Thermodynamics-informed neural network for recovering supercritical fluid thermophysical information from turbulent velocity data
Recent research has highlighted the potential of supercritical fluids under high-pressure transcritical conditions to achieve microconfined turbulence as a result of the thermophysical properties they exhibit in the vicinity of the pseudo-boiling region. This has led to increased interest in underst...
| Autores: | , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2023 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/393209 |
| Acceso en línea: | https://hdl.handle.net/2117/393209 https://dx.doi.org/10.1016/j.ijft.2023.100448 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep learning Supercritical fluids Thermodynamics-informed neural network Turbulent flow Aprenentatge profund Fluids supercrítics Àrees temàtiques de la UPC::Enginyeria mecànica::Mecànica de fluids |
| Sumario: | Recent research has highlighted the potential of supercritical fluids under high-pressure transcritical conditions to achieve microconfined turbulence as a result of the thermophysical properties they exhibit in the vicinity of the pseudo-boiling region. This has led to increased interest in understanding their hybrid thermophysical properties when operating near the pseudo-boiling transitioning region. However, despite the potential benefits of microfluidic systems working under transcritical conditions, limited experimental data is available due to the inherent challenges of performing experiments at high-pressure conditions. In addition, traditional experimental methods, such as particle image velocimetry and particle tracking velocimetry, are inadequate for measuring thermophysical properties under such conditions, since they are primarily designed for velocity-related data acquisition. In this regard, this work introduces an efficient thermodynamics-informed neural network framework for reconstructing thermophysical information from velocity data in high-pressure turbulent transcritical regimes. The proposed model incorporates thermophysical constraints through a thermodynamics-informed loss function consisting of the residual of the real-gas equation of state and integrates boundary conditions into the network’s architecture to ensure their satisfaction. The performance of the proposed framework is evaluated through the analysis of two test cases and compared against non-physically informed models. The results demonstrate the superior accuracy, robustness, and satisfaction of physical constraints achieved by the proposed model, as well as its ability to reconstruct averaged thermophysical profiles and preserve bulk quantities with a relative error reduction of approximately 2×. In addition, the physically-consistent predictions provided by the model enable a more accurate reconstruction of dependent thermophysical properties. |
|---|