46 TESTING
stages of the model building process foam properties that were not significant in improving the model performance were isolated and screened out. In the end, no single specific foam criteria was of critical importance; rather, we found that any meaningful analysis required a broader understanding of the overall variation generated by the entire dataset that the model is built on.
Among the foam property data available,
it was determined that the properties listed in Table 1 may be the most significant in describing the sensory perception of the quality of toothpaste foams. For the purposes of the current study,
the perceived sensory performance of the toothpastes based on their foam characteristics were evaluated and described using the following five properties: ease of foaming, amount of foam, foam consistency, creaminess of foam, aeration of foam. To gain a comprehensive understanding
of how toothpaste foam properties correlate with sensory panel scores, it was determined that data across the entire sensory panel space would need to be considered and analysed. Subsequently, a design of experiment (DoE) study was devised that efficiently minimised the number of experiments needed to explore the relationship between foam analytical properties and sensory panel scores across the entire sensory panel spectrum. This was achieved by analysing previous
sensory panel scores as provided by our external partner company and picking toothpastes that had responses as close to the corners of the design space as possible. The foam data for these toothpastes was subsequently generated with a foam analyser. Attempts were initially made to identify links
and predictive patterns between the two data sets via traditional statistical methods such as basic regression models, correlation analysis and ANOVA (analysis of variance). However, the inherent complexity and variability within both datasets made it difficult for these conventional methods to capture meaningful correlations and patterns.
Machine learning approach It was clear at this point that more advanced statistical methodologies and a more refined computational data analysis approach would be needed to achieve the desired outcome of using the foam characteristic data to model the sensory panel outcomes. Lucideon’s Integrated Materials Processing and Computational Techniques (IMPACT™) team subsequently reviewed the available data to determine whether data science and machine learning approaches could enhance the analysis. Recognising the limitations of traditional methods and the complexities inherent in the dataset, a constructive approach to machine learning model development was essential to achieve robust validation and reliable predictions. IMPACT’s Predictive Modelling Machine
Learning framework and data pipeline were specifically tailored in-house to handle the complexities of the current dataset. This framework’s capability to manage continuous
PERSONAL CARE November 2024
iterations and incorporate feedback significantly boosted the performance of the predictive model. Designed with a focus on materials datasets, the framework enabled fast and effective data processing and model experimentation, improving the model’s forecast performance and reliability. The model development journey began with an initial baseline model. Although not perfect in accuracy, it revealed some intriguing patterns within the data. These patterns suggested some potential pathways for improving the predictions. The initial model served as a valuable foundation, highlighting the possibilities for refinement and optimisations in subsequent iterations. Building on this foundation, a process
of thorough and iterative model refinement followed. To identify the most suitable machine learning algorithm, our team’s framework and data processing pipeline enabled experimentation through various combinations of machine learning algorithms, ranging from straightforward linear models such as regression to more complex neural networks. This flexibility enabled effective screening and identification of the best algorithms and parameters for the specific requirements of the current study. Each iteration involved integrating new data,
augmenting existing data, carrying out data pre-processing and cleaning, adjusting model parameters, and carefully validating the model’s predictions against sensory panel results. Given the high cost and effort associated
TABLE 1: KEY PHYSICAL FOAM ATTRIBUTE PARAMETERS Bulk Foam Properties
Time of maximum foam height Maximum foam volume
Liquid volume bound in the foam Maximum total volume
Microstructural Foam Properties Time of foam stabilisation Initial bubble count
Initial mean bubble area
Initial mean bubble radius Bubble count half-life time Final bubble count Final mean bubble area Final mean bubble radius
with acquiring quality datasets for these experiments, the data acquisition process was rigorously planned and strategised to maximise the model’s improvement per iteration. Additional foam analyser experiments were
conducted on toothpaste samples that the model had not previously seen, ensuring robust data validation. These samples were selected from a diverse range provided by our partner company, with characteristics deliberately chosen to ensure that they varied significantly in their sensory panel evaluations. This approach, combined with mathematical techniques verified by domain experts, was used to enhance the model’s accuracy and reliability. As the model development progressed,
the predictive model demonstrated significant accuracy improvements, increasingly aligning with user requirements and providing valuable insights into the interplay between various foam attributes and sensory scores. Accuracy in the context of the current
study was defined as how closely the model’s predictions match the experiential outcomes generated by the trained sensory panel. For example, if the model predicts a particular toothpaste will have a high score for ‘Amount of Foam’ and the sensory panel also rates it highly, the model is considered accurate. An accuracy score of around 80%
was achieved, which was a significant developmental milestone for the project. For the purposes of the current study, this indicates the model was sufficiently accurate to serve as
Figure 1: Predictive data analytics can be used to leverage large, high-quality datasets as a tool to forecast experimental outcomes
www.personalcaremagazine.com
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60 |
Page 61 |
Page 62 |
Page 63 |
Page 64 |
Page 65 |
Page 66 |
Page 67 |
Page 68 |
Page 69 |
Page 70 |
Page 71 |
Page 72 |
Page 73 |
Page 74 |
Page 75 |
Page 76 |
Page 77 |
Page 78 |
Page 79 |
Page 80 |
Page 81 |
Page 82 |
Page 83 |
Page 84 |
Page 85 |
Page 86 |
Page 87 |
Page 88 |
Page 89 |
Page 90 |
Page 91 |
Page 92 |
Page 93 |
Page 94 |
Page 95 |
Page 96 |
Page 97 |
Page 98 |
Page 99 |
Page 100
