(169bk) Utilizing Surfactant-Specific Graph Convolutional Networks to Predict Surfactant Adsorption Efficiency

Quantifying the adsorption efficiency of surfactants, measured by the surface area occupied per molecule, is vital for the development of new surfactants, particularly those for firefighting foams. Dense surfactant packing at interfaces is believed to significantly hinder oil movement through the surfactant layer due to high resistance. Nevertheless, deciphering the intricate link between a surfactant's structural makeup and its adsorption efficiency poses a significant challenge. To address this challenge, we introduce a graph convolution network (GCN) specifically designed for surfactants to predict their adsorption efficiency. This model categorizes the surfactant molecules into head and tail sections, incorporating both specific features of these sections and general atomic characteristics into the feature matrix. Some of these head and tail features are derived from low-cost molecular dynamics (MD) simulations. Furthermore, the model is trained on a comprehensive dataset of 98 different surfactant structures, covering all surfactant categories such as anionic, cationic, zwitterionic, and nonionic, thus ensuring its applicability across various surfactant types. Our findings indicate that this specialized model surpasses conventional deep learning models in accuracy when predicting surfactant adsorption efficiency across a wide array of data. Additionally, molecular saliency maps are explored to gain insights into how specific surfactant substructures affect adsorption efficiency.