Predictive model for wetness sensation in fabrics based on machine learning and SHAP

Currently, many fabric properties collectively influence the wetness sensation experienced upon contact; however, the individual contributions and underlying mechanisms of these properties remain unclear. This study examined 20 fabrics based on 10 physical properties relevant to wetness sensation: thickness, grammage, friction coefficient, surface roughness, bending stiffness, bending hysteresis, maximum instantaneous heat flow, wicking height, moisture regain, and saturated water content. Subjective evaluations were conducted at three relative water content (RWC) levels: 25%, 50%, and 75%. Using these properties and RWC as input features, and wetness sensation scores as output, eight machine learning models were developed. The Gradient Boosting Regression model demonstrated the best performance (R² = 0.904 and RMSE = 0.570). SHAP analysis identified RWC, wicking height, saturated water content, thickness, bending stiffness, and friction coefficient as the key factors influencing the results. The influence of fabric properties on wetness sensation is complex and nonlinear, with each property exhibiting a threshold beyond which its impact on wetness perception alters. Additionally, the effect of each property varies with RWC levels; for instance, wicking height has a greater effect at low RWC, while thickness, bending stiffness, and friction coefficient exert stronger effects at moderate RWC, and saturated water content at high RWC. These findings provide actionable insights for optimizing textile design to enhance human wet comfort, enabling the development of more comfortable and performance-oriented fabrics for diverse applications.