PAT, QbD and Process Performance Monitoring (The Impact for...
Transcript of PAT, QbD and Process Performance Monitoring (The Impact for...
Julian Morris1, Zeng-ping Chen2 and David Lovett3
Centre for Process Analytics and Control Technologies1Newcastle University, 2Strathclyde University, 3Perceptive Engineering
PAT, QbD and Process Performance Monitoring (The Impact for Process Systems Engineering)
CPAC Rome Workshop March 2008
Overview of Presentation
What is Process Systems Engineering?
Where are we now, where are we going and how important is variability?
Look at some novel methodologies for building fit-for-purpose calibration and statistical monitoring models
Highlight some process analytical technologies through work aiming to improve understanding of monitoring and controlling batch crystallisation processes using advanced Chemometrics
Show how some new closed loop process PAT control ideas are coming to fruition through innovative process systems engineering - Closing the Analytical Control Loop
Closure
The EU provides 32% of the worlds chemicals manufacturing through some 25,000 enterprises of which 98% are SMEs which account for 45% of the sectors
‘added value’, and 46% of all employees are in SMEs
What is Process Systems Engineering?
Process Systems Engineering (PSE) is a well established scientific area, concerned with the development of methods and computer-based tools for an integrated approach to all aspects of process modelling, simulation, design, operation, control, optimisation and management of complexity in uncertain systems across multiple time and scale lengths.
These tools enable Process Systems Engineers to systematically develop products and processes across a wide range of systems involving chemical, biological and physical changes from molecular and genetic phenomena to manufacturing processes and related business processes.
Where are we Now and Where are we Going
in Process Analytics and Control Technologies?
Impurities and Polymorphism (1)
Impurities and Polymorphism (2)
Impurities effect nucleation and growth processes, and hence canstabilise meta-stable polymorphic forms.
As product purity improves during process chemistry work-up, the “stable” polymorphic form can change !
e.g. RITONOVIR aids drug which changed from anhydrous to hydratecrystal after launch:
with lower solubility and hence bio-availability.
product was withdrawn for a year and reformulated.
new FDA approval needed – mega cost implication !
API Crystal Size Distribution
During the development and testing of a new drug an issue arose with the ‘smell’ of an API product.
Although the material was very pure (>99%) there was a residual ‘smell’ of the crystallisation solvent which persisted even after drying
This made the ‘placebo’ tablets difficult to disguise (blind)
There was also a problem with the final blending, drying and compressing operations where the formulation team wanted larger crystals which had better flow and compression characteristics for tablet making.
Courtesy of Gerry Steele, AstraZeneca. APACT07
API Crystal Size Distribution
Origin of the ‘Smell’:
Crystal size
Courtesy of Gerry Steele, AstraZeneca. APACT07
Temperature CyclingRather than carry out a re-crystallisation, temperature cycling the material causing the difference in the dissolution rates can be used increase crystal size.
T Cycle 1 T Cycle 2 T Cycle 3
Courtesy of Gerry Steele, AstraZeneca. APACT07
Where Are We Going?From thisFrom this
Reactor
TurbidityEnablirT, pH PC
Manhole
FTIR spectrometer
Turbidityprobe
FTIR probe
N2 purgingtube
pH probe Reactor
TurbidityEnablirT, pH PC
Manhole
FTIR spectrometer
Turbidityprobe
FTIR probe
N2 purgingtube
pH probe Reactor
TurbidityEnablirT, pH PC
Manhole
FTIR spectrometer
Turbidityprobe
FTIR probe
N2 purgingtube
pH probe Reactor
TurbidityEnablirT, pH PC
Manhole
FTIR spectrometer
Turbidityprobe
FTIR probe
N2 purgingtube
pH probe
…. to this
Reactor
TurbidityEnablirT, pH PC
Manhole
FTIR spectrometer
Turbidityprobe
FTIR probe
N2 purgingtube
pH probe Reactor
TurbidityEnablirT, pH PC
Manhole
FTIR spectrometer
Turbidityprobe
FTIR probe
N2 purgingtube
pH probe Reactor
TurbidityEnablirT, pH PC
Manhole
FTIR spectrometer
Turbidityprobe
FTIR probe
N2 purgingtube
pH probe Reactor
TurbidityEnablirT, pH PC
Manhole
FTIR spectrometer
Turbidityprobe
FTIR probe
N2 purgingtube
pH probe
…. to this…. to this…
Pharma Process Control
Pharmaceutical companies typically operate at around 2 to 2½ Sigmacompared with 6 Sigma in world class manufacturing
Variability (or PAT) by Edwards Deming
Cease reliance on mass inspection to achieve quality.
