Establishing predictive model for the spreadability of cosmetic formulations by Large Amplitude Oscillatory Shear (LAOS) and machine learning
Podium 38
Presented by: Jun Dong Park
Introduction
A direct approach to investigate the texture of cosmetics is the panel test that evaluates the texture perceived by trained panels. However, the panel test is time-consuming, expensive and can be easily affected by irrelevant factors. As an alternative, there has been attempts to predict the sensory texture from the instrumental measurements. Especially, the sensory texture has often been related to rheology given that the sensory texture is result of flow and deformation of cosmetics. While meaningful correlations between rheological properties and sensory texture is reported in previous studies, it looks like a long way to go to establish quantitative prediction model.
Such limitations stem from the following reasons. Firstly, rheological analysis has been conducted under flow conditions that are not consistent with the actual use process. Application of cosmetic products is repeated rubbing in a wide skin area, during which cosmetics can experience various rheological transition. However, previous studies have been conducted based on the rheological test that does not reflect actual application process. Such lack of sophisticated rheological analysis reflecting actual cosmetics application process leads to failure of finding physically meaningful input for prediction model. Secondly, systematic approach to build a universal prediction model has been absent. Although large amount of reliable dataset is prerequisite for building a predictive model, relatively small data sets have been used in previous studies. Furthermore, efficient techniques, such as machine learning, for deriving a predictive model have rarely been adapted.
This study aims to provide an effective strategy for establishing a predictive model for texture of cosmetics. As a model problem, prediction of the spreadability of essence is discussed. Spreadability of essence samples are scored by professionally trained panels. In terms of rheology, the essences are investigated via Large Amplitude Oscillatory Shear (LAOS) analysis, inspired by an analogy between the actual cosmetics application process and LAOS. Based on the LAOS analysis and panel test database, a machine learning model is trained to predict the spreadability. It is shown that the trained model can effectively estimates the spreadability without panel test.
Methods
A spreadability prediction model is established based on the spreadability score from panel test and LAOS–SPP (Sequence of Physical Process) analysis database. Firstly, spreadability of total 77 cosmetic formulations with different formulas are tested by 12 trained panels and scored on a scale of 150. Rheological property of each sample is characterized by three parameters of the maximum transient elastic modulus, minimum transient viscous, and maximum stress that is measured under the LAOS condition of strain amplitude 10 and frequency 1rad/sec. Here, the maximum transient elastic modulus and the minimum transient viscous modulus are calculated from the SPP analysis. A linear regression model with the three parameters is trained with the gradient descent algorithm and the exhaustive cross-validated.
Results
The established predictive model for spreadability, which is scored on a scale of 150, shows Root Mean Square Error of 10.1. If a prediction with error smaller than 15 (10% of maximum score) is classified as a correct prediction, spreadability of 85% of essence samples is successfully predicted. It is shown that our model can make more accurate prediction, compared with a prediction model based on the linear viscoelasticity and simple shear stress.
Discussion and conclusion
A predictive model for spreadability of essence is established based on the LAOS analysis and machine learning techniques. Exceptionally large number of samples and panel test data is accumulated for building a general and reliable prediction model. Additionally, the state-of-the-art technique LAOS-SPP is introduced to investigate rheological behavior of cosmetics in actual application process. It is shown that our model, which is trained by sufficient panel test-LAOS analysis data, can effectively predict the spreadability of essence. The decent performance of the prediction model in this study illustrates the importance of 1) input variables reflecting actual cosmetics application process and 2) effective model equation derivation using machine learning techniques. Potentially, similar approach can be proposed for establishing a predictive model for other sensory texture.
A direct approach to investigate the texture of cosmetics is the panel test that evaluates the texture perceived by trained panels. However, the panel test is time-consuming, expensive and can be easily affected by irrelevant factors. As an alternative, there has been attempts to predict the sensory texture from the instrumental measurements. Especially, the sensory texture has often been related to rheology given that the sensory texture is result of flow and deformation of cosmetics. While meaningful correlations between rheological properties and sensory texture is reported in previous studies, it looks like a long way to go to establish quantitative prediction model.
Such limitations stem from the following reasons. Firstly, rheological analysis has been conducted under flow conditions that are not consistent with the actual use process. Application of cosmetic products is repeated rubbing in a wide skin area, during which cosmetics can experience various rheological transition. However, previous studies have been conducted based on the rheological test that does not reflect actual application process. Such lack of sophisticated rheological analysis reflecting actual cosmetics application process leads to failure of finding physically meaningful input for prediction model. Secondly, systematic approach to build a universal prediction model has been absent. Although large amount of reliable dataset is prerequisite for building a predictive model, relatively small data sets have been used in previous studies. Furthermore, efficient techniques, such as machine learning, for deriving a predictive model have rarely been adapted.
This study aims to provide an effective strategy for establishing a predictive model for texture of cosmetics. As a model problem, prediction of the spreadability of essence is discussed. Spreadability of essence samples are scored by professionally trained panels. In terms of rheology, the essences are investigated via Large Amplitude Oscillatory Shear (LAOS) analysis, inspired by an analogy between the actual cosmetics application process and LAOS. Based on the LAOS analysis and panel test database, a machine learning model is trained to predict the spreadability. It is shown that the trained model can effectively estimates the spreadability without panel test.
Methods
A spreadability prediction model is established based on the spreadability score from panel test and LAOS–SPP (Sequence of Physical Process) analysis database. Firstly, spreadability of total 77 cosmetic formulations with different formulas are tested by 12 trained panels and scored on a scale of 150. Rheological property of each sample is characterized by three parameters of the maximum transient elastic modulus, minimum transient viscous, and maximum stress that is measured under the LAOS condition of strain amplitude 10 and frequency 1rad/sec. Here, the maximum transient elastic modulus and the minimum transient viscous modulus are calculated from the SPP analysis. A linear regression model with the three parameters is trained with the gradient descent algorithm and the exhaustive cross-validated.
Results
The established predictive model for spreadability, which is scored on a scale of 150, shows Root Mean Square Error of 10.1. If a prediction with error smaller than 15 (10% of maximum score) is classified as a correct prediction, spreadability of 85% of essence samples is successfully predicted. It is shown that our model can make more accurate prediction, compared with a prediction model based on the linear viscoelasticity and simple shear stress.
Discussion and conclusion
A predictive model for spreadability of essence is established based on the LAOS analysis and machine learning techniques. Exceptionally large number of samples and panel test data is accumulated for building a general and reliable prediction model. Additionally, the state-of-the-art technique LAOS-SPP is introduced to investigate rheological behavior of cosmetics in actual application process. It is shown that our model, which is trained by sufficient panel test-LAOS analysis data, can effectively predict the spreadability of essence. The decent performance of the prediction model in this study illustrates the importance of 1) input variables reflecting actual cosmetics application process and 2) effective model equation derivation using machine learning techniques. Potentially, similar approach can be proposed for establishing a predictive model for other sensory texture.