UC Irvine

In heat transfer problems, it is desirable to know when critical heat flux (CHF) has been reached. Exceeding CHF results in diminished heat transfer and can pose a significant risk since at this point not all superfluous energy can be removed. Equipment and boiling rigs are in danger of being damaged if CHF is exceeded for prolonged times.To prevent such events from happening, the temperatures of coolants in boiling rigs need to be known throughout time. Ideally, they should be monitored in real-time. While coolant temperatures can be measured using devices such as thermocouples, this approach is not always feasible in practice due to the small size of experimental boiling setups. When cooling computer chips, for instance, dimensions are on the millimetre scale. In addition, thermocouples only deliver point-wise temperature estimates and particularly temperatures in areas of high interest (e.g. near the surface of a computer chip) would have to be interpolated. This work examines the feasibility of solving part of this problem with physics-informed neural networks (PINNs). The trajectories of bubbles in boiling rigs are observable. As such, one can use these bubble velocities in conjunction with established physical laws for fluid flows and train neural models that can predict velocities in entire boiling rigs. Predicted velocities can then be used to predict temperature fields. The models in this work focus on predicting liquid velocities and show how PINNs can be used to recover hidden physical properties from simulated bubble observations.