AI-driven formulator skill augmentation
～ Co-formulation of Human and AI for cosmetic development ～
Podium 74
Presented by: Toshihiro Arai
Introduction
In the cosmetics industry, quickly delivering new products for the constantly diversifying and evolving market is essential. To quickly develop products that meet these needs, formulators are required to improve their ability to design novel formulations effectively. Presently, the only way to achieve this is to repeat hypothesis construction and verification by creating prototypes. However, it takes much time to create prototypes so effective training of formulators is crucial.
Recently, the applications of AI are advancing and the improvement of human ability by AI is attracting attention. In the world of chess, players are starting to use an AI for learning. Players can play multiple games against AI in a short period of time, and can quickly study optimal strategies and get inspiration throughout the games.
Therefore, we considered how to build an AI that can be trained using cosmetics information to accelerate the development of the formulator. To achieve this, we designed an AI that predicts the quality of the finished cosmetics and provides appropriate information for formulators. First, we acquired data that represents the tactile sensation of cosmetics, which is crucial for cosmetic’s quality. Then we developed prediction models that connect the tactile sensation data and formulation data using a machine learning algorithm, and designed an AI system that uses these prediction models.
Methods
To obtain training data for machine learning, a wide variety of data related to 148 in-house marketed skincare products (toner, milky lotion, cream, gel, oil) were used as candidate features. To develop a prediction model, a large amount of tactile sensation data that can be compared between multiple cosmetic formulations is needed. Data on the strength of 18 commonly used tactile sensation descriptors was obtained using the Check All that Apply (CATA) Method, which does not require training and can be done in a short time. The conventional methods (Quantitative Descriptive Analysis, etc.) are considered unsuitable because they take time and require training. To verify the validity of the data, we conducted correspondence analysis. To construct machine learning models that predict the tactile sensation, we used the XG-Boost algorithm. For accuracy confirmation, 5-fold cross-validation was used. An AI system interface was then built using the models. To check if it is helpful for skill development, the AI was tested by less-experienced formulators.
Results
For a training tool, the prediction accuracy is essential. We addressed the issue by efficiently collecting large amounts of tactile sensation data. Based on the correspondence analysis, the data is thought to appropriately reflect the tactile differences of each cosmetic product. Next, some features (ingredients, methods) and physical property values (viscosity, thermal conductivity, etc.) that were thought to affect the tactile sensation were extracted according to the correlation coefficient with the tactile sensation. By applying the extracted features to XG-Boost, we successfully created machine learning models that predict the strength of each tactile sensation. From the 5-fold cross-validation, each model was shown to have sufficient prediction accuracy. In addition, from each model, we found several novel features that affect the strength of each tactile sensation. During the test of AI system, formulators were able to generate a prototype in fewer steps using ingredients they would not have previously considered, thereby expanding their abilities.
Discussion/Conclusion
Our research reveals that the AI is helpful for formulators to expand their abilities through numerous hypothesis construction and verification. Information provided by the AI is also useful to inspire them. Therefore, by utilizing the AI as data-driven guidance, even knowledgeable experienced formulators should be able to augment their creativity. The AI is currently being used for developing a wide range of products and is in the process of being expanded to other product types by adding data. In the future, we aim at widening the application of the AI to other aspects of cosmetics development as well, such as in the DIY cosmetics setting to assist customers in designing their own products. By expanding the limits of human skills with AI, cosmetic chemists should be able to develop innovative and revolutionary formulations and give cosmetics possibilities of new functions and uses.
In the cosmetics industry, quickly delivering new products for the constantly diversifying and evolving market is essential. To quickly develop products that meet these needs, formulators are required to improve their ability to design novel formulations effectively. Presently, the only way to achieve this is to repeat hypothesis construction and verification by creating prototypes. However, it takes much time to create prototypes so effective training of formulators is crucial.
Recently, the applications of AI are advancing and the improvement of human ability by AI is attracting attention. In the world of chess, players are starting to use an AI for learning. Players can play multiple games against AI in a short period of time, and can quickly study optimal strategies and get inspiration throughout the games.
Therefore, we considered how to build an AI that can be trained using cosmetics information to accelerate the development of the formulator. To achieve this, we designed an AI that predicts the quality of the finished cosmetics and provides appropriate information for formulators. First, we acquired data that represents the tactile sensation of cosmetics, which is crucial for cosmetic’s quality. Then we developed prediction models that connect the tactile sensation data and formulation data using a machine learning algorithm, and designed an AI system that uses these prediction models.
Methods
To obtain training data for machine learning, a wide variety of data related to 148 in-house marketed skincare products (toner, milky lotion, cream, gel, oil) were used as candidate features. To develop a prediction model, a large amount of tactile sensation data that can be compared between multiple cosmetic formulations is needed. Data on the strength of 18 commonly used tactile sensation descriptors was obtained using the Check All that Apply (CATA) Method, which does not require training and can be done in a short time. The conventional methods (Quantitative Descriptive Analysis, etc.) are considered unsuitable because they take time and require training. To verify the validity of the data, we conducted correspondence analysis. To construct machine learning models that predict the tactile sensation, we used the XG-Boost algorithm. For accuracy confirmation, 5-fold cross-validation was used. An AI system interface was then built using the models. To check if it is helpful for skill development, the AI was tested by less-experienced formulators.
Results
For a training tool, the prediction accuracy is essential. We addressed the issue by efficiently collecting large amounts of tactile sensation data. Based on the correspondence analysis, the data is thought to appropriately reflect the tactile differences of each cosmetic product. Next, some features (ingredients, methods) and physical property values (viscosity, thermal conductivity, etc.) that were thought to affect the tactile sensation were extracted according to the correlation coefficient with the tactile sensation. By applying the extracted features to XG-Boost, we successfully created machine learning models that predict the strength of each tactile sensation. From the 5-fold cross-validation, each model was shown to have sufficient prediction accuracy. In addition, from each model, we found several novel features that affect the strength of each tactile sensation. During the test of AI system, formulators were able to generate a prototype in fewer steps using ingredients they would not have previously considered, thereby expanding their abilities.
Discussion/Conclusion
Our research reveals that the AI is helpful for formulators to expand their abilities through numerous hypothesis construction and verification. Information provided by the AI is also useful to inspire them. Therefore, by utilizing the AI as data-driven guidance, even knowledgeable experienced formulators should be able to augment their creativity. The AI is currently being used for developing a wide range of products and is in the process of being expanded to other product types by adding data. In the future, we aim at widening the application of the AI to other aspects of cosmetics development as well, such as in the DIY cosmetics setting to assist customers in designing their own products. By expanding the limits of human skills with AI, cosmetic chemists should be able to develop innovative and revolutionary formulations and give cosmetics possibilities of new functions and uses.