Computational fluid dynamics solves the Navier–Stokes equations numerically over a discretised tank, revealing the velocity, turbulence and concentration fields that empirical loading rates conceal. In water treatment it earns its cost on exactly one class of problem: where the flow field, not the chemistry or biology, is what limits performance — short-circuiting clarifiers, poorly baffled contact tanks, uneven flow splits and flotation contact zones.

What CFD is actually for in water treatment

Conventional design uses lumped parameters: surface overflow rate, hydraulic retention time, velocity gradient G. Each assumes an idealised flow pattern — perfect plug flow or perfect mixing — that no real tank achieves. CFD replaces the assumption with a computed field.

That is worth doing when the deviation from ideality is what limits the process. Four situations recur:

ProblemWhat the lumped model missesTypical CFD finding
Clarifier carry-over at nominal loadingDensity currents driven by the solids concentration differenceA bottom current reaching the wall and rising, short-circuiting 30–50% of the volume
Contact tank failing CT despite adequate HRTBaffle factor: t10/T may be 0.3 rather than the assumed 0.7Corner recirculation and a fast core stream; baffle changes recover most of it
Uneven flow split to parallel unitsMomentum effects in the distribution chamber10–25% maldistribution, derating the whole works to the worst-loaded stream
Flotation or flocculation underperformanceLocal G and bubble or floc contact time distributionDead zones and high-shear zones coexisting in the same tank

Conversely, CFD is poor value where the limiting step is a reaction rate, an adsorption isotherm or a biological growth rate. No mesh refinement will tell you the minimum sludge age for nitrification; that is the domain of process modelling.

The governing equations and why turbulence must be modelled

The flow of an incompressible Newtonian fluid is fully described by conservation of mass and momentum:

∇·u = 0
ρ(∂u/∂t + u·∇u) = −∇p + μ∇²u + ρg + S
where u = velocity vector, p = pressure, ρ = density, μ = dynamic viscosity and S = source terms (buoyancy, drag from a dispersed phase, porous resistance).

Water treatment flows are turbulent at any practical scale. The Reynolds number for a 4 m deep clarifier at a mean velocity of even 2 mm/s is Re = ρuL/μ = 1000 × 0.002 × 4 / 0.001 = 8,000, and inlet jets and channels run into the 105–106 range. Resolving all turbulent scales directly (DNS) would require a mesh of order Re9/4 cells — astronomically infeasible — so turbulence is modelled rather than resolved.

The standard approach is Reynolds averaging (RANS): decompose each variable into a mean and a fluctuation, average the equations, and close the resulting Reynolds stress term with a turbulence model.

ClosureCharacterUse in water treatment
Standard k–εRobust, cheap, poor in strong adverse pressure gradients and swirlThe default for tanks and channels; adequate for most clarifier work
Realizable / RNG k–εBetter for swirl, separation and round jetsVortex grit chambers, inlet jets, mixing zones
k–ω SSTAccurate near walls and in adverse gradientsNozzles, pump suctions, screens, wall-bounded detail
LESResolves large eddies; expensive, transient by natureResearch on mixing and floc breakage; rarely justified in design

Buoyancy must not be omitted. In a clarifier, mixed liquor at 3,500 mg/L is roughly 0.1–0.2% denser than the supernatant; that tiny density difference is enough to drive a bottom density current at several times the mean velocity, and it is the primary cause of clarifier failure at nominal loading rates. A CFD model that neglects the buoyancy source term will predict an idealised flow field that does not exist.

Multiphase methods: choosing the right representation

Water treatment is rarely single-phase. The choice of multiphase model is the single largest determinant of both accuracy and cost.

