Turbulence is a central challenge in classical physics and a critical barrier to accurate flow prediction in climate, aerospace, and energy systems. Despite the widespread reliance on Reynolds-averaged Navier-Stokes (RANS) solvers in industrial simulations, existing turbulence models lack the generalizability to handle diverse regimes, such as separation, secondary flows, and free-shear flows, without manual tuning or switching. We propose a unified data-driven turbulence modeling framework based on multi-objective learning. The goal is to achieve Pareto-optimal performance across heterogeneous flow datasets, each representing distinct mechanisms and quantities of interest. The resulting unified foundation model employs a parallel tensor basis neural network with automatic balancing and internal branching to adapt across flow regimes without explicit switching. The parallel architecture enables explicit regularization to promote model parsimony, while the tensor-basis formulation preserves physical symmetries. Trained on five representative flows, the model is evaluated on 27 test cases spanning attached, separated, and secondary flows, as well as two realistic three-dimensional flows of industrial relevance. It improves or matches the performance of the baseline k-ω model in all cases. For specific applications, we show that specialist models trained on tailored datasets can further improve accuracy in challenging configurations, such as three-dimensional diffuser flows common in gas turbine aerodynamics, which exhibit simultaneous separation and secondary flows. These results demonstrate that a generalized, deployable turbulence model unifying multiple flow mechanisms within a single architecture is achievable. This work marks significant progress toward unified turbulence modeling for scientific and industrial applications.
Liu et al. (Sun,) studied this question.