From Design of Experiments to closed loop control

Petter Mörée & Erik Johansson Umetrics

Umetrics, The Company

• • • • • • • • • • • Part of

~1Billion

conglomerate The

market leader

(DOE) in software for multivariate analysis (MVDA) & Design of Experiments

25+ years

in the market Off line analysis tools On-Line process monitoring and fault detection 700+ companies, 7,000+ users Pharmaceutical, Biotech, Chemical, Food, Semiconductors and more Worldwide Presence with MKS Offices: – – – – – Umeå, Malmo, Sweden York, England Boston, San Jose, USA Singapore Frankfurt, Germany Close collaboration with universities in USA, Sweden, UK and Canada Partnership with

Sartorius

; global marketing, distribution, development and integration.

Building a capable process

Manufacturing DOE Control Strategy DOE Analysis Design Space QRA:

Quality Risk Assessment

MVDA

• •

QFD

Quality Function Deployment

DOE is a knowledge building tool for process development MVDA is used both for process understanding and process monitoring

3

Processes and their data are never perfect

Delegates at this meeting are of course excluded

• Multivariate data analysis (MVA) is a tool to learn from data • Marek used MVA and NIR to predict glucose nad other parameters inside the reactor • This talk will focus on process parameters – Tightly controlled • pH, pO2, Temperature – Parameters used for keeping tightly controlled at their sepoint • Stirring, airation, cooling, base addition ..

• – Commonly measured • CER, OUR … Monitor, interpret, control 4

Is this chart familiar?

DJIA = x1 * Merck + x2 * J&J + x3 * Pfizer + x4 * DuPont + ....

5

MSPC – Multivariate Statistical Process Control

Evolution Level – Monitoring

• Example of a fermentation

t 1

= x1 * Temperature + x2 * Pressure Predicted Scores [comp. 1] + x3 * Agitation speed t[1] (Avg) -3 Std.Dev

+ x4 tPS[1] (Batch S0058_A_854826) * pO2 .

Control limits 4 3 0 -1 2 1 Average (signature) of all good experiments -2 -3 0 10 20 30 40 80 50 60 70 $Time (normalized) New run/experiment assessed by the model 90 100 110 120 SIMCA-P+ 11 - 01.08.2009 14:42:24 130

Variable, Batch: 617, Phase: 16

Statistical Process Control MULTIVARIATE CONTROL CHART

-15.4122 * 543071TT--607 - 15.4122

-2.84936 * 543071TT--600 - 2.84936

0.00281207 * 543071PT--644 + 0.00767695

-1.10938 * 543071TT--602 - 1.10938

0.00826856 * 543071-Acetone_43x10e13 1.0

0.9

0.8

0.3

0.2

0.1

0.0

0.7

0.6

0.5

0.4

2 1 0 8400 8500 8600 8700 8800 8900 9000 9100 9200 9300 Num 9400 9500 9600 9700 -1 9800 9900 10000 10100 SIMCA-P+ 11 - 09.03.2011 19:17:38 -2 Model Data Ciclo - Oct 2010 v5.M2:16

control limits (± 3

s +3 Std.Dev

from avg.)

-3 Std.Dev

t[1] (Aligned): 635 -3 -4 -5 -6 -7 -8 -9 t[Comp. 1]/t[Comp. 2] -100 0

Multivariate Process Signature

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 $Time (smoothed) 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600 2700 SIMCA-P+ 11 - 14.03.2011 17:53:19 20 10 0 -10 -20 614 612 615 624 635 637 622 616 639 638 628 627 631 618 634 620 630 623 617 -80 -70 -60 -50 -40 -30 -20 -10 0 t[1] 10 20 R2X[1] = 0.873728 R2X[2] = 0.0564179 Ellipse: Hotelling T2 (0.95) 30 40 50 60 70 SIMCA-P+ 11 - 10.03.2011 19:27:42 80

MVDA

Objectives for the pharmaceutical & biopharmaceutical industry

• • Increase of

process understanding

– Identification of influential process parameters – Identification of correlation pattern among the process parameters – Generation of process signatures – Relationship between process parameters and quality attributes Increase of

process control

– Efficient on-line tool for • Multivariate statistical control (MSPC) • Analysis of process variability – Enabling on-line early fault detection – Support for time resolved design space verification • real time quality assurance – Predicting quality attributes based on process data – Excellent tool for root cause, trending analysis and visualization – Fundament for Continued Process Verification (CPV)

