Heat Exchangers Fouling Theory

A stochastic fouling theory for heat exchangers - why one asymptotic fouling curve can hide different removal mechanisms.

Stochastic fouling theory

One fouling curve, four mechanisms

A fouling curve is usually read as the signature of a mechanism: a straight line as deposition alone, a plateau as shear removal. It is not. Physically different mechanisms draw exactly the same mean curve, and what tells them apart is the scatter around it.

What fouling is

This work is about heat exchangers, which pass heat from a hot fluid to a cold one through a metal wall, usually the walls of a bundle of tubes. Fouling is unwanted material building up on those walls: scale from dissolved salts in cooling water, silt and corrosion products, coke and gums formed by reactions in hot crude oil, biofilms. Deposits conduct heat poorly, so they add a resistance to the heat flow through the wall, the fouling resistance, defined through the overall heat-transfer coefficient \(U\) of the fouled and the clean exchanger:

\[R_f=\frac{1}{U}-\frac{1}{U_\mathrm{clean}}.\]

Small deposits matter. A layer \(\delta=0.1\) mm thick with a conductivity \(k_f=0.5\ \mathrm{W\,m^{-1}\,K^{-1}}\) adds \(R_f=\delta/k_f=2\times10^{-4}\ \mathrm{m^2\,K\,W^{-1}}\); on an exchanger with \(U=400\ \mathrm{W\,m^{-2}\,K^{-1}}\) that cuts the heat transferred by about 7 %. The lost heat is made up by burning extra fuel, and once the resistance reaches a limit the exchanger is taken offline and cleaned. The fouling resistance is not measured directly: operators reconstruct it from temperatures and flow rates.

The typical fouling curves

Plotted against time, fouling records tend to fall into a few familiar shapes, and each has a conventional reading.

induction time fouling resistance Rf linear falling-rate asymptotic sawtooth
Idealised shapes of fouling records after an induction period. Real records are noisier and often mix these shapes. The two drawn darker, linear and asymptotic, are the classical laws studied below.

The habit, in short, is to read the shape as the mechanism. This work shows why that does not work.

It does so by describing fouling as a stochastic process. Instead of one smooth deposition rate, material reaches the wall in random amounts at random times, and removal acts on whatever is there. The classical curves then describe the average over many such records, and the scatter of each real record around that average stops being a nuisance and becomes information.

Same straight line, different deposition

Start with the simplest record: deposition and no removal. The mean fouling resistance then grows in a straight line, the linear fouling law. Underneath that line, material can reach the wall in very different ways: in discrete bursts with quiet spells in between, as a dense deposition of mostly tiny events, or as a deposition that every so often delivers a much larger lump. Here the three are set to the same mean rate and the same variance. Pick one.

The mean and the band never move. Only the texture of the records does.

mean, identical for all three mean ± 2 standard deviations, identical too three simulated records

A straight fouling line says nothing about how the deposit is built up, and neither does its spread. The three laws differ only in how lopsided the deposit of a single period is, its skewness. That difference comes back at the end of the page.

The maths

1. The model

Split time into sampling periods of length \(\tau_0\), one reading per period (three days, say). In each period the deposit grows by a random amount \(\Delta D\ge0\) that can never be negative. As long as operating conditions do not change, we assume that the amounts deposited in different periods are independent of each other and all drawn from the same distribution. Mathematicians call a process built this way a Lévy subordinator; the word only means that it can go up but never down.

2. Why the mean is a straight line

After \(n\) periods the deposit is the sum of \(n\) increments, \(D=\Delta D_1+\Delta D_2+\dots+\Delta D_n\). The average of a sum is the sum of the averages, so

\[\mathbb{E}[D(t)]=n\,\mathbb{E}[\Delta D]=m\,t,\]

where \(m\) is the mean deposition rate and \(t=n\tau_0\). Nothing in this step depends on what the distribution of \(\Delta D\) looks like: bursts, fine deposition and lumps all give the same line. That is why the linear law cannot tell them apart.

3. Why the spread is the same too

Because the increments are independent, their variances also add: \(\operatorname{Var}[D]=n\operatorname{Var}[\Delta D]\). We describe how irregular one period's deposit is by its coefficient of variation, the standard deviation of \(\Delta D\) divided by its mean, \(c=\sigma_{\Delta D}/\mathbb{E}[\Delta D]\). Then

\[\operatorname{Var}[D(t)]=c^2m^2\tau_0\,t,\]

so the band of \(\pm2\) standard deviations widens in proportion to \(\sqrt{t}\). Giving the three laws the same \(m\) and the same \(c\) makes both the line and the band identical.

