Project Description
By including all of the muffin formulas together in one AI food model, ProSensus not only determined which were the key ingredients but also achieved a common platform for designing and evaluating new formulas.
The next time PepsiCo Foods wanted to reformulate muffins (to different design criteria), the AI model helped them get there in just two iterations.
Project Details
The Challenge
Our client wanted to achieve a reduction in a key quality property called AOI, without introducing new ingredients, and while maintaining the sensory properties of the existing product as much as possible. AOI was a poorly understood property for which a first-principles model was unavailable.
The Results
Four different products were targeted, and dramatic AOI reductions ranging from 47%-55% were achieved. Maintenance of the sensory properties was challenging because they had not been quantified, but the ProSensus approach can accommodate less-than-perfect data sets.
Our Approach
Assess Available Data
Ideally, recipe data and quantified outcomes are available for several variations on a product, or a family of products. Alternatively, the ProSensus approach can generate highly efficient designed experiments even with no prior data.
For the muffin batters, AOI had been quantified for all muffin formulas currently in production, and these were paired with the recipe data and label-stated baking parameters to build an initial PLS model.
Interpret the Data with AI-Driven Models
The following collection of complementary FormuSense plots demonstrate the power of using AI for this foods formulation challenge.
The model distilled over 60 variables down to a few latent (“underlying”) variables, with a reasonable quality of fit for the AOI (“Key Property 1”), as evident on the observed vs predicted plot (top right).
The score plot (top left) shows the distribution of muffin formulations in the first two latent variables. Products located near each other on the score plot are similar in both raw material and formulation properties (X) as well as in quality attributes (Key Property).
The W*C plot (bottom left) identifies the correlations between the X (orange) and Y (blue) variables in this dataset. Variables clustered together (such as Key Properties 1 & 2, ASN, AA, Leaven and Sugar) are positively correlated with one another, while variables on the opposite side of the plot (such as Grains & Flours, Salt and Color) are negatively correlated with the first cluster.
Note that muffin formulations that have low values for Key Property 1 are located on the far left of the score plot. This is an expected result, recalling that Key Property 1 is located on the opposite side of the W*C plot.
The coefficient plot (bottom right) indicates muffin batters with high values of ASN and Sugar are correlated to higher values of Key Property 1. Since PepsiCo wanted to reduce Key Property 1, ASN and Sugar should therefore be reduced.
As illustrated, intuitive plots associated with the AI model, provide an improved understanding of the underlying food chemistry and correlations and also forms the basis for discovering new products or re-formulating existing products.
Build Focused Models
While the initial model above was useful for understanding some overall correlations in the dataset, a focused model that uses individual ingredient ratios as input variables (as opposed to ingredient class ratios) can also be valuable in some situations such as when there is no desire to consider new unused ingredients in future formulations.
The W*C plot and coefficient plot below illustrate how individual ingredient information can be intuitively displayed for quick interpretation. In this particular model, which was focused on one type of muffin batter product, process conditions and water content have the highest correlation to Key Property 1. Interestingly, the coefficient plot also indicates that some ingredients from within the same ingredient class have opposite effects on Key Property 1. For example, Batter Thickener 1 is correlated to lower values of Key Property 1, while Batter Thickener 2 is correlated to higher values of Key Property 1.
Augment Models with Designed Experiments
When needed, ProSensus aims to augment the model with a few carefully selected runs. Design of experiments is a tool for making data balanced and representative [2], and this goal remains the same in latent variable spaces.
Specific runs are chosen to fill in any gaps in the latent variable space , extend the model to cover new regions or focus in on areas of interest. In this foods case study, the initial AI model was augmented by designing a 2-level factorial in the first three latent variables.
What’s Wrong with Traditional DOE’s?
Traditional use of factorial and/or mixture design of experiments can be problematic in product development. The number of experiments can quickly become intractable, especially if we want to explore a wide range of possible raw materials.
Formulating an optimal experiment in the latent variable space is an efficient alternative. Since the latent variables are an orthogonal representation of the dominant directions in the original variables, we can typically explore the important regions with far fewer experiments.
Optimization
ProSensus uses optimization techniques to invert the PLS model; this inversion selects raw materials and process conditions that best meet the project goals while meeting relevant constraints.
In this case study, optimization was used to find recipes that minimized the attribute of interest (AOI), met nutritional constraints, and achieved the required cost targets.
Multivariate constraints helped to ensure that sensory attributes would be in-line with the original muffin batters. Optimization produced several recipe alternatives for each of the four target products; these were baked and evaluated in the laboratory.
Completing the Cycle
The new data generated by the optimization step was used to update the model. This step completes the model-building process, and ensures that the model contains and represents all knowledge gained during the project.
When the next reformulation is undertaken, this model is used as the starting point for determining a recommended formulation straight off or a few strategic experimental runs, depending on the project goals.
Consider New Raw Materials
The inclusion of raw material characterizations greatly extends the capabilities of the ProSensus approach to product development. PLS models built with this data express the final product properties in terms of the raw material properties.
The optimization can therefore select raw materials that have never been tried before, based on their property values [3].
References
- E. Nichols, Latent Variable Methods: Case Studies in the Food Industry, Hamilton: McMaster University, 2011.
- S. Wold, M. Josefson, J. Gottfries and A. Linusson, “The utility of multivariate design in PLS modeling,” Journal of Chemometrics, pp. 156-165, 2004.
- K. Muteki, J. F. MacGregor and T. Ueda, “Rapid Development of New Polymer Blends: The Optimal Selection of Materials and Blend Ratios,” Industrial & Engineering Chemistry Research, 45, 4653-4660, 2006.
