(191aa) Theory-Based Quantitative Structure-Property Relationship Models for Phase Behavior Predictions
AIChE Annual Meeting
2010
2010 Annual Meeting
Engineering Sciences and Fundamentals
Poster Session: Thermodynamics and Transport Properties
Monday, November 8, 2010 - 6:00pm to 8:00pm
Multiphase equilibrium calculations are an integral part of the design and optimization of numerous chemical processes. Several accurate experimental techniques have been developed for measuring phase equilibrium data; however, experimental techniques are time consuming and costly. Hence, a need exists for reliable thermodynamic models capable of giving a priori predictions of the phase behavior of diverse systems in the absence of experimental data. Quantitative structure?property relationship (QSPR) modeling has the potential to provide reliable property estimates based on detailed chemical structure information. Although current QSPR models have been successful in providing reliable structure-based property predictions, they have been limited to estimating properties at a single temperature. Further, little work has been done on QSPR models for mixtures.
In this work, we refine our generalization of the universal quasi-chemical (UNIQUAC) excess Gibbs energy model. Improved structure-based UNIQAC model parameters were obtained using QSPR techniques and a modeling strategy that assures uniqueness in the binary parameter values generated. As such, an integrated UNIQUAC-QSPR model capable of a priori prediction of the vapor?liquid equilibrium (VLE) phase behavior was developed. The descriptors for the QSPR model were generated using CODESSA. The suitable descriptor subset for this model was selected based on a differential evolution (DE) optimizer that is wrapped around neural network mapping functions. The final model was built using an ensemble of the best 5 models identified using the wrapper algorithm. The current UNIQAC-QSPR model generalization produced VLE predictions for a wide variety of systems (329 binaries) with twice to three times the errors obtained from direct parameter regressions.