Robust Method for Property Prediction via Artificial Neural Networks: Incorporating Key Structural Features for...

Area of Science:

• Chemical Engineering
• Materials Science

Background:

• Existing data on carbon dioxide and ionic liquid mixtures, especially flow properties, is fragmented, limiting practical applications.
• Accurate prediction of mixture properties is essential for designing and optimizing industrial processes.

Purpose of the Study:

• To develop a robust and efficient method for predicting the density of carbon dioxide-ionic liquid mixtures.
• To establish a machine learning-based approach for property prediction validated by molecular dynamics simulations.

Main Methods:

• Artificial neural networks (ANNs) were trained using ionic liquid critical properties, structural descriptors, or a combination thereof.
• Models were validated using novel techniques, including molecular dynamics simulations and cross-comparison tests.
• A postprocessing outlier-handling method was employed to enhance model performance.

Main Results:

• ANN models achieved relative deviations below 3% for testing data.
• Combining critical and structural data significantly improved prediction accuracy (R² = 0.986).
• The combined ANN model demonstrated robust generalization, accurately predicting properties outside training ranges and for unseen ionic liquids.

Conclusions:

• Artificial neural networks provide a highly accurate and efficient method for predicting carbon dioxide-ionic liquid mixture densities.
• This computational approach offers a faster and broader alternative to traditional thermodynamic tools.
• The study establishes a solid foundation for future machine learning-based property predictions in chemical engineering.