How Predict Works

The tensile strength of cast iron is determined by its cooling rate as it solidifies its nucleation, and its chemistry.

In controlling the results of tensile testing, the cooling rate is removed from consider­ation by using standard test specimens. Most foundries negate the effect of nucleation by following the good practice of keeping a consistent inoculation practice. Therefore, controlling the chemistry should be the key to controlling tensile strength.

Foundrymen have known the effect of the chemistry for years. The carbon and silicon have such a strong effect that a great deal of effort was placed on controlling them. The problem has been trying to deter­mine the effects of all of the minor changes that are always occurring. The chromium goes up making the iron stronger while the titanium goes up as well making the it weaker - what's the end result? To make matters even worse, if the carbon and silicon change a little at the same time, out guessing the strength can almost be impossible.

The computer's ability to remember and rapidly deal with numerous variables is a very logical answer to this problem.  PREDICT starts by using a multiple regression analysis pertaining specifically to the operation and grade of iron in question. It provides the computer with the analog for handling all the subtle changes that occur during the operation.

A natural question is, why should a multiple regression analysis be performed when that information is already available in the literature?  The answer is that the published information just doesn't perform well enough.  There are many possible reasons for the lack of correlation with that data.  The most obvious simply being a bias in the opera­tion spectrometer as compared to the chemical analysis tools used to develop the information in the literature. 

It should be noted that in some cases, the multiple regression analysis provides results that are contradictory to what we as metallurgists believe we know.  Such contradic­tions can be explained by considering those elements which are not tested.  As an example, a foundry may purchase steel scrap that has lower chromium than normal but has a higher nitrogen content than normal.  The nitrogen isn't tested for but the chromium is.  Since the increase in nitrogen increases the strength more substantially than the reduced chromium lowers the tensile strength, the analysis may indicate reducing chromium increases the strength.  While that isn't what happens, as long as the relationship in the purchased scrap is maintained, the answer comes out right.

Of course, once the relationship is developed between the chemistry and tensile strength, it's not difficult to make mathematical manipulations so that an equation yields the necessary amount of an element to achieve a desired tensile strength.  From there, knowing the amount of iron being treated and compositions of the alloys being used, it's an easy step for the computer to develop the amount of the alloys that need to be added.

Target Adjustment

The above was the thinking behind the original versions of the program.  However, we found that it wasn’t good enough.  We found, for unknown reasons, the results would drift from the answers predicted.  In order to compensate for that the program tracks the difference between actual and prediction.  It averages that difference for the last ten readings and adjusts the tensile target accordingly.  As an example, if the average of the last ten tests showed the actuals were 500 psi over the predicted, the additions would be calculated towards a target that was 500 psi lower.

Compensation for Recoveries

Since we don't have the final chemistries until it's too late, we must make our predic­tions from base iron samples.  Again, it’s a very logical starting point and was how the program started.  As we began to analyze the results, we noticed, at times, there was a definite bias between the results from base samples as compared to final samples.  While some elements show the expected loss from being in the liquid state longer, others, unexpectedly, showed increases even though nothing was added.  (It's our supposition that the inoculation process may be changing the sample in some way causing the spectrometer to read slightly different.)

No matter what the cause, we found our results improved if we would take into account the expected recovery of all of the elements.  PREDICT now calculates the recoveries of all elements as part of the process.  When making the calculations of additions from the base samples, it averages the recoveries from the last ten recorded results and adjusts the base readings accordingly.

Step by Step

Another way of looking at how PREDICT works is to look at it step-by-step as it goes through the process.

1.         The user enters the base data.

2.         PREDICT looks up the target desired and adjusts it on the basis of the last ten tests difference between actual and predicted.

3.         The program then calculates the recoveries for all of the elements from the last ten tests.

4.        The program, using the calculated recoveries, results from the multiple regression analysis, and base iron information, calculates the alloy additions necessary to achieve the adjusted target.

5.        The program then checks all of this information to determine if any of the entries or any calculated results are statistically unusual.

6.         If a test is to be taken that will be entered later, the user then enters what additions will actually be made and the program stores the data in a file.

7.         Once the tensile results have been obtained, the user will enter the appropriate information into that part of the program.

8.         PREDICT re-calculates the expected tensile properties from the final chemistries results and makes a number of comparison calculations -- tensile/hardness ratio, comparison of laboratory hardness reading to foundry's, ultimate strength/yield strength ratio, etc.

9.         It once again compares all of the entries and calculated results to previ­ous entries to ascertain if anything is statistically unusual.

10.       It stores all of the information in a format that is readable by Excel and most other spreadsheet programs.