Surrogate mixture and fuel/biofuel blend design supported by GCA-EOS and machine learning models for properties prediction

AYLEN MERLO RAIMONDI; SELVA PEREDA; ANIBAL BLANCO; MARIANA GONZALEZ PRIETO

Congreso; XII IBEROAMERICAN CONFERENCE ON PHASE EQUILIBRIA AND FLUID PROPERTIES FOR PROCESS DESIGN / XI Brazilian conference on applied thermodynamics; 2021

Biofuels are oxygenated renewable additives that modify the final fuel properties. In that regard, the non-ideal phase behavior of mixtures comprising hydrocarbons and oxygen-bearing compounds has a significant impact on blends storage, transportation and use in engines. In particular, volatility properties correlate with numerous parameters of the fuel performance, thus many standardized tests regulate the quality of blends indirectly through this property. Additionally, the octane number is a non-equilibrium property, which measures the fuel ignition quality and its tendency to resist knocking. Since its experimental determination is expensive and time consuming, many correlations based on different pure compound properties and mixing rules have been proposed, most of which fail when challenged to predict new data. The octane number strongly depends on the molecular structure of species in the fuel and the interaction among them does not show an easily predictable response on the mixture octane number. Moreover, the complexity is higher when different functional groups like aromatic, olefinic or oxygenated are present. In this scenario, with complex and unknown underlying fundamentals, machine learning models provide a robust alternative that shows good performance for a broad scope of blends. In this work, we develop a software for fuel/biofuel blend design that aims to reduce the exploratory laboratory work. Since the detailed composition profile of the base fuels are rarely known, we also apply this tool to search for surrogate mixtures based on the fuel bulk properties and its PIANOX (the fraction of each type of hydrocarbon that comprises the fuel). We first present algorithms to simulate two standardized test that regulate the volatility of a fuel blend, the Reid Vapor Pressure (RVP) and the fuel Distillation Curve (DC). In both cases, equilibrium calculations are based on predictions of the Group Contribution with Association Equation of State (GCA-EOS). On the other hand, we assess the accuracy of an artificial neural network (ANN) proposed in the open literature to predict the octane number of mixtures. Finally, to characterize base fuels and design fuel blends, we combined stochastic and deterministic optimization techniques for a fast search of surrogate multicomponent mixtures that fulfill specified properties, approach that allows using rigorous complex non-linear thermodynamic models and sophisticated ANN models in the property prediction package.