A smile is all you need: predicting limiting activity coefficients from SMILES with natural language processing

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Machine Learning

Background:

• Accurate prediction of mixture phase equilibria relies on activity coefficients.
• Experimental determination of activity coefficients is costly and data is often limited.
• Existing machine learning models struggle with extrapolating predictions for novel molecules.

Purpose of the Study:

• To develop a novel machine learning approach for predicting binary limiting activity coefficients.
• To overcome data limitations by utilizing a hybrid training strategy.
• To improve the accuracy of activity coefficient predictions for unknown chemical compounds.

Main Methods:

• Introduction of the SMILES-to-properties-transformer (SPT), a natural language processing network.
• Initial training on a large synthetic dataset (10 million points) from COSMO-RS.
• Fine-tuning the model on a curated experimental dataset (20,870 points).

Main Results:

• The SPT model demonstrates high accuracy in predicting binary limiting activity coefficients.
• The model effectively predicts coefficients for previously unseen molecules.
• Achieved a 50% reduction in mean prediction error compared to COSMO-RS and UNIFAC Dortmund.

Conclusions:

• The SPT model offers a significant advancement in predicting activity coefficients for diverse chemical mixtures.
• Hybrid training overcomes experimental data scarcity, enabling accurate predictions for novel compounds.
• This approach enhances the efficiency and reliability of phase equilibria calculations in chemical process design.