Eliminate the need for mass inspection by building quality into the product in the first place.
Dr W. Edwards Deming (Circa 1980s)
“Learning is not compulsory, ……. Neither is survival”
Variability and Coping with Different Product Recipes,
Different Processing Units, Different Production Sites,
Different Spectroscopic Probes & Probe Locations,
……, etc
Variability with Different Product Recipes, Processing Units, Production Sites, Spectroscopic Probe Locations, …
The elimination of differences between different modelling or calibration applications is a major issue in flexible process manufacturing
Different spectroscopic probes in different vessels or on different sites
Different reactors (unit operations), multiple recipes, etc
Normally this would require the construction of separate models for each individual situation.
A simple but partial solution is the subtraction of local or global mean levels
A better solution is to develop a multi-group (sub-space) model or calibration such that the interest can then focus on within process (or product) variability.
Martin, E.B. and A.J. Morris, (2003), “Monitoring Performance in Flexible Process Manufacturing”, Proceedings of IFAC ADCHEM’2003, Hong KongLane S., et al, “Performance Monitoring of a Multi-Product Semi-Batch Process." Journal of Process Control, 2001, 11(1), 1-11.
Case Study: Between and Within Group VariationMulti product manufacturing with Unilever (Lever Fabergé).
Approximately 50 different products and five main production units, complicated by multiple recipes.
Monitoring the process using standard MSPC would mean the building and maintenance of approximately 250 statistical monitoring models.
Too complicated for ‘quality’ and operational personnel.
High model-maintenance costs
Composition of the Multi-Recipe Data SetsTo demonstrate the application of multiple group PLS the production of two recipes are considered. Each process mixer is also considered as a separate ‘Recipe’ and thus the process model contains four distinct groups.
Note that Recipe 2 has fewer raw materials and fewer process variables than Recipe 1.
Recipe Group Mixer Number ofBatches
RawMaterials
QualityVariables
1 1 ( ) 1 19 23 1
2 (+) 2 21 23 1
2 3 (o) 1 20 17 1
4 (x) 2 29 17 1
ProcessVariables
120
120
90
90
Impact of Different Recipes
Principal Component 1
-10 -5 0 105 15
5
0
-5
10
-15-10
Prin
cipa
l Com
pone
nt 2
This model contains both within group and between group variation.
Subtle process events cannot be detected; the greater the number of distinct groups in the data set, the greater the impact of between group variation.
Agitator
Location 1Location 2Location 3
Location 4
Spectroscopic Probe Location PCA Plot of Measured Spectra
Impact of Spectroscopic Probe Locationin a Reaction Vessel
PC1
PC
2
Location 3
PC
2
Location 2
Location 4
Location 1
Multi-group PLS – Pooled Covariance Model
-6 -4 -2 0 2 4 6 8
6
4
2
0
-2
-4
-6
Latent Variable 1
Late
nt V
aria
ble
2The pooled variance-covariance matrix is constructed as a weighted sum of the individual elements of the individual variance-covariance matrices.
Case Study – Dosing Valve Problem
The monitoring chart shows the scores jumping outside the actionand warning limits as the process malfunction impacts.
-6 -4 -2 0 2 4
-15
-10
-5
0
5
Prin
cipa
l Com
pone
nt 2
Principal Component 1
Fault Diagnostics
Variables (62 - 65) were identified as being related to raw material addition.A faulty dosing-control valve was located as being to source of the problem.
Variables0 20 40 60 80
-2
0
2
4
6
8
Con
tribu
tion
Variability – Modelling and Calibration ChallengesProcess Issues:
Equipment characteristics; site-to-site process differences, etc
Fluctuations in both control and external process variables
Cell improvement; cell line changes; media changes
Small data sets are an issue but can be enhanced through Bootstrap Aggregation
Analytical Issues:
Separating absorbance from multiplicative light scattering effects caused by the variations in optical path length
Inter probe variability: impact of component variance on PLS calibration – can probe differences be accomodated or eliminated?
Can calibration models be made generic for different production unit operations / production lines?
Variability - Coping with Minimal Experimental DataThe need for fit-for-purpose statistical models can be limited due to intrinsic data sparsity resulting from the small number of objects available, e.g. experiments, batches, runs, etc, compared with the large number of process variables wavelengths measured.
This has been successfully addressed in PLS calibrations and in building statistical monitoring and predictive models through the addition of Gaussian noise to the original data and combining multiple models using Bootstrap Aggregated Regression.
This allows the development of robust calibration or process models and has been shown to lead to a decrease in prediction errors as a consequence of the increase in the data density.
