Predicting the firefighting efficacy of surfactants prior to synthesis via ensemble artificial neural network modeling of a foam performance database

Abstract

Research efforts incorporating machine learning (ML) are currently focused on developing replacements for the toxic and bio-accumulative per- and polyfluorinated alkyl substances in fire suppressing foams. In the following work, ensembles of 10 artificial neural networks (ANN) were trained on a fire suppression database, described by Sudol et al., correlating area under the curve values obtained from 19-cm gasoline and heptane pool fire extinction curves to the molecular descriptors of surfactants within various firefighting foams. These ANN model ensembles were then used to evaluate proposed surfactant structures to predict the firefighting effectiveness prior to laboratory synthesis. The two most promising surfactants were a tetrasiloxane diglucoside and a chlorotrisiloxane-polyethyleneoxide (PEO). These surfactants were synthesized, and their fire extinction performances were assessed via 19-cm gasoline and heptane pool fire experiments to validate the ANN predictions. The synthesis of the demonstrably high-performing tetrasiloxane diglucoside surfactant is considered a successful ML application in the context of fluorine-free firefighting surfactant research and development. Meanwhile, the synthesis of the low-performing chlorinated PEO surfactant, which failed to meet predicted performance expectations, demonstrates the need for both comprehensive training data sets and the proper consideration of modeling redundancies to safeguard against unreliable ML-derived performance predictions.