Physics-guided discrete cosine transform neural network for fast prediction of cavitation and pressure fluctuations in micro axial piston pumps

Evaluating the coupled evolution of internal cavitation and pressure fluctuation in micro axial piston pumps (MAPPs) across diverse operating conditions remains computationally expensive for conventional computational fluid dynamics (CFD). To reduce this cost within a defined operating envelope, a deep-learning-based surrogate model that combines discrete cosine transform-based frequency-domain representation with time-domain physical regularization is introduced for rapid prediction of key transient responses. Driven by CFD reference data, nonlinear mappings between operating parameters and key physical quantities, including cavitation volume, piston chamber pressure, and outlet pressure, are established by the proposed model. The results show that the principal dynamic features of the target variables can be reconstructed with good agreement with CFD reference results under the tested operating conditions. A conventional CFD simulation requires approximately 3 h for a single operating condition, whereas the proposed model requires 35 s for training and 0.7 s for a single inference, yielding an inference speedup of about 15 429×. These results indicate that the proposed method can substantially reduce computational cost for CFD-based multi-condition assessment of MAPPs under the fixed geometry, fluid-property, gas-content, temperature, and outlet-boundary conditions considered in this study.