Umetrics Suite Blog

What-If tool helps you identify and predict parameters that affect quality

December 7, 2017

In manufacturing and other industries that have complex processes, knowing which variables have the most impact on quality and at what point, or knowing which combination of variables to change in order to improve your process, can have a huge impact on the overall quality or profitability of your manufacturing process. But without making expensive and time-consuming changes in the physical processes in order to test all possible scenarios, how can you identify and predict the variables that have the most significant impact on your outputs?

SIMCA What IF predicts changes in batch process outputs

The What-If tool in SIMCA uses OPLS to predict how changes in input variables (such as time or ph) will affect process outputs, such as the amount of ethanol in alcohol.

One way is to use a data analytics tool that simulates possible outcomes when changing critical process parameter settings. The What-If tool in SIMCA allows you to systematically identify and predict how changes in various input variables will affect your outputs. With it, you can answer questions such as:

  • What if I change the critical process parameters to a slightly higher value?
  • What if I could find an alternative variable giving the same rise in quality and which is easier to manipulate?
  • What if the pH probe starts failing during batch progress?
  • What if another set-point exists with a higher throughput and which would stay within the borders of the existing model/design space?

What-If tool simulates changes

The What-If tool in SIMCA allows you to graphically increase or decrease the values of selected variables and observe the effects within prediction plots based on factors of your active model. The exact functionality depends on whether you have a continuous process model, a Batch Evolution model (BEM) or a Batch Level Model (BLM), but all three can be examined.

SIMCA What-if tool impose changes to critical process factorsUsing the SIMCA What-If tool, you can impose changes to critical process factors and then graphically visualize the consequences of increasing or decreasing the selected variables on score plots and DMoDx plots.

In the What-If tool, the prediction set of the currently defined model is duplicated and a simulation is created.

You can very easily and transparently change the values of the various selected critical process factors using sliders. The original prediction set items are displayed in one color and the What-If prediction set in another.

Using the What-If tool in SIMCA

To turn on the What-If tool, you go to the Predict tab, and click on the What-If button. The What-If pane will open up. Inside the pane, you can selectively check or activate the variables you would like to use, and for each one, you will have a slider to control the upper and lower values.

Continous process example simca what if

A continuous process example: mineral sorting

To start, we can look at a continuous process example, using a dataset from a mineral sorting plant. In this process, grinders divide raw iron ore into finer materials. The material is sorted and concentrated by magnetics separators. The concentrated material is then divided into two product streams:

- PAR (that goes to a pelletization process)

- FAR (fines, which are sold as is)

In this example, there are 12 input variables (X-block in illustration below), of which 3 can be manipulated using design of experiments (DOE), and the other 9 reflect intermediate process variables. And there are 6 output variables (Y-block in the illustration below).

use design of experiments to manupuliate three of 12 input variables and  9 process variables with OPLS

We use regression modeling to determine if there are overlaps between the process inputs (X) and process outputs (Y), and whether the process outputs can be predicted by future settings of the process inputs. More specifically, for this example, we’re using OPLS to accomplish the regression modeling.

When we use OPLS and link these two blocks together, we obtain a model with four latent variables: with these four components we are able to model and predict around 75% of the variance in the six process outputs. The first two components dominate the OPLS model and that is why we will concentrate on them in this example.

If you want to know more about the OPLS tool, watch this webinar.

Presenting results with score and loading plots


On the score plot above (above, left), you can see a grouped or semi-grouped nature of the process time points. In the loading plot (above, right), you can see the process inputs (green) and outputs (blue). The red stars show the critical process factors, which can be controlled. These are:

  1. Total feed (tonnage in, Ton_In)
  2. Speed of the first magnetic separator (HS_1)
  3. Speed of the second magnetic separator (HS_2)

While there are a number of process outputs, the two we are concerned with here have to do with the quality of the final product, which relates to the amount of iron in the product (FAR) stream, and the amount of phosphorus in the product (FAR) stream.