Martin, E.B. and A.J. Morris, “Enhanced bio-manufacturing through advanced multivariate statistical technologies”, Journal of Biotechnology 99, 2002, 223-235Conlin, A. et al, “Data augmentation: an alternative approach to the analysis of spectroscopic data”, Chemometrics and Intelligent Laboratory Systems 44,1998,161-173Zhang, J. et al, “Estimation of impurity and fouling in batch polymerisation reactors through the application of neural networks”, Computers and Chemical Engineering 23, 1999, 30, 301-314
Advanced Chemometric Methods
FTIR
Supersa-turation
Size
GrowthkineticsUSS
BatchBatchProcessprocess
monitoringmonitoring& control& control
Processconditions
Heat transferLDA/PIV Reaction
CalorimetryMixing &scale-up
CFD
Shape Video microscopy
MSZWUVvis
Nucleationkinetics
Reactantrheology
Polymor-phic form XRD
FTIR
Supersa-turation
Size
GrowthkineticsUSS
BatchBatchProcessprocess
monitoringmonitoring& control& control
Processconditions
Heat transferHeat transferLDA/PIVLDA/PIV Reaction
CalorimetryMixing &scale-upMixing &scale-up
CFDCFD
Shape Video microscopyVideo microscopy
MSZWUVvis
Nucleationkinetics
Reactantrheology
Polymor-phic form XRD
Chemicals Behaving Badly (CBBII): Integrated In-Process Analytics and Closed Loop Control
CBB#2 Project: collaboration with
Leeds, Heriot-Watt and Newcastle University
PartnersAEA Technology
AstraZenecaBede Scientific Instruments
BNFLClairet Scientific
DTIEPSRC
GlaxoSmithKlineHEL
Malvern InstrumentsPfizer
Syngenta
Smoothed Principal Component Analysis (SPCA)Enhancing signal-to-noise ratio
iii k rQQIXrX )( TT ×+×= λ+×= ],,[],,[ 2121 cc rrrrrrF LL nsF FxFxFxx +==
mi ,,2,1 L=
SPCA
2θ16.0 16.5 17.0 17.5 18.0 18.5 19.0
Inte
nsity
1200
1250
1300
1350
1400
1450
16.0 16.5 17.0 17.5 18.0 18.51200
1250
1300
1350
1400
1450
16.0 16.5 17.0 17.5 18.0 18.5 19.01200
1250
1300
1350
1400
1450
16.0 16.5 17.0 17.5 18.0 18.51200
1250
1300
1350
1400
1450a
2θ2θ16.0 16.5 17.0 17.5 18.0 18.5 19.0
Inte
nsity
1200
1250
1300
1350
1400
1450
16.0 16.5 17.0 17.5 18.0 18.51200
1250
1300
1350
1400
1450
16.0 16.5 17.0 17.5 18.0 18.5 19.01200
1250
1300
1350
1400
1450
16.0 16.5 17.0 17.5 18.0 18.51200
1250
1300
1350
1400
1450a b
1.000%0.800%
0.533%0.389%
0.178%0.000%
16.0 16.5 17.0 17.5 18.0 18.5 19.0In
tens
ity
1250
1300
1350
1400
1450b
1.000%0.800%
0.533%0.389%
0.178%0.0%
b
1.000%0.800%
0.533%0.389%
0.178%0.000%
16.0 16.5 17.0 18.0 18.5 19.01250
1300
1350
1400
1450
1.000%0.800%
0.533%0.389%
0.178%0.0%
b
2θ
b
1.000%0.800%
0.533%0.389%
0.178%0.000%
16.0 16.5 17.0 17.5 18.0 18.5 19.0In
tens
ity
1250
1300
1350
1400
1450b
1.000%0.800%
0.533%0.389%
0.178%0.0%
b
1.000%0.800%
0.533%0.389%
0.178%0.000%
16.0 16.5 17.0 18.0 18.5 19.01250
1300
1350
1400
1450
1.000%0.800%
0.533%0.389%
0.178%0.0%
b
2θ2θ
Raw (a) and Processed (b) XRD profiles (by SPCA) for 6 XRD data sets of mannitol-methanol suspensions with the contents of mannitol equalling to 0.0%, 0.178%, 0.389%, 0.533%, 0.8% and 1.0% g/ml, respectively
SPCA in Morphology Monitoring (β - Form)
Concentration (%)0 2 4 6 8 10
Pea
k he
ight
-400
-200
0
200
400
600
800
R= 0.9540
2θ =29.8272
0 2 4 6 8 10-400
-200
0
200
400
600
800
R= 0.9540
2θ =29.8272
Concentration (%)0 2 4 6 8 10
Pea
k he
ight
-400
-200
0
200
400
600
800
R= 0.9540
2θ =29.8272
0 2 4 6 8 10-400
-200
0
200
400
600
800
R= 0.9540
2θ =29.8272
Concentration (%)0 2 4 6 8 10
Peak
hei
ght
-400
-200
0
200
400
600
800
R = 0.9951
2θ =29.8272
0 2 4 6 8 10-400
-200
0
200
400
600
800
R = 0.9951
2θ= 29.8272
Concentration (%)0 2 4 6 8 10
Peak
hei
ght
-400
-200
0
200
400
600
800
R = 0.9951
2θ =29.8272
0 2 4 6 8 10-400
-200
0
200
400
600
800
R = 0.9951
2θ= 29.8272
The raw and pre-processed XRD profiles of the β-form with concentration varying from 0.02% to 8.00%; relationships between concentrations and peak heights at peaks B3 of the raw and processed spectra.
