The Uncountable platform excels at improving the performance of a recipe under ambitious but reasonable requirements. The canonic use is to a) efficiently explore the input space defined by ingredients and process parameters with the goal of b) increasing the performance of a product across all of its specification goals.
Projects are often combinations of more than one use case where the aggregation of data of individual projects produce results greater than the sum of the parts in terms of development speed. This, however, is too general, so a non-exhaustive list of use cases and examples are listed below.
- Cost optimization (and any other linear property of a recipe)
- Ingredient replacement
- Exploration around a benchmark or established product
- “Application change”
- “Product expansion”
- Prediction of performance before the actual implementation of an experiment.
- Bringing a recipe to Scale-up or Production stages
- Tackling several routes at the same time
- Cost optimization (and any other linear property of a recipe)
The platform is designed and optimized to produce recommendations that follow linear constraints, which cost is a well-known example of.
This means that we can reduce or keep the cost of a formulation while replacing ingredients and/or improving performance. We can do so by considering that prices -or any other attribute- depend on different locations, providers, lots, etc.
If a project falls in this category, implement the following:
- To guarantee formula recommendations within a specific range: Scroll to the bottom of the project Constraints-> Click on “+ Calculation Constraint” -> select the pertinent (previously defined) calculation -> set a reasonable value under the Maximum cell.
- To push the platform to not only guarantee a range, but also minimize the desired cost or quantity: In the project Spec scroll down -> click on “+ Add Goal” -> click on Add Calculation -> select the pertinent (previously defined) calculation -> search and select the pertinent (previously defined) calculation -> in the newly added row, select Minimize and set a reasonable Upper Limit and Goal.
Note that implementing options A and B might produce recommendations that are less varied, as opposed to setting only A.
2. Ingredient replacement
Removing specific unfavorable raw materials is common during the process of designing experiments. This could be for sustainability, cost, supply chain, toxicity or other reasons.
A common example is “green replacement” where we contemplate green raw materials only, selected either by feedstock type, total natural content, carbon emission, vegan rating etc. This workflow tends to be more exploratory since attention to this use case is fairly recent and the purpose tends to obtain ‘new and innovative’ solutions with experimental raw materials that, inherently, have not been used much before.
If a project falls in this category, it is assumed that most of the formulation will be similar and some few ingredients will be replaced. For this reason, we recommend:
- Creating a set of constraints based on an existing recipe. To do so, in the constraints page, click on the wheel button (next to “+ Blank Constraints”) -> click on “Load Recipe as Base” in order to select the recipe to explore around -> Ban and/or enforce ingredients by setting the “Use” column to “Always” or “Never”. Make sure to add the new ingredients by scrolling down and clicking on the blue “+ Add Ingredient” -> filter by Category or clicking the funnel to filter by a common attribute -> add the list of ingredients in batch and set a reasonable Minimum and Maximum for the added ingredients.
- In the Suggest Experiments page, we recommend using the option of “3. Optimize With Chosen Inputs” and the consequent option of “Select Constraints” (instead of the “Automatically Selected” ingredients option) and in the dropdown menu, select the Constraints you defined. Click on “Advance Settings” and make sure that “Suggest Additional Ingredients” is unchecked. If the list of ingredients is large, consider checking the box “Adjust Ingredient Ranges and Probabilities.”
3. Exploration around a benchmark or established product
It is common to ask how a recipe would perform by changing it slightly, maybe just replacing a couple of ingredients or a group of ingredients.
Usually, constraints tend to be tight and inflexible in this situation given that most of the ingredients remain fixed (set to Always use, with same min. and max.). For this reason, it may be the case that the Learn About Inputs option provides the desired diversity in recipes as opposed to optimizing given that asking a model to produce suggestions near its maxima/minima will often return very similar recipes.
We recommend testing the methodology described in 2), and if recommendations look very similar, then try using “1. Learn About Inputs” instead of “3. Optimize With Chosen Inputs”.
4. “Application change”
When an existing product can be spin for a similar application, we can use the platform to take advantage of and learn from background data that is alike (similar set of ingredients and overlapping set of properties, but different goals).
For example:
- fulfilling different regulatory requirements
- meeting regional preferences
5. “Product expansion”
We can think of product expansion in a similar way to the application change case, where most success metrics have a lot of background data and we just want to extend the specifications with some new types of output -maybe exclusive to a project/product/lab-.
When a new property and a commonly measured one that is already available in the platform are correlated, the analysis and visualization tools make it trivial to find the relationship and use the existing one as a “proxy output” for the new one. A frequent situation is utilizing values that correspond to the same property but that have been measured in different units or grading systems, which can later be transformed into a standardized output.
6. Prediction of performance before the actual implementation of an experiment.
The platform provides the ability of creating state-of-the-art Machine Learning models carefully selected for DOE, as escribed in (1) – (7) which can then be used to generalize for and predict the performance of untested recipes.
When the accuracy of a model is high enough, or the tested recipe is close to the current data set, it is recommended to use the prediction tools to save resources in the lab.
Predictions should be considered only in the context of a population by comparing relative performance across suggestions and/or with respect to predictions of existing recipes used in the training set. Please refer to (27) – (30) for additional explanation, nuances and instructions on how to obtain predictions.
Similarly, direct predictions from other sources can be implemented in the platform in the form of dynamic calculations. Examples of this are:
- when a commonly existing output is found to be highly correlated to a new output
- when there exists supporting theory to predict an output based on calculations that take in process parameters and/or ingredients along their quantities and attributes
7. Bringing a recipe to Scale-up or Production stages
After a recipe performs as desired in the laboratory, it may undergo additional changes as it proceeds to scale-up and, consequently, production stages. This adds complexity as this usually brings changes in batch size, environment, precision of measurements, etc. often causing that the performance in the lab is no longer replicated.
The platform is suited to seamlessly incorporate new process parameters and outputs that correspond to these later stages in order to tune the recipe. This use case can be thought of as an extension of the ones listed above; there is no different paradigm in the methods used as the Process Parameters are treated as yet another type of inputs with the difference that they can be categorical and are not part of some constraints, such as the 100% total when summing ingredients.
8. Tackling several routes at the same time
Some products require R&D on more than one front at a single time. It may be that a recipe requires an ingredient to be replaced immediately, and that it becomes greener, but not necessarily at the same time or with the same deadline.
The way the platform allows to encode different sets of both specifications and input constraints for a single project permits that iteration of rounds can be performed in parallel. It is possible to make a copy of a Constraint/Spec set by clicking on the green arrow next to “Overwrite”, clicking “Save As” and naming the copy. To load a different set, click on “Load Constraints/Spec”.