Basic working mechanism of What-If

Using the iron-ore example, we’ll first get familiar with the basic working mechanism of What-If. We can see from the loading plot above that tonnage-in (feed) has a strong influence in the first component of the OPLS models. As a consequence, when we increase the tonnage in (feed), we move from left to right on the score plot, and the reverse if we reduce the tonnage in.


Looking at the second dimension of this OPLS model, the speed of the magnetic separator has an inversed influence in regard to the positive direction of the latent variable of the second component. That means as you increase the numerical value (the speed of the magnetic separator), you move downwards in the score plot and the other way around. If you drag the second magnetic separator to higher values, (higher speeds) you are moving toward lower values on the score plot.


From this, you can see how changing a process factor that has a strong influence on the first latent variable of the model forces a migration onto the process behavior in the horizontal direction. And if you are working with a process variable that has a strong influence in the second dimension of the model, then you are going to have changes in the position of the observations in vertical dimension.

What-If to simulate desirable outcomes for the mineral sorting plant


A joint interpretation of the scores and loadings (diagram above) indicates that an optimal combination of high throughput and quality arises in the fourth quadrant. FAR and PAR are throughput parameters that we want to maximize, a fact which points to the right-hand part of the plots. The percentage of iron in the FAR and the percentage of phosphorus in the FAR are quality parameters and have strong loadings into the second dimension. You want as high concentration of iron in the final product as possible and as little phosphorus as possible, a fact which points to the lower part of the plots.

Using the What-If tool, it appears possible to move closer to ideal position by increasing the Ton in and HS_2 and not adjusting the HS_1. This is interesting and encouraging and can now be tested in the production area.


Simultaneous changes in Ton_IN (increase), HS_1 (same) and HS_2 (increase)

Applying What-If to batch processes

You can use the same approach for batch process. The objective is to establish a model for process evolution based on well-performing and behaving batches. Batch data has a time dependency. We stack the data for each batch underneath each other and calculate a batch evolution model (BEM). To give the data a direction, we use a Y variable that can be time or some progress indicator a k a ‘maturity’ variable. We then use OPLS to combine the information overlapping in the process inputs, the batch evolution measurements and process time.

Read more about basics of batch process data analytics 

Creating a batch evolution model with OPLS

A batch process example

Below is an example of baker’s yeast production in which we have 33 batches represented by 9 process variables and 5 batch conditions. (Batch conditions include: X parameter Innoculum and Y parameters QP1, QP2, amount and yield).

We have the set point for temperature and the set point for air. Using What-If we can see what happens if we start to change these two variables. Here is the result from applying What-If to the temperature set point at 12 hours.


Increase in temperature set point at maturity (= 12 hours) is predicted to move the trajectory closer to target trajectory.

The question is can we get this batch closer to the ideal trajectory (the green line)? (The red lines indicated the upper and lower limits (given by +/- 3 standard deviation settings) and the batch that is currently being charted is blue.)

We can see that further on into the lifetime of the batch we are still relatively close to the lower border (control limit) particularly in the time space of 50-55 hours. So following the indications, we have identified that an interesting test is around 37 hours and other around 55 hours. We made a complimentary decrease in the air set point at 37 hours and we see it is predicted to move the trajectory even closer to the target.


Here the What-If is applied to Air as well as the Temperature.


The What-If feature is a simulation based on the active model. It lets you graphically increase or decrease the values of selected variables and observe the effects on prediction plots. Sometimes in process modeling it’s not so easy to acquire additional observations, and you may want to do some range finding simulations in order to determine the best way to operate the process in the future. It is in this setting that the What-If technology becomes most interesting.

Want to know more?

Want to try the What-If tool for yourself? Download a 30-day free trial version of SIMCA

SIMCA free trial


Topics: Data Visualization, Manufacturing Quality Control, Process Validation

Ing-Marie Lindström

Written by Ing-Marie Lindström

Product Manager at Sartorius Stedim Data Analytics