Ruiz-Botella, M, Eusébio, T. M., Sales-Pardo, M, Guimerà, R, Rodriguez-Reartes, S. B., Llovell, F.
Chem. Eng. J. 550 , 183129 (2026).

Download PDF or Visit journal

The search for working fluids that balance thermodynamic performance, safety, and environmental sustainability remains a pressing challenge for refrigeration and heat pump technologies. Here, we present a hybrid computational workflow combining deep generative learning with classical thermodynamic modelling to identify novel working fluid candidates. Using a graph-based diffusion model, we generate and screen molecules against a quantitative physicochemical profile derived from known working fluids. We characterize five selected candidates in depth through molecular dynamics simulations, the polar soft-SAFT EoS, COSMO-RS calculations, and two neural network predictors for flammability and global warming potential (GWP). We find that the predicted thermophysical and transport properties of the candidates fall within the range of established hydrofluorocarbons. Benchmarking against R-245fa and R-134a identifies a new compound (3,3,3-trifluoro-2-methylpropene) as the most promising candidate for high-temperature heat pump applications, and confirms another (1,1,1,2,2-pentafluoropropane) as a viable refrigerant for vapour compression refrigeration cycles. Our work establishes a scalable framework for the rational, data-driven discovery of next-generation working fluids.