A comprehensive survey of mathematical models for water engineering (arXiv:2512.16351) documents a field-wide shift: purely physics-based solvers are no longer sufficient for managing real hydraulic systems. The gap between idealized mathematical models and field applications requires hybrid methodologies combining classical CFD with data assimilation, uncertainty quantification, and machine learning.
The survey covers the full spectrum from depth-averaged shallow water equations to three-dimensional Navier-Stokes models, noting that practical challenges — source-term treatment, wetting and drying, turbulence modeling, free-surface representation — significantly degrade performance in real-world applications. These aren't numerical challenges; they're physical ones. A dam spillway with air entrainment, debris, and time-varying inflow conditions violates the assumptions of every standard turbulence model.
The surprising recognition: machine learning doesn't replace physics-based solvers — it fills the gaps between what physics predicts and what measurements show. Data assimilation corrects model state using observations. ML turbulence closures learn from high-fidelity simulations what low-fidelity models can't resolve. Uncertainty quantification wraps predictions in honest error bars that classical CFD typically omits.
This mirrors a broader pattern in computational engineering. The era of pure physics-based simulation is giving way to physics-informed hybrid methods, not because physics failed but because the real world contains more complexity than any single equation set can capture. The hybrid approach is pragmatic rather than elegant: it uses physics where physics works and data where physics doesn't, accepting theoretical inconsistency in exchange for practical accuracy.
The lesson: maturity in computational engineering isn't building better solvers — it's knowing where each solver stops being trustworthy and having something ready to take over.