Artificial intelligence model for the prediction of cleansing foam formulations with excellent make-up removability ～Is an “in silico formulator” superior to a human formulator?～
Podium 72
Presented by: Masugu Hamaguchi
Introduction
Cleansing foam is used to wash excess sebum and dirt from skin and make-up remover is used to remove make-up cosmetics. Recently, from the viewpoint of shortening time and eco-friendliness, there is an increasing need for a single product to serve as cleansing foam and make-up remover with one product.
Solvent-based cleansing agents such as make-up remover oil exhibit excellent make-up removal performance, because oil is the main component. However, problems are associated with solvent-based cleansing agents such as high environmental loads and material costs, in addition to the feeling of residual oiliness after rinsing. Conversely, surfactant-based cleansing agents such as cleansing foams have excellent rinsing properties but weak oil removability, because they are mainly water-based. In this study, the latter approach was adopted to improve the cleansing performance of cleansing foams.
Notably, cleansing foams are composed of many ingredients, making formulation optimization difficult to find the best formulation from an infinite number of combinations. Therefore, artificial intelligence (AI) using machine learning was introduced into the formulation design to construct a cleansing capability prediction system that considers the effects of surfactant self-assembly and chemical characteristics of components.
Methods
Samples: Cleansing foam samples (> 500 formulations), consisting of ionic surfactants, nonionic surfactants, polyols, and water, were prepared by mixing thoroughly while heating and stirring.
Evaluation of cleansing capability: The test samples were prepared by applying a waterproof eyeliner pencil on a piece of white artificial leather. The cleansing foam was added dropwise onto the sample, rubbed a certain number of times, and then allowed to dry after rinsing. The cleansing capability was evaluated using the eyeliner-pencil residual ratio with a colorimeter.
Data Preprocessing: To predict the cleansing capability, we utilized molecular descriptors, which are a set of numeric representation of a molecule calculated by chemical structure, such as molecular weight and the number of esters in a molecule. We gathered 2,144 descriptors per ingredient. The cleansing performance of each formulation was estimated using the weighted average (in terms of mol% or wt%) of the descriptors of each ingredient. Important descriptors were selected to reduce a noise from irrelevant descriptors.
Artificial Intelligence (AI) prediction: We applied several predictive AI models (two tree-based and three linear-based models) were applied, and their hyperparameters were optimized. The prediction performance was evaluated based on 10-fold cross validation with the indices of validated R2, which is a statistical measure that represents the proportion of the variance for a dependent variable that is explained by independent variables.
Results and Discussion
An AI model was established to predict cleansing capability. The best prediction accuracy was obtained with R2=0.765 (that of conventional multiple linear regression was below zero). The prediction accuracy increased significantly with the use of descriptors. Overall, for the calculation of weighted average, mol% was suitable for tree-based models, whereas wt% was suitable for linear-based models. The mol% of the weighted average is considered potentially more accurate based on stoichiometry, since water constitutes >97 mol% on average. Therefore, the influence of water is more dominant. Linear-based models are more affected by this effect, whereas tree-based models are not. In the prediction of cleansing capability, nonlinear behavior should be considered due to the interplay between surfactants and water molecules, and their self-assembly. Tree-based models are usually more suitable for non-linear prediction; therefore, their prediction accuracies are higher than those of linear-based models.
Applying this prediction method to formulation development, a high cleansing capability of >89 % for waterproof eyeliner removal was obtained. Notably, the combination of eicosaglycerol hexacaprylate and specific polyols contributed to the high cleansing capability.
Conclusion
An accuracy of R2=0.765 was obtained during cleansing performance predictions. Mixtures of cosmetic ingredients showed interactions, and non-linear behavior made it more difficult for formulators to predict their performance. However, high accuracy was obtained by incorporating chemical characteristics with descriptors. This AI prediction model based on the molecular structure of the ingredients and surfactant self-assembly showed higher accuracy and was better than conventional approaches such as multiple linear regression. This system may contribute to a significant reduction in the effort required for cosmetics development. Lastly, cleansing foams with high cleansing capability and foamability were achieved.
