2025 Spring Meeting and 21st Global Congress on Process Safety
(92b) Evaluating Graph-Based Models for Logcmc Prediction in Surfactants: A Comparative Study of Pharmhgt, GCN, and Gat.
Authors
Gabriela Theis Marchan - Presenter, Louisiana state university
Andrew N Okafor, Louisiana State University
Jose Romagnoli, Louisiana State University
Accurate prediction of the critical micelle concentration (LogCMC) of surfactants is crucial for optimizing their use in a range of industrial applications, such as pharmaceuticals, detergents, and emulsions. This study presents a comparative analysis of three prominent graph-based machine learning models: Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and the transformer-based model PharmHGT, to assess their performance in predicting LogCMC values. Our investigation explores each model’s ability to capture the intricate structural and physicochemical characteristics of surfactants. The findings underscore the comparative advantages and limitations of these approaches, emphasizing the strengths of transformer-based models like PharmHGT in processing molecular graph representations over traditional graph neural networks. This work advances the understanding of model capabilities in LogCMC prediction and supports the development of more precise tools for surfactant design and application.