Data-efficient surrogate modeling of thermodynamic equilibria using Sobolev training, data augmentation and adaptive sampling

Abstract

Modern thermodynamic models, such as the PC-SAFT equation of state, are very accurate but also computationally intensive, which limits their applicability to process design optimization, for example. Surrogate models, which can be evaluated quickly, can be used to approximate the thermodynamic equilibria. However, this requires many data points from the flash calculation routine. In this paper, we investigate three approaches to reduce the number of samples and thus the effort needed to train the surrogate models. First, Sobolev training is used, where the surrogate model is trained not only on the output values, but also on derivative information. Second, data augmentation along the tie lines in LLE systems is proposed to generate samples without additional flash calculations. Third, adaptive sampling is revisited with a novel quality criterion. It is shown that the combination of these techniques can be used to significantly reduce the number of samples required.