Revealing the bridge from surface pressure to ventilated cavity dynamics via a pressure distribution-informed neural network

This study presents an integrated experimental and data-driven investigation of the instantaneous prediction of ventilated cavity morphology based on water tunnel measurements. The multiphase flow characteristics of ventilated cavities around an axisymmetric body are examined to elucidate the intrinsic correlation between the fluctuating surface pressure distribution and the cavity evolution process. A pressure distribution-based method is proposed to identify physically admissible intervals of cavity length from surface pressure measurements. Building upon this interval information, a pressure distribution-informed neural network (PDINN) model is developed to perform uncertainty-constrained prediction of cavity length under high-frequency shedding conditions. Furthermore, the influences of cavity-shedding scale and sensor arrangement density on the model prediction accuracy are systematically analyzed. The results indicate that the pressure near the cavity interface exhibits periodic alternation between positive and negative values, from which the effective cavity-length intervals of the cavity can be inferred. By integrating this interval information, the PDINN model achieves high-accuracy predictions. Moreover, the prediction accuracy is found to improve markedly with increasing shedding scale and sensor arrangement density. When the shedding scale is not smaller than the sensor arrangement density, PDINN achieves relative errors below 10% throughout all stages of cavity evolution within non-violation regions.