2θ
10 12 14 16 18 20 22 24 26 28 30 32 34
Inte
nsity
1000
2000
3000
4000
5000β - form
B3
2θ
10 12 14 16 18 20 22 24 26 28 30 32 34
1000
2000
3000
4000
5000β - form
B3
B2B1
2θ
10 12 14 16 18 20 22 24 26 28 30 32 34
Inte
nsity
1000
2000
3000
4000
5000β - form
B3
2θ
10 12 14 16 18 20 22 24 26 28 30 32 34
1000
2000
3000
4000
5000β - form
B3
B2B1
10 12 14 16 18 20 22 24 26 28 30 32 34
Inte
nsity
1000
2000
3000
4000
5000
6000
B1B2
B3
? - form
2θ
10 12 14 16 18 20 22 24 26 28 30 32 34
1000
2000
3000
4000
5000
6000B1
B2
B3
β - form
10 12 14 16 18 20 22 24 26 28 30 32 34
Inte
nsity
1000
2000
3000
4000
5000
6000
B1B2
B3
? - form
2θ
10 12 14 16 18 20 22 24 26 28 30 32 34
1000
2000
3000
4000
5000
6000B1
B2
B3
β - form
Bede MONITORTM In-process XRDCrystal Polymorph Monitoring & Control Using SPCA
Typically circa 1 wt % detectable via in-process XRD, much lower with advanced chemometric analysis (Smoothed PCA)
Temperaturereaders
Temperaturemonitors
0
2000
4000
6000
8000
10000
12000
15 18 21 24 27 302Theta [degree]
Inte
nsity
t = 200 s
t = 600 s
t = 1000 s
t = 1400 s
t = 1800 s
t = 2200 s
t = 2600 s
t = 3000 s
t = 3400 s
t = 3800 s
t = 4200 s
t = 4400 s
t = 4600 s
αpeaks
βpeaks
0
2000
4000
6000
8000
10000
12000
15 18 21 24 27 302Theta [degree]
Inte
nsity
t = 200 s
t = 600 s
t = 1000 s
t = 1400 s
t = 1800 s
t = 2200 s
t = 2600 s
t = 3000 s
t = 3400 s
t = 3800 s
t = 4200 s
t = 4400 s
t = 4600 s
αpeaksαpeaks
βpeaksβpeaks
System provides capability to monitor polymorphic form “in-process”unaffected by product separation prior to analysis.
Loading Space Standardisation (LSS)Correcting non-linear shift and broadening in spectral bands caused by temperature fluctuations
)()()(1
ii
K
kikkiii ttct esx += ∑
=,
)()(
)()(
m
11
tt
tt
PCAm
PCA
'
'
TPX
TPX
⎯⎯→⎯
⎯⎯→⎯
MMM LSStt ii ⇒~)(P )()( refLSS
test tt xx ⎯⎯→⎯
LSS
Wavelength (nm)850 900 950 1000 1050
Abso
rban
ce (A
U)
-0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Wavelength (nm)850 900 950 1000 1050
Wavelength (nm)850 900 950 1000 1050
Abso
rban
ce (A
U)
-0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Abso
rban
ce (A
U)
-0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Abs
orba
nce
(AU
)
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Wavelength (nm)850 900 950 1000 1050
Abs
orba
nce
(AU
)
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Abs
orba
nce
(AU
)
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Wavelength (nm)850 900 950 1000 1050
Wavelength (nm)850 900 950 1000 1050
Optical Path-Length Estimation and Correction (OPLEC)Separating absorbance from multiplicative light scattering effects caused by the variations in optical path length
], , ,[] , , ,[),( 2121 mjjjjmji cccOPLECb LL === cxxxXcX ,,
i
J
jjjiii cb εsx += ∑
=1, multiplicative parameter-ib
OPLEC
OPLEC has recently been applied for Raman Scattering applications incomplex crystallisation processes.
