Prediction of infinite-dilution activity coefficients with neural collaborative filtering

Accurate prediction of infinite dilution activity coefficient (γ^∞) for phase equilibria and process design is crucial. In this work, an experimental γ^∞ dataset containing 295 solutes and 407 solvents (21,048 points) is obtained through data integrating, cleaning, and filtering. The dataset is arranged as a sparse matrix with solutes and solvents as columns and rows, respectively. Neural collaborative filtering (NCF), a modern matrix completion technique based on deep learning, is proposed to fully fill in the γ^∞ matrix. Ten-fold cross-validation is performed on the collected dataset to test the effectiveness of the proposed NCF, proving that NCF outperforms the state-of-the-art physical model and previous machine learning model. The completed γ^∞ matrix makes solvent screening and extension of UNIFAC parameters possible. Taking two typical hard-to-separate systems (benzene/cyclohexane and methyl cyclopentane/n-hexane mixtures) as examples, the NCF-developed database provides high-throughput screening for separation systems in terms of solvent selectivity and capacity.