MI(Materials Integration) Research with MIRAI

There are so many industries that can’t apply ordinal MI(Materials Informatics) to their job.
Here is one example listed in patent.

Beta column is Sensory Evaluation (feel very good=5, feel very bad=1). A, B, and C are the ingredients of the formula. Each ingredient is sold as a commercial product, and the exact contents are not known. You can imagine something like Carbon Black. CB(A-1) is bought from W-company, A-2 is from X, A-3 is from Y.

All we know is how much of each ingredient was used to create the formula and how many sensory evaluation values were obtained.
In fact, there are multiple types of sensory evaluation.

What we want to do is to use MI to design formulations that will result in multiple sensory evaluations, all of which will result in a 5-point rating.

But we can neither use the RDKit identifier nor the results of molecular orbital calculations for this case.
It is not easy to increase the data since hundreds of people are gathered for sensory evaluation.

What can MI researchers do about this problem?

If you are confident in your skills, create a prediction equation with the following data and try to predict the value of the predicted formulations.

CompoundBetaA-1A-2A-3A-4A-5A-6A-7B-1B-2B-3B-4B-5C
E6500001500000201049
E1250001200430000049
E1500150000200001049
CE1215000000200001049
CE1210001500000300049
E4500150000000201049
CE821400100030000049
CE5401500000000201049
CE1040001500060000019
E94000150000.1000078.9
CE142000150000000079
E850002500030000039
E1050001500050000029
E115000900630000049
CE1120001500003000049
CE1320001500000003049
CE6200000150000201049
CE2401500000200001049
CE950003000030000034
CE720001500000001069
E750001500030000049
E5500015000000201049
Predict
CE4215000000000201049
E2500015000200001049
E1350001400230000049
E3500001500200001049
CE3200000150200001049

Pirika has new analytical tool, MIRAI(Multiple Index Regression for AI).

This time, I have made this tool into a web application.

A typical multiple regression analysis would look like this.

One disadvantage of the multiple regression method is that it cannot be used for those with interactions between explanatory variables or those that are nonlinear.

In contrast, MIRAI is a feed-forward type neural network method, designed to maintain predictive performance even with such a small number of data.

beta= C0 + Σ[(Ca*Xa+Cb*Xb・・Ci*Xi+1)^α
  *(Cm*Xm+・・・ +Cq*Xq+1)^γ 
  *(・・・    +Cz*Xz+1)^ω ]

In this analysis, we have constructed an equation that reproduces the value of beta by multiplying the power function and estimate the results of the formulations that were not included in the training.

Once you have constructed the four different estimating formulas, you can then have the computer look for a formula that will give you an all-five rating.

I will put this MIRAI into GROVE.

Many examples are available at https://www.pirika.com/Education/JP/MAGICIAN/Examples/Formulation-Top.html