Wavelength (nm)850 900 950 1000 1050850 900 950 1000
Abs.
(AU
)
1.8
2.0
2.2
2.4
2.6
2.8
3.0
3.2
2.0
2.2
2.4
2.6
2.8
3.0
3.2
Wavelength (nm)850 900 950 1000 1050850 900 950 1000
Wavelength (nm)850 900 950 1000 1050850 900 950 1000
Abs.
(AU
)
1.8
2.0
2.2
2.4
2.6
2.8
3.0
3.2
2.0
2.2
2.4
2.6
2.8
3.0
3.2
Abs.
(AU
)
1.8
2.0
2.2
2.4
2.6
2.8
3.0
3.2
2.0
2.2
2.4
2.6
2.8
3.0
3.2
Abs.
(AU
)1.85
1.9
1.95
2.0
2.05
2.1
2.2
2.15
Wavelength (nm)850 900 950 1000 1050850 900 950 1000
Abs.
(AU
)1.85
1.9
1.95
2.0
2.05
2.1
2.2
2.15
Abs.
(AU
)1.85
1.9
1.95
2.0
2.05
2.1
2.2
2.15
Wavelength (nm)850 900 950 1000 1050850 900 950 1000
Wavelength (nm)850 900 950 1000 1050850 900 950 1000
250l Pilot Plant Batch Agitated Vessel
Outlet (8mm) port Inlet (12mm) port
α-sizer
Supersaturation Control System Upgrade to PI Capability
INPUTBlock
User DefineModel
User Define “S”set point
HEL PC
Solubility Model
Macro
S(t)
PI ControllerControl Block
Data Processing Block
C(t)
ENABLIR
SOFTWARE
WINISO
SOFTWARE
C(t)
FTIR PC Spectra
PLS model
FTIR Spectrometer
Crystallisation Vessel
4 – 20 mA
INPUTBlock
User DefineModel
User Define “S”set point
HEL PC
Solubility Model
MacroMacro
S(t)
IMC Based PI ControllerControl Block
Data Processing Block
C(t)
ENABLIR
SOFTWARE
WINISO
SOFTWARE
C(t)
FTIR PC Spectra
PLS model
C(t)
FTIR PC Spectra
Cal. model
FTIR Spectrometer
Crystallisation Vessel
4 – 20 mA
Supersaturation Control of L-Glutamic Acid250 litre Plant Crystalliser
0
10
20
30
40
50
60
70
80
90
100
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300Time (min)
Tem
pera
ture
, Con
cent
ratio
n, S
olub
ility
& T
urbi
dity
0.0250.0750.1250.1750.2250.2750.3250.3750.4250.4750.5250.5750.6250.6750.7250.7750.8250.8750.9250.9751.0251.0751.1251.1751.225
Supe
rsat
urat
ion
(S =
C/C
*)
Temperature (°C) Concentration (g/500ml)
Solubility (g/500ml) Turbidity (%)Supersaturation SlimitSlimit
Smax = 1.125S = 1.1
Smin = 1.075Started SupersaturationControl
5% seedsadded
0
10
20
30
40
50
60
70
80
90
100
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300Time (min)
Tem
pera
ture
, Con
cent
ratio
n, S
olub
ility
& T
urbi
dity
0.0250.0750.1250.1750.2250.2750.3250.3750.4250.4750.5250.5750.6250.6750.7250.7750.8250.8750.9250.9751.0251.0751.1251.1751.225
Supe
rsat
urat
ion
(S =
C/C
*)
Temperature (°C) Concentration (g/500ml)
Solubility (g/500ml) Turbidity (%)Supersaturation SlimitSlimit
Temperature (°C) Concentration (g/500ml)
Solubility (g/500ml) Turbidity (%)Supersaturation SlimitSlimit
Smax = 1.125S = 1.1
Smin = 1.075Started SupersaturationControl
5% seedsadded
L-Glutamic Acid Crystals (a) Seeds (b) Product
(4x 1x)
100 μm
(4x 1x)
100 μm100 μma b α β
Multivariate Statistical Process Control (MSPC)or
Process Performance Monitoring
Is the Data Unfolded and Scaled Properly?
Batch Data Unfolding (Nomikos & MacGregor) - MWPCAN&MB
atch
es
Variables
Time
1
IJ1
K
1J
I
JK
. . . . . .T1 T2 T3
Batch Data Unfolding (Wold et al) - PCAW
Bat
ches
Variables
Time
1
IJ1
KJ
B1
1
K
IK
B3
B2
Key References
Nomikos and MacGregor (N&M). Batch-wise unfolding.“Multivariate SPC charts for monitoring batch processes”. Technometrics. 37, 41-59, 1995.
