Machine learning for fluid thermodynamics. State of the art and open challenges

This paper presents a critical survey on adopting machine learning in solving complex real fluid thermodynamics problems. After reviewing the primary computational machine learning frameworks employed in thermodynamic modelling, we have analysed current research with a particular emphasis on properly estimating gas and liquid properties, vapour-liquid equilibrium, and supercritical fluids, focusing on pure gases. While ML offers a powerful paradigm for augmenting or even replacing traditional methods, its application faces significant open challenges. Key issues include the persistent trade-off between model accuracy and computational efficiency, the difficulty in capturing highly non-linear behaviour, especially near critical points or under extreme conditions, and the pervasive problem of data scarcity. We conclude the paper by introducing the main datasets available for thermodynamic property computation, such as the results of the GERG2008 project, and others relevant to turbomachinery applications. This survey provides a unified perspective on machine learning architectures used in thermodynamics and identifies open challenges and potential future advancements for enhancing predictive accuracy and efficiency while reducing execution time.