4. What does differ: the skewness

Skewness measures how lopsided a distribution is: zero for a symmetric bell curve, positive when there is a long tail of large values. For the deposit of one period it is \(\gamma_1=1.5\,c\) for burst deposition, \(2c\) for fine deposition and \(3c\) for deposition with lumps; the larger the number, the more of the deposit arrives in rare, large events. Over a window of \(k\) periods the skewness shrinks as \(1/\sqrt{k}\), because adding many independent pieces always tends towards a bell curve (the central limit theorem). That fading is what makes the last panel hard.

5. Quiet periods

Only burst deposition delivers material in separate events, so a whole period can pass with nothing deposited. The chance of that is \(e^{-2/c^2}\), about 61 % here. The other two laws deposit a little in every period, however short.

Simulated with \(m=1\), \(c=2\), \(\tau_0=1\), as in Fig. 2 of the paper.

Same curve, different mechanism

Now add removal. The straight line bends over to a plateau, the Kern–Seaton curve, and there are several physically different ways to get there. Pick a mechanism. The solid line, the mean fouling resistance, does not move. The individual records and the band they fill do.

The mean never moves. Watch the scatter on the plateau.

mean, identical for all four mean ± 2 standard deviations three simulated records

Fitting the curve returns a plateau and a time constant, and nothing else. A Kern–Seaton fit is not evidence of shear removal.

The maths

1. The Kern–Seaton equation

Kern and Seaton balanced two rates: deposition at a constant rate \(m\), and removal proportional to what is already on the wall, at a rate \(\psi R\):

\[\frac{dR}{dt}=m-\psi R.\]

Its solution, starting from a clean surface, is \(R(t)=R_\infty\left(1-e^{-\psi t}\right)\). The resistance levels off at the plateau \(R_\infty=m/\psi\), where deposition and removal balance, and gets there in a few multiples of the time constant \(1/\psi\).

2. Why all four mechanisms give that same curve

Each mechanism takes away, on average, the same fraction \(\psi\) of the deposit per unit time; they differ only in how:

  • Continuous erosion shaves every piece of deposit a little all the time.
  • Per-element detachment leaves each piece intact, then removes it whole; any piece goes with probability \(\psi\) per unit time.
  • Sloughing removes a fraction \(\eta\) of the whole layer at random moments, \(r_s\) times per unit time on average, so \(\psi=r_s\eta\).
  • Surface saturation removes nothing; instead attachment slows down as the free surface fills up. In the averaged equation this enters exactly like removal.

Averaging over many records turns each of them into the Kern–Seaton equation,

\[\frac{d\,\mathbb{E}[R]}{dt}=m-\psi\,\mathbb{E}[R].\]

The mean curve is therefore the same, and fitting it returns only \(m\) and \(\psi\) (equivalently, the plateau and the time constant).

3. Where they differ: the scatter on the plateau

Let \(s^2=c^2m^2\tau_0\) be the variance that deposition adds per unit time (step 3 of the previous panel). Once the plateau is reached, the variance of \(R\) settles at a value that depends on the mechanism:

  • Continuous erosion: \(\dfrac{s^2}{2\psi}\). Every deposited piece shrinks gradually, so its contribution to the scatter fades smoothly.
  • Per-element detachment: \(\dfrac{s^2}{\psi}\), twice as much. A piece keeps its full size until the moment it leaves, so the scatter it causes lasts longer.
  • Sloughing: \(\dfrac{s^2}{(2-\eta)\psi}+\dfrac{\eta R_\infty^2}{2-\eta}\). The second term does not involve deposition at all: it comes from the random timing of the drops, the sawtooth. With \(\eta=0.5\) it equals \(R_\infty^2/3\), a scatter of almost 60 % of the plateau, even if deposition were perfectly smooth.
  • Surface saturation: the variance is \(\ell R_\infty\,\theta(1-\theta)\), where \(\theta\) is the fraction of the surface covered and \(\ell\) the typical size of one deposition event. It peaks at half coverage and is zero once the surface is full.

4. Reading it from a record

Fit the Kern–Seaton curve, subtract it, and compute the variance of what is left on the plateau. Subtract the variance of the sensor noise. A leftover that is a sizeable fraction of \(R_\infty^2\) points to sloughing; a leftover close to zero is consistent with continuous erosion or saturation, which are then told apart by how the scatter evolves before the plateau.

Simulated with \(m=1\), \(c=1\), \(\psi=0.2\), \(\eta=0.5\), as in Fig. 3 of the paper. The tiles show the plateau variance of each mechanism for these values.

Can you measure the deposition law?

