Collaborative Research: Explainable deep learning for cavitation control: from first-principles simulations to reinforcement-learning strategies

Cavitation occurs when small vapor bubbles form in a liquid as the pressure drops. When the bubbles move into regions of higher pressure, they collapse and can create noise, vibration, and damage in engineering systems. Cavitation affects many technologies, including ship propellers, water turbines, medical ultrasound, and drug delivery devices. Yet it remains difficult to predict and control because it arises from complex links between fluid motion, pressure changes, and phase transitions. These gaps in understanding can reduce efficiency, shorten equipment life, and slow progress in healthcare and biotechnology. This award will develop a new method to study and control cavitation by combining advanced physical models with modern artificial intelligence. The goal is to improve cavitation prediction tools and support safer, more efficient, and more sustainable engineering designs. The award also advances the national interest by strengthening U.S. leadership in computational science and engineering and supporting economic competitiveness in marine, energy, and biomedical technologies. Educational and outreach efforts will help train the future science and engineering workforce. This award will develop a framework by integrating physics and artificial intelligence for predicting and suppressing cavitation in turbulent flows. High-fidelity simulations based on first-principles thermodynamics will be used to model the formation and evolution of vapor bubbles. These simulations will generate detailed data linking fluid motion to cavitation events. Explainable-deep-learning methods will then be applied to identify the flow patterns that most strongly influence cavitation. The award will establish causal links between turbulence structures and cavitation onset by systematically quantifying which features of the flow promote cavitation inception. Building on this understanding, the project will develop adaptive-control strategies using deep reinforcement learning. These controllers will act on the flow in real time, targeting the specific structures responsible for cavitation. The expected outcomes include improved predictive capability for cavitation; interpretable control strategies grounded in physics, and broadly applicable tools for managing complex fluid systems. These advances will contribute to more reliable engineering systems and deeper understanding of multiphase-flow phenomena. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2535653 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Hector Gomez | Institution: Purdue University, WEST LAFAYETTE, IN | Award Amount: $270,522 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2535653 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2535653.html