Modern industrial liquids-including coolants, lubricants, solvent blends, cryoprotectant cocktails, etc.-frequently employ complex, many-component formulations, but contemporary viscosity models and datasets are overwhelmingly limited to simplified binary or ternary compositions, leaving the most application-relevant compositional spaces broadly unexplored. This gap is attributable both to the limitations of classical viscosity correlations, which typically require either untenably idealized interaction assumptions or an untenably larger number of interaction parameters, and to the lack of systematic many-component viscosity datasets. Here, we provide a first-of-its-kind dataset of 44 316 viscosity measurements spanning 100 aqueous and organic solutions of up to 17-component complexity across temperatures from -20 to 35 °C, and we use it to power a physics-informed neural network (PINN) model that provides unprecedented predictive power and physical insight into many-component solution viscosity. We first show that predictive implementations of two prominent classical models (Katti-Chaudhuri and augmented Adam-Gibbs) systematically fail to describe these solutions but retain valuable trend-level information. We then use these models as physical guides for our PINN, embedding classical model insight as a training aid while machine learning the residual non-ideal contributions that dominate in many-component composition space. The resulting model, which we also provide as a software application, maintains stable performance across solution complexity, outperforming both classical correlations and data-only artificial neural networks on fully withheld test data. Finally, we compare residuals of both the classical and PINN models as a function of component number and entropy of mixing, suggesting that entropic phenomena unaccounted for in classical models may dominate the viscosity of many-component solutions.
Kavian et al. (Mon,) studied this question.