Deposition laws with the same mean and variance differ only in the skewness of their increments. Measurement noise hides it at short windows, and averaging erases it at long ones, because the sum of many periods tends to the same Gaussian shape whatever the law (the central limit theorem). Set your instrument and see how long a record you would need.

Noise forces a longer window, and a longer window erases the shape.

skewness left in your record noise-free record
Record needed to tell Gamma from inverse Gaussian
Signal-to-noise ratio, one period
Best window keeps

The record needed grows as the fourth power of the measurement noise. Improving the instrument helps far more than waiting.

The maths

1. What is actually measured

The fouling resistance is not measured directly: it is calculated from temperatures and flow rates through the overall heat-transfer coefficient, \(R=1/U-1/U_\mathrm{clean}\). If \(U\) is known only to within a relative error \(\epsilon\) (first slider), each reading of \(R\) carries an error of about \(\epsilon/U\). In the same units, one sampling period of fouling changes \(U\) by a relative amount \(\zeta\) (second slider).

2. Signal and noise for one period

The deposition statistics are in the increments, the difference between two readings. That difference contains two things: the deposit laid down in between (the signal) and the errors of the two readings (the noise). Comparing their variances gives the signal-to-noise ratio of one sampling period,

\[\mathrm{SNR}=\frac{\text{variance of one period's deposit}}{\text{variance of the difference of two errors}}=\frac{c^2}{2}\left(\frac{\zeta}{\epsilon}\right)^{2}.\]

With the default settings (\(\epsilon=1\%\), \(\zeta=0.4\%\), \(c=2\)) this is 0.32: the noise is about three times larger than the signal.

3. Averaging helps the noise but hurts the skewness

The obvious remedy is to take increments over a longer window of \(k\) periods. The deposit then grows \(k\) times while the noise of the two end readings stays the same, so the window's signal-to-noise ratio becomes \(k\,\mathrm{SNR}\). But two things now eat into the skewness that distinguishes the laws:

  • Aggregation: the skewness of a \(k\)-period deposit falls as \(1/\sqrt{k}\) (step 4 of the first panel).
  • Dilution: the noise is symmetric, so it adds spread but no skewness, and divides the skewness by a further factor. The fraction left is \(\left(\dfrac{k\,\mathrm{SNR}}{1+k\,\mathrm{SNR}}\right)^{3/2}\), close to zero when noise dominates and close to one when signal dominates.

Multiplying the two gives the curve in the chart, the fraction of the one-period skewness that survives:

\[\text{retained}=\frac{1}{\sqrt{k}}\left(\frac{k\,\mathrm{SNR}}{1+k\,\mathrm{SNR}}\right)^{3/2}.\]

It rises while averaging still beats down the noise, then falls as aggregation takes over. The best window is where the window's signal-to-noise ratio equals 2, i.e. \(k=2/\mathrm{SNR}\) (the orange dot), and even there only \(\tfrac{2}{3\sqrt{3}}\sqrt{\mathrm{SNR}}\approx0.385\sqrt{\mathrm{SNR}}\) of the skewness is left. With an SNR of 0.01 that is under 4 %.

4. How long a record you need

Skewness estimated from \(n\) independent windows has a statistical uncertainty of about \(\sqrt{6/n}\). Fine deposition and deposition with lumps differ in one-period skewness by \(c\) (from \(2c\) to \(3c\)). Demanding that the surviving difference be at least two uncertainties, and choosing the window that best trades retained skewness against the number of windows, gives a record of

\[N\approx\frac{324}{c^2\,\mathrm{SNR}^2}=\frac{1296}{c^6}\left(\frac{\epsilon}{\zeta}\right)^{4}\]

sampling periods, converted to years here at three days per period. The record grows with the fourth power of the sensor error: halving \(\epsilon\) shortens it sixteen-fold, which is why a better instrument helps far more than a longer wait.

5. Caveats

The estimate counts only random reading errors that are independent from one reading to the next. Errors that drift or are correlated in time make things worse, not better. A constant calibration offset cancels in the increments and does not matter. When the signal-to-noise ratio exceeds \(1/c^2\) the formula is outside its range of accuracy and the number is indicative only.

What this means in practice

  • A fitted fouling curve gives a deposition rate and, if it levels off, the time constant of the approach to the plateau. It does not identify the mechanism.
  • To identify the removal mechanism, watch the scatter on the plateau. Sloughing keeps a large variance even when deposition is smooth.
  • The deposition law is out of reach for a thermally reconstructed resistance at realistic accuracy.
Read the paper on arXiv S. Casolo, Classical heat exchanger fouling-resistance laws as first-moment identities of additive stochastic deposition, arXiv:2611.22808 [eess.SY] (2026).