A graded filter has pores that vary in size from one end to the other — coarse at the inlet, fine at the outlet, or the reverse. Soil horizons are natural graded filters. Engineered water treatment systems are designed ones. The question is: which gradient is best?
Srinivasan and Gómez (arXiv:2603.20745) show that the answer depends entirely on which performance metric you choose. If you optimize for minimum outflux concentration — the cleanest possible output — one porosity gradient is optimal. If you optimize for maximum total adsorption — the most contaminant captured — a different gradient wins. The two objectives do not merely give different optima. They give qualitatively different designs: the direction of the porosity gradient can reverse depending on which metric you care about.
The mechanism is a coupling between fluid velocity and solute concentration that the standard theory ignores. Classical homogenization assumes incompressibility — the fluid velocity field is divergence-free. But in a graded medium where porosity varies spatially, the mixture of solvent and solute violates this assumption near the clogging limit. The authors derive a modified incompressibility condition from non-equilibrium thermodynamics that introduces a coupling parameter: the solute concentration affects the local fluid velocity, which affects how the solute is transported, which changes the concentration. This feedback loop is invisible under the standard solenoidal assumption and becomes dominant precisely when the filter matters most — near clogging, where every pore volume counts.
The optimal filter under the corrected model differs from the optimal filter under the classical model. The correction matters most when performance matters most.
The through-claim: when a system has multiple legitimate performance metrics, the optimal design is not robust to the choice of metric — it is a function of that choice. There is no “best” design, only a best design for a specified objective. This is not a philosophical point about values. It is a mathematical one: the optimization landscape itself changes shape when you change what you are optimizing, and near the system's limits, the landscapes can point in opposite directions. Anyone who presents an optimal design without specifying the objective has not solved the problem. They have hidden it.