  • Eulerian–Eulerian (two-fluid). Both phases treated as interpenetrating continua with a volume fraction field. The right choice for high dispersed-phase fractions: sludge blankets, dense bubble plumes, fluidised media. Requires drag, lift and turbulent dispersion closures.
  • Mixture / drift-flux. A simplified two-fluid formulation solving one momentum equation with an algebraic slip velocity. Cheap and usually sufficient for settling tanks where the solids follow the water closely.
  • Lagrangian particle tracking (DPM). Individual particles or parcels tracked through the computed flow field. The correct method for UV dose distribution, where the quantity of interest is the distribution of exposure across trajectories, not its mean. Valid only at low volume fractions.
  • Volume of fluid (VOF). Tracks a sharp interface. Used for weirs, free-surface launders, hydraulic jumps and tank filling.
  • Population balance (PBM). Coupled to the above to track a size distribution that evolves by aggregation and breakage — essential for genuine flocculation modelling and for bubble coalescence in flotation.
Stokes number decides whether particles follow the flow:
St = τpu/L = (ρpdp2/18μ)·(u/L)
St « 1: particles follow streamlines, and a mixture model suffices. St ≳ 1: particles have independent inertia and must be tracked. For a 100 µm floc at Δρ = 50 kg/m³, τp ≈ 2.8×10−5 s — St is minute, so flocs follow the water and settle only through the slip velocity. That is why clarifier CFD is really a density-current problem.

Meshing, discretisation and the sins that invalidate a model

A CFD result is a numerical approximation, and the approximation error must be quantified rather than hoped away.

  • Mesh quality. Aim for low skewness (<0.85) and reasonable aspect ratios; resolve boundary layers with prism layers at a wall-function-appropriate y+ (typically 30–300 for standard wall functions, y+ < 5 for low-Reynolds models).
  • Discretisation scheme. First-order upwind is stable and wrong: numerical diffusion smears exactly the gradients that matter, and it will make a short-circuiting tank look well mixed. Converge on first order if necessary, then finish on second order.
  • Convergence. Residuals alone are insufficient. Monitor a physical quantity — outlet concentration, t10, a flow split — and require it to be steady, with global mass and species imbalances below about 0.1%.
  • Transient stability. For time-accurate work, keep the Courant number Co = uΔt/Δx at or below 1 for explicit schemes; implicit schemes tolerate more but lose temporal accuracy.
Grid Convergence Index (Roache):
p = ln[(f3−f2)/(f2−f1)] / ln r,  GCIfine = Fs|(f2−f1)/f1| / (rp−1)
with Fs = 1.25 for three-mesh studies and r the refinement ratio.

Worked illustration. Three meshes at r = 2 give a predicted removal efficiency of f1 = 0.912 (fine), f2 = 0.905, f3 = 0.885 (coarse). Then p = ln(0.020/0.007)/ln 2 = 1.050/0.693 = 1.52, and GCIfine = 1.25 × (0.007/0.912)/(21.52−1) = 1.25 × 0.00768/1.858 = 0.52%. The discretisation uncertainty on the fine mesh is therefore about half a percentage point — small enough to report the result as mesh-independent. Without that calculation, a stated efficiency of 91.2% is an unquantified claim.

Validation: the tracer test is the arbiter

A CFD model must be validated against physical measurement, and the most accessible validation in water treatment is a tracer residence-time distribution. Inject a conservative tracer, measure the outlet response, and compare with the simulated response from a passive scalar released identically.

The comparison should be made on RTD metrics, not on visual similarity of curves:

  • t10 — time for 10% of tracer to appear; the basis of the disinfection baffle factor.
  • Mean residence time t̄ = ∫tE(t)dt, which must equal V/Q if the model conserves mass and the tank has no dead volume.
  • Variance σ² and the Morrill index t90/t10, which characterise dispersion.
  • Dead volume fraction inferred from the shortfall of t̄ against V/Q.

Agreement within roughly 10–15% on t10 and mean residence time is a reasonable acceptance criterion for design work. The interpretation of those metrics is developed in our guide to residence time distribution and tracer testing. Where a tracer test is impossible, validate against measurable proxies: sludge blanket level, measured flow splits, velocity profiles from an ADV, or dye visualisation at model scale.

The professional standard. A CFD study reported without a mesh-independence statement and without validation against measurement is an illustration, not an engineering calculation. Say which it is.

Worked application: diagnosing a clarifier that fails at design loading

A 30 m diameter secondary clarifier carries over solids at a surface overflow rate of 0.6 m/h, comfortably within the conventional 1.0–1.3 m/h design range. The lumped model says it should work. CFD explains why it does not.

