Data-Driven Imputation of Miscibility of Aqueous Solutions via Graph-Regularized Logistic Matrix Factorization

Area of Science:

• Computational Chemistry and Materials Science
• Chemical Engineering and Process Design

Background:

• Aqueous, two-phase systems (ATPSs) are vital for numerous separation, purification, and extraction processes.
• Accurate prediction of ATPS miscibility is essential for accelerating the discovery of new systems and reducing development costs.
• Existing machine learning methods often rely on detailed physicochemical properties of individual solutes, which may not always be available.

Purpose of the Study:

• To develop a novel machine learning approach for predicting ATPS miscibility directly from incomplete pairwise experimental data.
• To demonstrate the efficacy of graph-regularized logistic matrix factorization (GR-LMF) in imputing missing miscibility outcomes.
• To compare the performance of GR-LMF against conventional methods using solute physicochemical features.

Main Methods:

• Implementation of graph-regularized logistic matrix factorization (GR-LMF) to learn latent representations of solutions.
• Utilizing observed pairwise miscibility data and a solute category graph (polymer, surfactant, salt, protein) as input for GR-LMF.
• Comparison with ordinary logistic matrix factorization and random forest classifiers trained on solute physicochemical features.

Main Results:

• GR-LMF achieved higher accuracy in predicting missing miscibility outcomes compared to baseline methods.
• The GR-LMF model successfully imputed miscibility data without requiring explicit physicochemical features for each solution.
• The model's predictive capability for new, unobserved solutions is dependent on having some prior miscibility data within the training set.

Conclusions:

• GR-LMF offers a powerful and data-efficient approach for predicting ATPS miscibility, particularly when experimental data is sparse.
• This method obviates the need for extensive solute characterization, streamlining the discovery process for new ATPS applications.
• Future work may explore strategies to enhance the prediction of entirely new solutions not represented in the initial training data.