Machine learning-based prediction of octane number of fuels supported by the thermodynamic model GCA-EOS

MERLO RAIMONDI, A.; GONZÁLEZ PRIETO, MARIANA; BLANCO, A.; PEREDA, S.

Buenos Aires

Congreso; 11th World Congress on Chemical Engineering; 2023

World Chemical Engineering Council - Asociación Argentina de Ingeniería Química

According to the International Energy Agency, bioenergy is an essential component of the future low-carbon global energy system if global climate change commitments are to be met, playing a particularly important role in the decarbonization of heavy transportation sectors such as aviation, maritime and long-haul road transport. In line with this, the annual global demand for biofuels is expected to grow about 30% by 2026. Additionally, the development of new fuels requires high quality products not only for the simple role of replacing hydrocarbons with biocomponents but also, and even more challenging, these are expected to perform efficiently on current and next-generation engines. Consistently, advanced engines employ technologies that lead to less fuel consumption, such as higher compression ratios and power densities, turbocharged engines (downsizing) and operation at lower speeds (downspeeding), which call for fuels with significantly higher knock resistance than commonly available today[1]. The octane number (ON) of a fuel is a non-equilibrium property that 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, mixing rules and Machine Learning (ML) techniques have been proposed, most of which fail when challenged to predict new data. The ON strongly depends on the molecular structure of species in the fuel and the interaction among them. In addition, the complexity increases when aromatic, olefinic and oxygenated compounds are present. Therefore, the development of a fuel and biofuel blend design tool is required to reduce the exploratory intensive laboratory work needed to assess the interaction between fuel blend properties and new candidate biofuels. To this end, in previous works, stochastic and deterministic optimization techniques (Particle Swarm Optimization and Levenberg-Marquardt, respectively) were combined for a rapid search for multicomponent surrogate mixtures that meet specific properties of the fuel. In this work, a new model to predict octane number is presented and integrated into the design tool, which already takes into account volatility properties of the fuel, such as Reid vapor pressure (RVP), distillation curve (DC) and the PIANOX classification. The accuracy and applicability of many ON models proposed in the open literature were evaluated concluding that, none of them are capable of accurately predicting ON for pure compounds and mixtures simultaneously. In the evaluation process, we compiled a database comprising the ON of 398 pure compounds and 448 mixtures. Since ML models have shown to be flexible enough to describe the complex antagonistic and synergistic effect of multicomponent mixtures [2], in this work, a new artificial neural network (ANN) model, based on supervised learning, is developed for the prediction of ON of pure components and mixtures. The ANN inputs are the concentration of functional groups (as defined in the Group Contribution with Association Equation of State, GCA-EOS) and other descriptors such as volatility properties and molecular size are also considered. The proposed approach allows not only to estimate ON for hypothetical blends, but also help to identify which descriptors, and to which extent, the different inputs affect the predictions. References1. R.L. McCormick, G., Fioroni, L. Fouts, E., Christensen, J., Yanowitz, E., Polikarpov, K., Albrecht, D.J., Gaspar, J. Gladden, A.G. (2017). SAE Int. J. Fuels Lubr. 10, 868.2. F. vom Lehn, B. Brosius, R., Broda, L. Cai, Pitsch H. (2020) Fuel. 281, 118772.