Machine learning approach to predict the viscosity of perfluoropolyether oils

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Chemistry

Background:

• Perfluoropolyethers (PFPEs) exhibit excellent chemical stability and thermal resistance.
• PFPEs are vital lubricants in demanding sectors like aerospace and semiconductors.
• Experimental viscosity determination for PFPEs is costly and time-intensive.

Purpose of the Study:

• To develop accurate predictive models for PFPE viscosity.
• To compare the performance of various intelligent models against traditional correlations.
• To establish a reliable computational method for PFPE lubricant characterization.

Main Methods:

• Utilized Multilayer Perceptron (MLP), Support Vector Regression (SVR), Gaussian Process Regression (GPR), and Adaptive Boost Support Vector Regression (AdaBoost-SVR).
• Input parameters included temperature, density, and average polymer chain length.
• Evaluated model performance using statistical error analysis and the leverage technique.

Main Results:

• Gaussian Process Regression (GPR) demonstrated superior accuracy with an RMSE of 0.4535 and R² of 0.999.
• The GPR model outperformed traditional Waterton and Vogel-Fulcher-Thamman (VFT) correlations.
• Leverage analysis confirmed the robustness of the GPR model, with 98.33% of data points within the valid range.

Conclusions:

• GPR provides a highly accurate and reliable method for predicting PFPE viscosity.
• The developed GPR model captures essential physicochemical trends, offering insights beyond simple prediction.
• This computational approach significantly reduces the need for expensive experimental viscosity measurements.