Fig. 2: Extracting the sorptivity of paste specimens using machine learning.
From: Automated estimation of cementitious sorptivity via computer vision
a Predicting paste sorptivity via ASTM C1585 shows consistent and repeatable water penetration dynamics across systems. The error bar represents the standard deviation of six replicates. b Initial and secondary sorptivity values in paste specimens are inversely correlated. c A heatmap scatter plot illustrates the relationship between wetted area ratio, time, and mass change. d Three random paste specimens highlight that porous matrices with larger wetted areas at shorter durations show higher mass change, while denser matrices with smaller wetted areas over longer durations exhibit lower mass change. In this subplot, ‘pre’ and ‘post’ refer to original and FPN-analyzed images, respectively. e The machine learning model accurately predicts penetration patterns for the training dataset (R² = 0.97). f The model also shows strong performance on the testing dataset, matching predicted and measured penetration values (R² = 0.93). g Indirectly estimating sorptivity from predicted penetration-versus-time data yields high accuracy (R² > 0.97), while (h) direct estimation using only wetted area and time data results in lower accuracy (for Both Si = blue points and SS = red points), particularly for secondary sorptivities (R2 = 0.83). Source data are provided as a Source Data file.
