Integrating Entropy Scaling into a Deep Neural Network Architecture to Predict Viscosities

One significant challenge in the development of new sustainable processes and materials is the scarcity of availability of property data. Besides, data driven models are often not able to handle thermodynamic constraints adequately. Integrating advanced machine learning methods with physically-based modeling techniques allows to combine advantages of both approaches leading to models that perform better than the approaches on their own. Here, a neural network architecture incorporating the entropy scaling approach is proposed to predict shear viscosities over a large range of species and thermodynamic state points. The resulting models demonstrate high prediction accuracy even for complex molecules with various functional groups, outperforming most traditional group contribution methods and most molecular simulations using current classical force fields.