Modeling the Property–Performance Relationship for Distillation-Based Binary Zeotropic Separation

The first key question to answer in separation engineering is which unit of operation should be employed. Due to the wide applications, distillation is always prioritized as a viable option. In general, given a specific task, rigorous process optimization needs to be performed for identifying the real potential of distillation. To prevent the time-consuming optimization and reveal the optimal performance rapidly, property–performance relationship models for distillation-based binary zeotropic separation are developed. For separating light and heavy components, based on the analysis of the rigorous mathematical model of distillation, six physical properties of components and their vapor–liquid equilibrium (VLE) are found to have major impacts on the process performance. Then, the properties and VLE of 97 real and 103 hypothetical binary mixtures are collected and utilized as input for carrying out rigorous process optimization. This generates the minimal total annualized cost, optimal reflux ratio, and number of trays directly. With those input and output data, three convolutional neural network (CNN) models are built to represent the property–performance relationships and predict the optimal performance of distillation operations. Meanwhile, by integrating the in-house database, property prediction models, and CNN models, a toolkit, DistBin, is constructed. This enables a quick estimation of the distillation performance and only provides the two components and their compositions. The toolkit formed can serve as an agent and be integrated into a large language model directly, so that the economic viability of distillation-based separation can be known in a question-and-answer form immediately.