Modeling flash points of biofuels using thermodynamically consistent neural networks

The flash point (FP) is an essential property for assessing flammability and ensuring safety in combustion processes. However, its experimental measurement is resource-intensive and data in the literature remains limited, especially for biofuel mixtures. To address this, predictive modeling has emerged as a promising alternative. This study investigates the capability of thermodynamically consistent neural network models to estimate FP, as well as hybrid approach that embeds a neural network constrained by Gibbs-Duhem equation within a thermodynamic model. The performance of these models was compared with that of a purely data-driven model and a widely used thermodynamic approach. The evaluation was conducted using a dataset comprising binary mixtures of 1-butanol and fatty acid ethyl esters (FAEEs). The data-driven and physically constrained neural network approaches achieved RMSE (Root Mean Squared Error) values of 2.518 K, 1.975 K, 2.798 K, and 2.470 K, while the thermodynamic model using NRTL providing a RMSE of 0.587 K, as expected given that its binary interaction parameters were fitted to the experimental FP data. In addition, incorporating physical constraints into the neural network models for FP prediction did not improve RMSE performance compared to the purely data-driven model, despite achieving improved consistency for the embedded physics equations as expected.