Predicting Foam Characteristics of Body Wash Products Using Machine Learning

Abstract Foam generation and structure strongly influence the consumer perception of body wash products, yet systematic optimization of foaming properties is challenging because formulations are complex multicomponent mixtures. Here, we develop machine learning (ML) models to predict key foam attributes of body wash formulations from detailed composition data. A data set of 545 proprietary formulations, comprising 362 distinct ingredients, was assembled by combining internal formulation records with foam measurements. Six foam descriptors were quantified: maximum foam volume, initial and final bubble count, initial and final mean bubble radius, and bubble count half-life. Five ML algorithms─decision tree, k-nearest neighbors, extra trees, gradient boosting machines, and random forest were trained and evaluated by stratified 5-fold cross-validation repeated with five random seeds. The results indicate that ensemble tree-based models outperformed the single-tree and k-nearest neighbor approaches, with random forest achieving cross-validated coefficients of determination (R2) between 0.62 and 0.86 across the six attributes. Shapley additive explanations (SHAP) analysis identified sodium laureth sulfate, cocamidopropyl betaine, opacifiers, fatty acids, and xanthan gum as the most influential contributors to foam behavior. Finally, we demonstrate an ML formulation capability using the trained models to explore trade-offs between foam volume and foam density. This study illustrates how ML can accelerate formulation development and enable data-driven design of personal care products.