Wold, Kettaneh, Friden and Holmberg. Variable-wise unfolding“Modelling and diagnostics of batch processes and analogous kinetic experiments”. Chemometrics and Intelligent Laboratory Systems. 44, 331-340, 1998.
• Unfolding in this direction followed by mean centering and scaling does not remove the mean trajectories from the data and hence only captures the covariance among the mean trajectories of the variables, which is not of interest in process performance monitoring.
• Rather, it is the covariance structure of the deviations of the variables about these mean trajectories that is of interest.
• This unfolding approach also only provides a ‘static’ model that captures only the local covariance structure amongst the variables and does not account for the dynamic behaviour of batch processes.
Non-linearity and Process Dynamics (MWPCANM)
Scaling can remove major non-linear component in the data.
Scaling may remove major dynamic component in the data.
0.5 1 1.5 20.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99
0.9970.999
Data
Prob
abili
ty
Normal Probability Plot
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.50.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99
0.9970.999
Data
Prob
abili
ty
Normal Probability Plot
Raw Data
Scaled Data
0 20 40 60 80 100-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time lag0 20 40 60 80 100
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time lag
Raw Data
Scaled Data
Non-linearity and Process Dynamics (PCAW)
Scaling cannot remove major non-linear component in the data.
Scaling cannot remove major dynamic component in the data. 0.5 1 1.5 2
0.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99
0.9970.999
Data
Prob
abili
ty
Normal Probability Plot
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5 10.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99
0.9970.999
Data
Prob
abili
ty
Normal Probability Plot
Raw Data
Scaled Data
0 20 40 60 80 100-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time lag0 20 40 60 80 100
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time lag
Raw Data
Scaled Data
Model Based Closed Loop Process Control
The Impact of Process Dynamics (Auto-correlated Data)
In the presence of auto-correlation, the action and warning limits calculated for a process performance monitoring scheme based on the usual assumptions that the observations are Independent and Identically Distributed (IID), are inappropriate making the monitoring charts unreliable, if not useless.
Developing an MSPC representation for the monitoring of processes, exhibiting time varying characteristics, will result in an increase in the number of false alarms and operational changes not being detected.
So, how do we deal with auto-correlated data?
Model-based Performance Monitoring
Plant
FirstPrinciples Model
orDynamic Empirical
Model
DynamicEmpirical
‘Error’ Model
+-
+-
Multivariate Monitoring
Charts
Inputs OutputsPlant
First Principles Modelor
Dynamic Empirical Model
Dynamic Empirical‘Error-whitening’
Model‘
+-
+-
+-
MultivariateMonitoring Charts
Inputs Outputs
Plant-ModelMismatch
Plant
FirstPrinciples Model
orDynamic Empirical
Model
DynamicEmpirical
‘Error’ Model
+-
+-
+-
Multivariate Monitoring
Charts
Inputs OutputsPlant
First Principles Modelor
Empirical Model
Dynamic Empirical‘Error-whitening’
Model‘
+-
+-
+-
MultivariateMonitoring Charts
Inputs Outputs
Plant-ModelMismatch
At the heart of a model-based control or monitoring system is a DYNAMIC model
Perc
enta
ge
99
95
80
402010
1
Process Data
60
5
80 130 180
Data
Model ResidualsDynamic Non-linear PLS
-1 0 1
9995
80
402010
60
5
1
Perc
enta
ge
-0.5 0.0 0.5 1.0
Perc
enta
ge
Model ResidualsLinear PLS
1
99
95
80
402010
60
5
Data
-1Plant-Model Residuals
Perc
enta
ge
99
95
80
402010
1
60
5
0 1-1 2-3 -2-4
Impact of Non-linearity – Normal Probability Plots
2 12 22 32 42 52 622 12 22 32 42 52 62
Process Data0.81.0
0.4
0.0
-1.0-0.8
-0.4
2 12 22 32 42 52 62
ResidualsPlant-Model Mismatch0.8
1.0
0.4
0.0
-1.0-0.8
-0.4
2 12 3222 5242 62
Model ResidualsDynamic PLS
0.81.0
0.4
0.0
-1.0-0.8
-0.4
2 12 22 32 42 52 62
Model ResidualsTime Series Model
0.81.0
0.4
0.0
-1.0-0.8
-0.4
Impact of Non-linearity – Normal Probability Plots
Closing the Analytical Control Loop
Challenges for Real Time Closed Loop Control and Real Time Release
At least 15 years worth of academic control theory, simulations and occasionally pilot studies have reported on the benefits and challenges of improving operational performance through improved control, more recently in the Pharmaceutical Industry.
