A Machine Learning-Fueled Modelfluid for Flowsheet Optimization

TL;DR

The paper tackles the data scarcity barrier in process-fluid optimization by introducing a ML-driven modelfluid built from physically interpretable VLE-based features. This modelfluid is mapped to simple thermodynamic models and embedded into distillation flowsheet simulations, enabling gradient-based optimization for solvent/entrainer design. The authors demonstrate the approach with an entrainer search for an acetone–chloroform azeotrope, combining baseline Pareto-frontier optimization with hypothetical entrainer optimization and a practical ML-driven mapping to real candidates. A local objective-prediction scheme ranks real candidates from a large pool, with rigorous validation showing strong correlation between predicted and true performance. Overall, the work provides a data-efficient pathway to leverage large-scale property predictions in practical process-design workflows, enabling scalable solvent and entrainer screening while maintaining thermodynamic consistency.

Abstract

Process optimization in chemical engineering may be hindered by the limited availability of reliable thermodynamic data for fluid mixtures. Remarkable progress is being made in predicting thermodynamic mixture properties by machine learning techniques. The vast information provided by these prediction methods enables new possibilities in process optimization. This work introduces a novel modelfluid representation that is designed to seamlessly integrate these ML-predicted data directly into flowsheet optimization. Tailored for distillation, our approach is built on physically interpretable and continuous features derived from core vapor liquid equilibrium phenomena. This ensures compatibility with existing simulation tools and gradient-based optimization. We demonstrate the power and accuracy of this ML-fueled modelfluid by applying it to the problem of entrainer selection for an azeotropic separation. The results show that our framework successfully identifies optimal, thermodynamically consistent entrainers with high fidelity compared to conventional models. Ultimately, this work provides a practical pathway to incorporate large-scale property prediction into efficient process design and optimization, overcoming the limitations of both traditional thermodynamic models and complex molecular-based equations of state.

Figures (10)

• Figure 1: Starting by a set of entrainer candidates, data from both literature databases and ML-based prediction methods are used to obtain the modelfluid features. A process fluid optimization based on those features is combined with an objective approximation method to yield a list of sorted entrainer candidates. This workflow demonstrates the embedding of the proposed modelfluid representation into a process fluid optimization task.
• Figure 2: Entrainer distillation flowsheet for the separation of a binary mixture with maximum-boiling azeotropic phase behavior. The columns are denoted by $Co_1$, $Co_2$, and $Co_3$, and the mixer is denoted by $M$. The liquid flow rates and molar fractions of the streams are represented by $L$ and $\ell$, respectively. The subscripts of these variables correspond to the stream numbers shown in the figure. The variables $RR$, $BR$, $\dot{Q}^{Con}$, $\dot{Q}^{Reb}$, and $\dot{Q}$ represent the reflux ratio, boilup ratio, condenser duty, reboiler duty, and heat duty, respectively.
• Figure 3: Schematic workflow for the identification of hypothetical optimal entrainer components. Starting from a known reference entrainer's NQ curve, we compute the optimal entrainer features for each point on the NQ curve, fixing the number of stages in the column and further minimizing the heat duty in the flowsheet. This yields a set of hypothetical optimal entrainer features, each representing an individual hypothetical entrainer.
• Figure 4: Comparison between the NQ curve obtained for the reference entrainer Benzene (lime-colored dash line with circle markers) and those obtained for the hypothetical optimal entrainers (black-colored solid lines with square markers) found in step 2. All NQ curves are computed by solving the optimization problem of step 1, but replacing the reference entrainer with the hypothetical entrainers found in step 2.
• Figure 5: Workflow of how the modelfluid features for each of the entrainer candidates is obtained using a PCP model parameter database and predictions of the activity coefficients at infinite dilution. The $\gamma_i\vert_j$ predictions are obtained using the Matrix Completion Method as proposed in Jirasek2020.