops are capable of intermediate product quality control. L2 control based on MPC is beneficial but the maintenance cost is also high.