Cleansing foam is used to wash excess sebum and dirt from skin and make-up remover is used to remove make-up cosmetics. Recently, from the viewpoint of shortening time and eco-friendliness, there is an increasing need for a single product to serve as cleansing foam and make-up remover with one product.
Solvent-based cleansing agents such as make-up remover oil exhibit excellent make-up removal performance, because oil is the main component. However, problems are associated with solvent-based cleansing agents such as high environmental loads and material costs, in addition to the feeling of residual oiliness after rinsing. Conversely, surfactant-based cleansing agents such as cleansing foams have excellent rinsing properties but weak oil removability, because they are mainly water-based. In this study, the latter approach was adopted to improve the cleansing performance of cleansing foams.
Notably, cleansing foams are composed of many ingredients, making formulation optimization difficult to find the best formulation from an infinite number of combinations. Therefore, artificial intelligence (AI) using machine learning was introduced into the formulation design to construct a cleansing capability prediction system that considers the effects of surfactant self-assembly and chemical characteristics of components.
Methods
Samples: Cleansing foam samples (> 500 formulations), consisting of ionic surfactants, nonionic surfactants, polyols, and water, were prepared by mixing thoroughly while heating and stirring.
Evaluation of cleansing capability: The test samples were prepared by applying a waterproof eyeliner pencil on a piece of white artificial leather. The cleansing foam was added dropwise onto the sample, rubbed a certain number of times, and then allowed to dry after rinsing. The cleansing capability was evaluated using the eyeliner-pencil residual ratio with a colorimeter.
Data Preprocessing: To predict the cleansing capability, we utilized molecular descriptors, which are a set of numeric representation of a molecule calculated by chemical structure, such as molecular weight and the number of esters in a molecule. We gathered 2,144 descriptors per ingredient. The cleansing performance of each formulation was estimated using the weighted average (in terms of mol% or wt%) of the descriptors of each ingredient. Important descriptors were selected to reduce a noise from irrelevant descriptors.
Artificial Intelligence (AI) prediction: We applied several predictive AI models (two tree-based and three linear-based models) were applied, and their hyperparameters were optimized. The prediction performance was evaluated based on 10-fold cross validation with the indices of validated R2, which is a statistical measure that represents the proportion of the variance for a dependent variable that is explained by independent variables.
Results and Discussion
An AI model was established to predict cleansing capability. The best prediction accuracy was obtained with R2=0.765 (that of conventional multiple linear regression was below zero). The prediction accuracy increased significantly with the use of descriptors. Overall, for the calculation of weighted average, mol% was suitable for tree-based models, whereas wt% was suitable for linear-based models. The mol% of the weighted average is considered potentially more accurate based on stoichiometry, since water constitutes >97 mol% on average. Therefore, the influence of water is more dominant. Linear-based models are more affected by this effect, whereas tree-based models are not. In the prediction of cleansing capability, nonlinear behavior should be considered due to the interplay between surfactants and water molecules, and their self-assembly. Tree-based models are usually more suitable for non-linear prediction; therefore, their prediction accuracies are higher than those of linear-based models.
Applying this prediction method to formulation development, a high cleansing capability of >89 % for waterproof eyeliner removal was obtained. Notably, the combination of eicosaglycerol hexacaprylate and specific polyols contributed to the high cleansing capability.
Conclusion
An accuracy of R2=0.765 was obtained during cleansing performance predictions. Mixtures of cosmetic ingredients showed interactions, and non-linear behavior made it more difficult for formulators to predict their performance. However, high accuracy was obtained by incorporating chemical characteristics with descriptors. This AI prediction model based on the molecular structure of the ingredients and surfactant self-assembly showed higher accuracy and was better than conventional approaches such as multiple linear regression. This system may contribute to a significant reduction in the effort required for cosmetics development. Lastly, cleansing foams with high cleansing capability and foamability were achieved.