Work and Data flow

For Method Development

Reduction of Dimensionality

Batch Level Evolution Level All Process Parameters

Individual Probes Individual Probes Recorded Process Parameter during granulation ObsID(Obs ID ($PhaseID)) Mixer Power rate of change precss variable 0.01 * Mixer torqute process variable 0.1 * Mixer speed process variable 0.1 * Product temperature process variable Mixer power process variavle (electrical) 1 0 4 3 2 14 13 12 7 6 5 9 8 11 10 320 330 340 350 360 370 Num 380 390 400 410 420

…

Aims: - Creation of batch signature Identify correlation patterns Fundament for CPV

Batch Level Evolution Level

Work and Data flow

For Routine Use in Production

PO_WST10332_EXJADE_GRAN_Steintraining - batch level (scores).M1 (PCA-X), All Batches, PS-Complementory tPS[Comp. 1]/tPS[Comp. 2] -80 -100 -120 -140 -160 -180 40 20 0 -20 -40 -60 -20 -10 S0042_A_85 S0044_A_85 0 S0040_A_85 S0027_B_85 S0010_A_85 S0023-A_85 S0022_A_85 10 20 30 40 50 tPS[1] 60 70 R2X[1] = 0.402627 R2X[2] = 0.341738 Ellipse: Hotelling T2PS (0.95) 80 90 100 110 S0028-B_85 120 SIMCA-P+ 11 - 08.02.2009 17:15:01 Aims: -Conformity check - Real time release testing - Trend analysis - Root cause analysis Identification of responsible Parameter(s) Investigation on process data PO_WST10332_EXJADE_GRAN_Steintraining.M2:7 Predicted Liquid feed pump speed +3 Std.Dev

XVar(Liquid feed pump speed) (Aligned) (Avg) -3 Std.Dev

XVarPS(Liquid feed pump speed) (Batch S0007_B_854825) 20 0 -20 -40 0 120 60 40 100 80 10 20 $Time (normalized) 30 40 SIMCA-P+ 11 - 08.02.2009 17:17:46 50

Increased of level of detail Answers: What? When? How?

What makes Multivariate-SPC so powerful?

• The SIMCA product family uses a data compression technique – Multivariate data analysis • PCA and or PLS • Data from all relevant process parameters are concentrated to a few highly informative graphs – Simplifies overview, analysis and interpretation – Enable use of data by increasing ease of use • Simple drill-down functionality to transfer compressed information back to raw data for analysis

Drill-down for analysis

Monitor

• Early fault detection – SIMCA-online technology is acknowledged for its ability to detect process issues before they become critical • Project dashboard – Full drill-down to raw data for cause analysis • Knowledge building – Instant analysis of process changes improves understanding • Process visibility – Easy-to-grasp graphics makes the process status accessible to colleagues at all levels

Prediction and Continued Process Verification

• • • • Product quality information – Indirect information based on process behavior – As long as a process behaves well, product should be according to specification Soft sensor modeling – Predict hard-to-get process properties from online process data, spectral data etc.

Predictive analytics – Online prediction of product quality and properties Continued Process Verification – Ongoing assurance is gained during routine production that the process remains in a state of control.

Motivation for QbD

• Reducing process variability is not necessarily desirable 1 0,8 0,6 0,4 0,2 0

Input

1 0,8 0,6 0,4 0,2 0

Process

1 0,8 0,6 0,4 0,2 0

Output

• • • With variation in inputs Initial material qualities Environment Equipment Static process Results in variability in outputs

15

QbD and PAT Strategies

• Control strategy b) feedforward control 1 0,8 0,6 0,4 0,2 0

Input

1 0,8 0,6 0,4 0,2 0

Process

1 0,8 0,6 0,4 0,2 0

Output

• • • Adjusting the process based on variations in the input Media and feed composition Used in pulp and paper and other industries with natural products with high variability Cheese production

16

QbD and PAT Strategies

• Control strategy c) PAT control 1 0,8 0,6 0,4 0,2 0

Input

1 0,8 0,6 0,4 0,2 0

Process

1 0,8 0,6 0,4 0,2 0

Output

• Adjusting the process based on measurement of quality in the process Used in many processing industries using various methods • • Direct measurement of material quality Inferential control – estimation of quality from process measurements • Spectral calibration

17

• Monitoring is used to

detect

and

diagnose

process deviations

Monitoring Important Process Parameter

UMETRICS CONFIDENTIAL 18

Model Predictive Control (MPC) Important Process Parameter

• MPC is used to

predict

UMETRICS CONFIDENTIAL 19

Model Predictive Control (MPC)

• MPC is used to

predict

and

optimize

the process

Important Process Parameter

UMETRICS CONFIDENTIAL 20

Model Based Control Manipulated Variables Important Process Parameter

UMETRICS CONFIDENTIAL 21

Novartis Biopharmaceutical

22

23

Chemometric portfolio

27

Thank you for your attention!

28