Toothpaste is composed of multiple substances, including humectants, cleaning agents, foaming agents, medicinal ingredients and viscosity modifiers. To develop high-quality toothpaste that meets consumer needs, we draw on experimental data and researcher knowledge to refine the composition until our targets are met, carrying out repeated experiments with various ingredients, amounts and combinations. When developing a composition with new quality parameters or new ingredients, because there is a lack of research data in the earlier stages, and additional study is needed. To address this, we established an experiment design method that allows researchers to simultaneously consider multiple targets with fewer rounds of experimentation. The method effectively incorporated researcher knowledge using Bayesian optimization, a machine learning method that allows a composition search that starts with limited, known data (Figure 1).
By applying this method, we are able to search for the optimal modulus of viscosity and elasticity, which indicate physical properties such as the softness or firmness of toothpaste. As a result, although it is not unusual for us to repeat an experiment more than one hundred times, we were able to derive a composition that met all of our targets after just sixteen attempts, completing composition development in about half the anticipated time, including the time required for the additional processes involved. After the sixth experiment during the study, we saw that deviations between predicted and observed values tend to narrow with repeated testing, indicating greater efficiency in identifying the optimal composition (Figure 2). Going forward, we plan to expand this method to the composition development of more toothpaste and other products.
Based on ten sets of known data, we are able to use this method to derive the optimal formulation for this composition and create toothpaste prototypes. The figure shows the predicted and observed values obtained by measuring the viscosity and modulus of elasticity of the prototype, incorporating known data, deriving another optimal composition and conducting further testing.