Since then the FDA has opened the way to a technology-led quality assurance regime there has been hectic activity directed towards defining an implementation methodology for measurement, modelling and data analysis.
Many of the larger automation companies have focussed on the overall infrastructure to ensure that the integration of data from many sources is recorded securely and a traceable batch record is generated.
Hopefully one of the key outcomes is that the integration of the data will also integrate the users of the data such as the chemists, biologist, engineers, to provide a process process systems engineering approach to design, build, operate, monitor and improve production facilities.
History:> 15 years of academic control theory, simulations & pilot studies in Pharma2004 - FDA opened the way to a technology-led quality assurance regime Hectic activity to define methodology for measurement, Design of Experiments, modelling, data analysis, etc.Large automation companies focus on overall infrastructure for data management.Aim to ensure integration of data from many sources is recorded securely and a traceable batch record is generated.
Recent:Aim to also integrate the users of data (chemists, engineers, Lab analysts, process development scientists)Provide a process systems engineering approach to design, build, operate, monitor and improve production facilities – aligning with Lean and Six sigma initiativesUnderstanding the challenges with data compatibility, data fusion and data validity. Exploration of the benefits and challenges of improving operation through improved control.Beginning to investigate the use of Advanced Control techniques on unit operations and production lines.
Challenges for Real Time Closed Loop Control
Why is Advanced Process Control required?
Why and How to verify the integrity of data used in PAT applications
Using Continuous, Discrete and Spectral data in a co-ordinated way
Batch Control – Advisory or Closed Loop?
Challenges of Real Time Closed Loop Pharma Control
Why is Advanced Control Required?
PAT and Advanced Process ControlWhy we’re here
PAT can be used as part of a tool box to optimise the way pharmaceuticals are manufacturedProvide greater understanding of what to controlPotentially provide a means to control “Critical Attributes” by monitoring and adjusting “Critical Parameters” in real timeProvides some of the ability, to reduce the risk of process variability, effecting process capability and product quality
Current Model
Variable Process
Variable Process Model
Fixed ProcessVariable Raw MaterialInput to process Variable Output
Consistent OutputVariable Raw Material
Input to process
Continuous Quality Verification
Understand process constraints and complex process interactions.
Build multivariate correlation model between variables and actuators, causes and effects.
Predict impact of known disturbances on operations.
Predict, Advise, Make co-ordinated moves on multiple actuators.
Exploit opportunities to push quality / throughput close to constraint limits
DISTURBANCE(FEED FORWARD VARIABLES)
CONTROLLEDVARIABLES
ACTUATORS(set -points
of PID loops)
MODELPREDICTIVECONTROLLER
DISTURBANCE
CONTROLLEDVARIABLES
ACTUATORS(set -points
of PID loops)
MODELPREDICTIVECONTROLLER
£ $ €
Time
Proc
ess
Varia
ble
eg M
oist
ure
Upper Constraint
σ
Predictive Control& Optimisation
Predictive Engine:Predicts the impact of process disturbances
Multivariate Process Modeling &Model Predictive Control
How Do You Verify the Integrity of Data for PAT Real Time Control & Real Time Release
Univariate Data Quality Monitor:
Individual Signal ValidationLogical ChecksStatistical checks
Multivariate: Using Robust PCA
Outlier detectionOutlier IdentificationData Quality Monitoring
Data Quality Monitoring
Data Quality record underpins the validity of the system – critical in a validated environment
21 CFR part 11 Records
Raw Process Data Processed Data
All signals used to infer (model) the Critical To Quality Parameters are included and monitored individually within a Data Quality Monitor
whose output is automatically logged
Robust Data Cleansing / Pre-processing
Outlier Removal(Robust PCA
HampelFiltering)
21 CFR part 11 Record
Clean Data Set
Real Time Quality Control (Using Spectral Data)
Spectral Data
Spectral Data
Data Quality Monitor
Process Data
Discrete Data
21 CFR part 11 Records
Imported fromModel Development File
Real TimePre-
Processing
Design Space
Control Space
PLS/PCA Calibration
Model
DynamicPCA
Controller
Spectral Data
Process Data
Discrete Data
Real time pre-processed data • Continuously measured OR• A-periodically measured OR• Real time value Inferred from
calibration model OR• End-point value inferred from
calibration model OR• Scores of calibration model are
CTQ parameters
CtQ Parameters
Real Time Quality Control (Integrated Data Management)
Key Challenges for Pharma Batch Control
Availability of Critical to Quality (CtQ) MeasurementsAre CtQ parameters continuously measured?Are CtQ parameters measured by lab assay at end or during batch?Can CtQ parameters be inferred from PAT sensors with calibration model?
