Physics-Informed Machine Learning for Predicting Oxygen Transfer and Mixing Efficiency in Vortex-Dominated Flows
Sanjith Balamurali
11/08/2026
Vortex shedding behind bluff bodies enhances oxygen transfer and mixing in aquatic systems through increased turbulence, expanded interfacial area, and enhanced boundary layer renewal. However, accurately predicting these mass transfer enhancements across varying Reynolds numbers and geometries remains challenging due to complex nonlinear coupling between fluid dynamics and transport phenomena. This study develops a physics-informed machine learning framework to predict oxygen transfer coefficient and mixing efficiency from computational fluid dynamics simulations. Using 51 vortex wake simulations spanning Reynolds numbers from 1,000 to 10,000, we extract 111 hydrodynamic features capturing velocity statistics, vorticity characteristics, and wake geometry at six downstream locations. Physics-based proxy targets are constructed using surface renewal theory incorporating vortex shedding frequency, wake width, and turbulence metrics; because these targets are deterministic, closed-form functions of a subset of the same input features (principally the velocity parameter U*, wake width at the 6D location, and maximum vorticity at the 6D location), this study is best understood as a test of how efficiently machine learning can recover a known physical relationship rather than as an independent discovery of new physics. Random Forest and Gradient Boosting models achieve test-set accuracy exceeding 96% for both targets, substantially outperforming a linear regression baseline using velocity alone (R-squared of 0.88) and a classical empirical correlation (R-squared of 0.76). Feature importance analysis reveals that flow velocity accounts for 89% of predictive contribution, with wake geometry contributing 2.3% and vorticity characteristics under 2% individually; this outcome is expected given that vortex shedding frequency, and therefore the oxygen transfer proxy, is defined as a direct function of velocity in the target construction. Residual analysis demonstrates unbiased errors with no systematic patterns across Reynolds regimes, though the small test set (11 simulations) limits the precision of these estimates. The framework demonstrates that machine learning can efficiently approximate the intended physical formula, offering a computationally efficient tool, with millisecond-scale predictions versus hours for full CFD simulations, for rapid design screening in aeration systems, wastewater treatment facilities, and stream restoration projects pending experimental validation.