Biofuels flash point modelling using neural networks

Currently, there is an intense effort to find new renewable energy resources able to reduce the dependence on fossil fuels. This search is motivated by several factors, such as environmental hazards caused by petroleum derivatives, price fluctuations, and primarily because fossil fuels are less sustainable when compared to other energy sources. For instance, in the Brazilian energy matrix, biodiesel and bioethanol are established renewable fuels. In addition, other biofuels, such as butanol, have shown potential as alternative energy resources, driving a growing demand for research into their safety and efficiency. To characterize the flammability of liquid fuels, the Flash Point (FP) is an important property and its assessment plays a relevant role in safety and combustion process design. However, its experimental measurement takes time, and resources, making the development of mathematical models for predicting this property a compelling approach. The usage of artificial neural networks (ANNs) for predicting thermodynamic properties has been explored in recent years, posing them as feasible data-driven alternatives to estimate FP values for a range of physical conditions. In this context, the main goal of this work is to evaluate the ANN predictive capacity of FP for biofuels and their respective blends. For this purpose, 456 experimental points from the literature were used for different systems involving esters, alcohols, and hydrocarbons, along with 34 newly acquired experimental points. Accordingly, a robust model was developed using an 80/20 train-test splitting strategy during a cross-validation loop. The model relied exclusively on three crucial input features: the mixture's average molar mass, the vapor pressure natural logarithm and the experimental method. The final architecture featured three hidden layers, with 16, 32 and 16 neurons, in which the Rectified Linear Unit (ReLU) activation function is employed. This model resulted in Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) values of 4.22 K and 3.09 K, respectively, for 98 unknown values, i.e., those not used during model training. Despite the simplicity, the model achieved a satisfactory level of accuracy, with the best and worst MAE values of 1.51 K and 3.63 K. The model also performed decently when compared to thermodynamical models that utilizes UNIFAC activity coefficient model. This fact demonstrates the approach's potential, considering that with almost 500 experimental points, the ANN algorithm obtained a generalized model for most of the analyzed data.