Process Non-linearityAs a function of process operating point (e.g. Arrhenius reaction rate, etc)Time varying dynamics as batch evolvesChanges in sign of process gain can be especially challenging
Trajectory trackingMany control systems excellent at regulating to fixed set-pointMany control systems excellent at rejecting disturbancesFew are good at tracking a time varying set-point profile
Key Challenges for Pharma Batch Control
Jacket Temperature SPReactor Temperature
Typically the CtQ parameter isn’t directly measuredControlled indirectly by making process variable track a pre-defined trajectory
Successful if: process variable trajectory guarantees CTQ parameter will hit target end-point within acceptable quality limitsAND variability in initial conditions/raw material is low or does not impact end-point AND process dynamics are ‘sufficiently linear’
Manipulated VariableControlled Variable
Temperature lags the ramping SP trajectory
Temperature overshoots
Traditional PID Control
Batch Model Predictive Tracking
Jacket Temperature SP
Manipulated VariableControlled Variable
No temperature overshoot
Temperature tracks ramping SP very well
Reactor Temperature
CtQ parameter is controlled indirectly by controlling a process variableDynamic process model captures dynamic multivariable process interactions.Model predicts trajectory of controlled variable(s) over future horizon.Controller computes trajectory of MV moves over future horizon to minimise current and future Set-Point error.Only the first move is applied, process repeats at next control intervalThis is a ‘moving window’ approach.
End-point value of CtQ parameter continuously estimated throughout the batchMV trajectories modified by controller to make estimated end-point hit targetUnfolded PLS model - predicts end-point value based on MV and process variable trajectories over the entire batchAt each point in the batch, computes trajectory of MV moves over entire batch to minimise the error between the predicted end-point value and end-point target
Start of batch End of batch
TIME
Controller calculates future MV moves over the
whole future window to the end of the batch
Current time
Trajectory of CtQ parameter
CtQ end-point target
Manipulated variable (MV)
Batch End-Point Control
Delivering Real Time ReleaseParametric or Real Time Release is a consequence of the process being:
Well understood,Engineered out and residually measured,Characterised and controlled,VARIABILITY understood and monitored and controlled.
Given that an appropriate measurement platform supports understanding and control, achieving this demands:
Robust data quality and control of whole system suitability – not just analytical but conventional measured systems must be quality monitored to ensure that behaviour outside the bounds of understanding is identified.Appropriate understanding of system sensitivity builds confidence and early warning systems must therefore themselves be sensitive.Periodic calibration an observation in operation alone are unlikely to pick up small errors introduced by sensor performance in process.Control to multivariate trajectories and end points demands a level of precision and understanding through the system that supports confident statements about quality and uncertainty – nothing else will be acceptable to the FDA.
Practicalities of Real Time ReleaseContrast testing between unit operations and final Quality Assay:
Where is the bulk of the testing for a complex product which undergoes (say) 10 processing operations?What is the overall lead time associated with stage-to-stage and final product testing?Is the quality / cost / productivity burden most affected by the ability to reduce variability and get predictable stage to stage testing or by targeting final product tests?With more potent compounds in play, the ability to obviate some stage-to-stage testing can have far reaching benefits beyond simple fixed cost reduction including reduced occupational exposure.
Final product assay testing is going to be with us for some time yet …..Multi stage modelling is the challenge to be met in deriving analogues for final product quality test parameters, these are not yet well understood.Data collected must be robust and with only one chance to produce quality at minimum cost, a focus on characterising product and controlling variability in unit operations is logical and builds a platform for Real Time Release.
PAT Sensors in Closed Loop Process Control– Some Challenges
Real-Time Management of Process Data - maintaining a traceable record of Data Quality.
Real time pre-processing – needs to be consistent with and traceable to the unique calibration model.
No control system is going to control a spectrum of several hundred simultaneous values; so what is important?
Is there is a fit-for-purpose calibration model to infer specific product properties?
Are there particular features / segments of the spectrum of interest?
Should the scores of the PCA/PLS calibration model be controlleddirectly using Latent Variable controllers?
Closure and Thanks
In this talk, I have …
Reviewed variability and coping with different product recipes / formulations, different processing units, different production sites, different spectroscopic probes & probe locations
Reviewed some novel methodologies for building robust calibration and statistical monitoring models.
Highlighted some process analytical technologies through work aiming to improve understanding of batch crystallisation processes and its scale-up, and the provision of enhanced process understanding through advanced chemometrics.
And shown how some new closed loop process PAT control ideas arecoming to fruition through innovative process systems engineering.
High Throughput Experimentation and Process Intensification in Product and Process Development
I will be happy to attempt to answer your questions
Many thanks to CPAC for their kind invitation
and of course you for your kind attention