Bayesian optimization is a powerful statistical approach that accelerates discovery in molecules, formulations, products, packaging, and processes by efficiently guiding sequential experiments, making it ideal for costly or time-consuming studies. This presentation demonstrates the applications of JMP Bayseian Optimization to optimize product and formulation design. We showcase a complex case study with more than 100 formulation ingredients and many categories of ingredients with both category-level and ingredient-level constraints. By generating a candidate formulation set that satisfied these complex formulation constraints and then applying JMP Bayesian Optimization, we successfully identified and experimentally confirmed improved formulation candidates.
Hello. My name is Fangyi Luo. I'm a statistician and also a data scientist from Procter & Gamble. Today, Chris Gotwalt and I are going to present our poster at the JMP Discovery Summit. We're going to talk about a case study of JMP Bayesian Optimization to improve the formula with complex formulation constraints. I will turn it to Chris.
Hey, Fangyi Luo, thank you. Hello. My name is Chris Gotwalt. I'm the JMP chief data scientist. Bayesian Optimization was a very important project for us, and I want to thank Fangyi Luo both for the opportunity to speak today, and also for the collaboration that she and I had as we were developing the platform. We worked closely to ensure that the platform met both P&G's needs and then also to make sure that it enables people to do Bayesian Optimization from all different industries that are JMP customers.
This particular application that she's going to be talking about today, I heard about when it was in progress, and it was, and still is, the largest application of JMP Bayesian Optimization I've ever heard of, both in terms of the number of rows of data that were used, the number of columns, and the very complex constraints that the problem had.
When I heard that they were talking about this and as we were consulting on it, I was a little bit nervous because I didn't really know if JMP was going to be able to succeed at this task. It was really exciting to hear that the results came out as well as they did. It's a really great use case, and I'm going to hand it back over to Fangyi Luo so she can go ahead and talk about what she did. Thank you.
Thank you, Chris. I think most of you already know that Bayesian Optimization is a very powerful sequential learning approach, which will accelerate either the formulation or product discovery by using fewer experiments. This is especially very effective when your experimentation is expensive or time-consuming.
At P&G, we make many consumer products, and many of the products are formulated products. For example, shampoo, conditioner, body wash, dish soap, laundry detergent, so you name it. One time, one of our formulators, he wanted to improve the performance of a certain formulation. I'm not going to disclose the detail of the formulation because of confidentiality. He wanted to improve the performance of the formulation. He has tried another software called Aisot AI for Bayesian Optimization, and he was not able to get the results due to the limited computing capacity.
Then he approached me and wanted to try JMP Bayesian Optimization for his problem. In his case, he had collected the performance data of 640 historical formulations. All of the formulations had very complex formulation constraints. Each of the formulations, it contains up to 124 ingredients. They're from 16 different ingredient categories, and 1 of the categories actually had 3 subcategories.
There are category-level constraints as well as ingredient-level constraints. Within each category, there's constraints on the maximum number of ingredients per formula, as well as minimum and maximum level of sum of the ingredients within each ingredient category. Also, for each of the ingredients, there's constraints on the minimum and maximum level of the ingredients that can be contained in the formula. It is very complex formulation constraints.
What we did, we used JMP Bayesian Optimization. In JMP Bayesian Optimization, it uses Gaussian process model. We had three key performance endpoints. One of the endpoints we want to maximize, and two of the other endpoints we want to minimize.
In JMP Bayesian Optimization, we select the best Gaussian process model by trying different kernel functions, which is also the correlation type for modeling the responses. Then we were able to select the best model with the highest leave-one-out R-squared. You can see our models actually is good. The R-squared is about like 0.9, 0.8, or around 0.6. It's not bad.
In order for JMP to handle formulation with complex constraints in JMP Bayesian Optimization, the approach is that we need to generate a candidate set satisfying all the complex constraints, and then upload this candidate set to JMP as a candidate formulation set.
Thank you to one of my colleagues, Haichao, and she was able to use R and also parallel computing and generated over three million formulas. Then, using JMP, I found that out of those formulas, about 120,000 formulas satisfied all complex formulation constraints I mentioned earlier. I loaded this candidate set into JMP Bayesian Optimization, and then we were able to identify three best formulation candidates from this candidate set using Bayesian desirability.
Then I gave those best candidates to our formulators, and then they tested those formulas experimentally. They found that the best candidate among the three I proposed using JMP Bayesian Optimization actually is confirmed as a top overall performer, and it outperformed the current top benchmarking formula. It's a success. We were able to accelerate our innovation by approximately 6 months, and also eliminate needs for hundreds of formulation tests. As you may know that each test typically would cost about a few hundred dollars, so we were able to save time and cost.
Our case study actually showed that JMP provides a fast, credible approach to identify superior formula under very complex formulation constraints. In our case, we used R and parallel computing and also JMP to generate candidate formulation satisfying complex constraints.
Recently, we worked with Jasmine Hatcher from JMP, and she was able to develop a JMP add-in to generate candidate formulation under complex constraints. This add-in will be available in JMP Marketplace soon. Thanks, Jasmine, and Chris, and others in JMP in developing this JMP add-in.
Also, our success story supports the broad application of JMP Bayesian Optimization to many industry innovation problems. Last, I would like to thank my colleagues, Brennan Schilling and also Haichao Zhang, for close collaboration on this work. Also, thank Chris and his team for developing the JMP Bayesian Optimization feature in JMP. I really appreciate the JMP team for developing all the features and also the add-in and also collaborating with P&G on the problems. Thank you. I'll stop here and see if Chris has anything to add.
No, I just think this was a really amazing, challenging problem that you guys were able to solve. It really showcases JMP Bayesian Optimization in a great light because of the size and the problem and the complexity of the constraints. It's great that you guys were able to save so much time and resources in developing this new formulation.
It was also good to hear that JMP Bayesian Optimization has performed very well when placed in a competitive situation with other Bayesian Optimization products that are out there. It was really great to hear that we were able to come up with a solution where other products had had challenges. I think this is a great example for us, and it was really great working with you on this project and throughout our time working on Bayesian Optimization together. Thank you.
Thank you.