  1. Set up. Axisymmetric or 30-degree sector mesh; mixture multiphase with a settling velocity function fitted to measured zone-settling data; buoyancy active; realizable k–ε.
  2. Result. Mixed liquor at 3,800 mg/L entering the centre well is roughly 0.15% denser than the tank contents. It plunges, forms a radial bottom density current travelling at 8–12 mm/s — some fifteen times the nominal overflow velocity — reaches the wall, and rises as a wall jet directly into the effluent launder.
  3. Diagnosis. The effective settling volume is perhaps half the geometric volume; the local upflow velocity beneath the launder is several times the design overflow rate; and the solids escaping are not a settling failure but a hydraulic one.
  4. Interventions tested in the model. A deeper, larger-diameter flocculating centre well to dissipate the inlet momentum; a peripheral baffle (a Stamford or McKinney baffle) to deflect the wall jet inward; and an inboard launder position.
  5. Outcome. The combination typically recovers a large fraction of the lost effective volume, with predicted effluent solids falling by half or more — for the cost of a baffle rather than a new tank.

The economics are the point. A modelling study costing a small fraction of a new clarifier can convert a capacity problem into a baffle installation. The same reasoning applies to flotation tank hydraulics, where CFD is used to keep the contact zone in genuine plug flow and to avoid short-circuiting the white water blanket — the basis of CFD-led optimisation of flotation plant.

Where CFD misleads, and how to keep it honest

The failure modes of CFD in this sector are well known and almost entirely avoidable.

  • Wrong boundary conditions. A uniform inlet velocity where the real inlet is a pipe with swirl and a developing profile will produce a plausible and wrong answer. Model enough upstream geometry to develop the flow.
  • Rheology ignored. Sludge above roughly 2–3% dry solids is strongly non-Newtonian — shear-thinning with a yield stress. Treating it as water overpredicts mixing everywhere in a digester or thickener.
  • Steady state assumed for an unsteady process. Flotation, flocculation and storm response are all inherently transient; a steady solution may be a numerical artefact of an oscillating physical flow.
  • Chemistry pasted on. A first-order decay rate applied to a converged flow field is only as good as that rate constant; the CFD adds precision to the transport, not to the kinetics.
  • Colourful fluid dynamics. Contour plots persuade far beyond their evidential value. Insist on numbers — t10, flow split percentages, predicted effluent concentration — with an uncertainty attached.

Used with that discipline, CFD is one of the few tools that can retire a capital scheme in favour of a modification. Used without it, it produces attractive pictures of a tank that does not exist.

Frequently asked questions

When is CFD worth doing on a treatment plant?

When the flow field is what limits performance: clarifiers short-circuiting through density currents, contact tanks with poor baffle factors, uneven splits to parallel streams, and flotation or flocculation zones. It adds little where the limit is a reaction rate, an isotherm or a biological growth rate.

Which turbulence model should be used?

Standard or realizable k-epsilon for most tanks and channels, k-omega SST where near-wall accuracy or adverse pressure gradients matter such as nozzles and pump suctions. Large eddy simulation is generally too expensive to justify in routine design work.

How is a CFD model validated?

Preferably against a tracer residence-time distribution, comparing t10, mean residence time and variance rather than curve appearance, with agreement of roughly 10 to 15 per cent as an acceptance criterion. Sludge blanket level, measured flow splits and velocity profiles are useful secondary checks.

What is the grid convergence index?

A formal estimate of discretisation uncertainty derived from three systematically refined meshes using Richardson extrapolation. It converts the qualitative claim of mesh independence into a percentage uncertainty on the reported result, and should accompany any CFD figure used for design.

Why do clarifiers fail at loading rates that should work?

Because mixed liquor is slightly denser than the tank contents, and that fraction of a per cent drives a bottom density current at many times the mean velocity. It rises at the wall into the launder, short-circuiting a large part of the volume, so the failure is hydraulic rather than a settling deficiency.

Can CFD replace a pilot trial?

No. CFD resolves transport; a pilot resolves treatability. They are complementary: use the pilot to establish the chemistry and achievable removal, and CFD to make sure the full-scale tank delivers the hydraulic conditions the pilot enjoyed.

Sources & further reading