Making thermodynamic models of mixtures predictive by machine learning: matrix completion of pair interactions

Abstract

Predictive models of thermodynamic properties of mixtures are paramount in chemical engineering and chemistry. Classical thermodynamic models are successful in generalizing over (continuous) conditions like temperature and concentration. On the...

Database for liquid phase diffusion coefficients at infinite dilution at 298 K and matrix completion methods for their prediction

“…4,17 In particular, they have been applied very successfully for predicting activity coefficients and Henry's law constants. 4,[28][29][30][31] In the present work, we extend the MCM approach to the prediction of diffusion coefficients.…”

Prediction of parameters of group contribution models of mixtures by matrix completion

“…The MCM that was used in the present work is based on Bayesian matrix factorization 31 and similar to the ones used in our previous works. [22][23][24][25]27 In principle, we could have applied the MCM directly to the matrix of the A-type parameters, i.e., the matrix containing the group-interaction parameters A GG 0 and A G 0 G . However, this option was discarded for the following reasons: firstly, the available values for A GG 0 and A G 0 G are inconsistent with eqn (1).…”

“…y G , b G , y G 0 , and b G 0 are parameters of UNIFAC-MCM, while K is a hyperparameter that controls the number of features considered per main group and thereby determines the flexibility of the model. Based on results of our prior work, 22 K was set to K = 3 here. The form of eqn ( 2) was chosen to ensure that all resulting group-interaction energies are symmetric, as required by the lattice model.…”

“…In a series of recent papers, we have demonstrated the capabilities of MCMs for predicting different types of thermodynamic data of mixtures using various component-based approaches. [22][23][24][25][26][27] However, these component-based approaches are inherently limited regarding the number of components that are covered; the respective models complete a matrix spanned by the components that are part of the mixtures in the training set. This is not the case for the group contribution methods, which we consider in the present work: as the groups form building blocks from which components can be created flexibly, the scope of the group contribution methods for mixture properties is inherently extremely large -and it can now be extended substantially by using an MCM to complete the set of group-interaction parameters.…”

“…Following an idea developed in a recent paper, 22 in which we have applied an MCM for estimating the component-specific pair-interaction parameters of UNIQUAC, we do not use the asymmetric group-interaction parameters ( A GG ′ ≠ A G ′ G ) directly, but rather the symmetric group-interaction energies U GG ′ = U G ′ G . The parameters of the two types ( A and U ) are connected by:Hence, according to eqn (1), A GG ′ and A G ′ G are not independent but correlated ‡…”