A theory-informed machine learning approach for cryogenic cavitation prediction

Inferring cryogenic cavitation features from the boundary conditions (BCs) remains a challenge due to the nonlinear thermal effects. This paper aims to build a fast model for cryogenic cavitation prediction from the BCs. Different from the traditional numerical solvers and conventional physics-informed neural networks, the approach can realize near real-time inference as the BCs change without a recalculating or retraining process. The model is based on the fusion of simple theories and neural network. It utilizes theories such as the B-factor theory to construct a physical module, quickly inferring hidden physical features from the BCs. These features represent the local and global cavitation intensity and thermal effect, which are treated as functions of location x. Then, a neural operator builds the mapping between these features and target functions (local pressure coefficient or temperature depression). The model is trained and validated based on the experimental measurements by Hord for liquid nitrogen and hydrogen. Effects of the physical module and training dataset size are investigated in terms of prediction errors. It is validated that the model can learn hidden knowledge from a small amount of experimental data and has considerable accuracy for new BCs and locations. In addition, preliminary studies show that it has the potential for cavitation prediction in unseen cryogenic liquids or over new geometries without retraining. The work highlights the potential of merging simple physical models and neural networks together for cryogenic cavitation prediction.