Can Machine Learning Predict the Phase Behavior of Surfactants?

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Materials Science

Background:

• Surfactant phase behavior is crucial for various industrial applications.
• Predicting phase diagrams is complex and often relies on extensive experimental data.
• Machine learning (ML) offers potential for accelerating these predictions.

Purpose of the Study:

• To evaluate the efficacy of ML methods for predicting nonionic surfactant phase behavior.
• To identify limitations and areas for improvement in ML-based phase diagram prediction.
• To assess the impact of data sampling strategies on prediction accuracy.

Main Methods:

• Utilized a dataset of twenty-three nonionic surfactants.
• Tested various state-of-the-art machine learning classifiers.
• Investigated the performance of ML in data imputation versus *de novo* phase diagram prediction.
• Explored strategies for enhancing *de novo* prediction accuracy.

Main Results:

• ML classifiers effectively filled missing data points in existing datasets.
• Significant challenges were observed in predicting entire *de novo* phase diagrams due to data bias and limited chemical space.
• Prediction accuracy was largely independent of the specific ML algorithm chosen.
• Incorporating data from experimental-analogy sampling improved *de novo* prediction outcomes.

Conclusions:

• ML is adept at data imputation for surfactant systems.
• Robust *de novo* phase diagram prediction requires addressing data bias and chemical space limitations.
• Strategic data acquisition, inspired by experimental protocols, is key to advancing ML-driven surfactant phase behavior prediction.