Hybridizing physical and data-driven prediction methods for physicochemical properties.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Predicting physicochemical properties is crucial in chemical research and development.
• Existing physical methods often lack accuracy, while data-driven methods require extensive experimental data.
• There is a need for hybrid approaches that leverage the strengths of both physical insights and empirical data.

Purpose of the Study:

• To develop a generic framework for hybridizing physical and data-driven methods.
• To improve the prediction of physicochemical properties, specifically activity coefficients at infinite dilution.
• To demonstrate the superiority of the hybrid approach over traditional methods.

Main Methods:

• A novel approach that 'distills' physical model predictions into a prior model.
• Integration of the prior model with sparse experimental data using Bayesian inference.
• Application of the hybrid method to predict activity coefficients at infinite dilution.

Main Results:

• Significant improvements in prediction accuracy for activity coefficients at infinite dilution.
• Outperformance compared to standalone physical methods.
• Outperformance compared to standalone data-driven methods and established ensemble machine learning techniques.

Conclusions:

• The proposed hybrid approach offers a powerful and versatile tool for predicting physicochemical properties.
• This method effectively combines physical understanding with experimental data for enhanced accuracy.
• The approach represents a significant advancement in the field of predictive chemistry.