5.

RTD TREATMENT AND MODELING

5.1. CALCULATION OF MOMENTS

The analysis of the residence time distribution data depends upon the specific aim for which the
radiotracer experiment has been carried out. The simples RTD data treatment is the calculation of
moments. Moments are used to characterize the RTD functions in terms of statistical parameters such as
mean residence time and standard deviation. The moments around the origin are defined as:

M i   t i E t dt
0
Where, i = 0, 1, 2, 3, ....

The zeroth moment of the normalized RTD gives the area under the curve, which is equal to unity:

M 0   E t dt  1
0
Mean residence time (MRT) is equal to the first moment:

M 1  t   t E t dt
0

The “spread” in the RTD is characterized by standard deviation (σ) or variance (σ2):


 2  M 2  M 12   t  t 2 E t dt
0
where

M 2   t 2 E t dt is the second moment around the origin.
0

Higher order moments allow calculating quantities like skewness and kurtosis (measurements of
the asymmetry and flattening) of the RTD function, but they are often difficult to estimate and not
t
frequently used. RTDs are often expressed in terms of dimensionless time   . Thus:
t
E    t E t 

If N reactors are connected in series then MRT and variance of the cascade can be obtained from
the following relations:
tcascade  t1  t2  t3  ...  t N

 cascade
2
  12   22   32  ...   N2

For a constant density fluid flowing in a system of volume V at flow rate Q, the MRT of the fluid
(holding time) is theoretically defined as:
V
t
Q
 For all normally operating systems, the experimentally measured MRT is the same as the holding
time but may differ in case of abnormal performance of the system.

5.2. RTD SYSTEM ANALYSIS

The concept of the residence time distribution (RTD) is fundamental to industrial reactor design.
The real time experimental RTD tracing is simple and reliable; it provides various important
hydrodynamic parameters.

5.2.1. Methodology

From a well-conducted radiotracer experiment, it is expected to obtain, in the best case, the true
RTD of the traced material in a system or in part of a system. Under less favorable circumstances, the
experimental curve simply represent an impulse response function between two detection points in the
system. This impulse response function does not possess the same conceptual power as a true RTD, but
contains valuable information all the same. Appropriate experimental RTD curve is crucial for system
analysis or modeling.

Modeling a flow from RTD experimental data means to represent the experimental curve by a
known theoretical function. Flow model is the quantitative description of hydrodynamic characteristics of
the transported material. The model helps to understanding of a process and its prediction (simulation) for
other conditions. Modeling of experimental RTD curve with theoretical functions of different flow
patterns is performed using different software. The arrangements of basic flow elements are used to
provide a proper model that gives a response identical, or as close as possible, to the signal from the tracer
experiment in the system under study. This approach is sometimes known as system (or systemic)
analysis. One therefore needs:

 A set of elementary flow models that describe the basic phenomena of fluid flow,
 A set of rules to combine these elementary models,
 An optimization procedure that makes model response fit with data from the tracer experiment.

It must be emphasized that, apart from these purely mathematical or numerical tools, some amount
of intuition, experience and self-criticism is also required to make a sound “systemic analysis” of a tracer
experiment.

5.2.2. Elementary models

Here are present a few elementary models that can be used as “building bricks” to build the
systemic model of a system. For each model are indicated:

 the physical basis of the model

 its parameters and their meaning
 a graph of its impulse response E(t) (i.e. response to a unit Dirac stimulus) and, whenever possible,
the corresponding mathematical expression.

Ideal models: plug flow and perfect mixer

 In plug flow, it is assumed that matter flows without any dispersion. In other words this flow is pure
convection. A Dirac injection is therefore transported without any deformation and shifted by a time-lag
, which is the only parameter of the model (Fig. 33). Mathematical expression of plug flow model is:
E t     t   
where δ is the Dirac impulse function.

E(t) 

0
0
Time

FIG.33. Plug flow model.

In case of a perfect mixer (or perfect mixing cell), tracer is assumed to be mixed instantaneously
and uniformly in the whole volume of system (Fig. 34). This model has one parameter, time constant τ
which is equal to the ratio of system volume V and volumetric flow rate Q.

FIG.34. Perfect mixer model.

The mathematical expression for the perfect mixer model is:

1  t
Et  exp  
  
The first moment (mean residence time) is: t =, and the second moment (variance) is σ2 = 2.

Plug flow and perfect mixer can be seen as two extreme cases, where mixing is either non existent
or complete instantaneously.

5.2.3. Models for non-ideal flows

Real flows often behave as intermediates between pure convection (plug flow) and pure mixing
(perfect mixer). Among these flows, many can be seen as the superposition of a pure transport
(convective) effect and a dispersive effect that blurs out the concentration gradients. It is often necessary
to characterize these effects; convection is related with velocities and flow rates; dispersion has an
adverse effect on heat and mass transfer which is important to quantify.

Two types of “dispersed models” can be used for the purpose:

 Axial dispersion model,

 Perfect mixers in series model - also „called tanks in series‟ or „perfect mixing cells in series‟ model.

1. Axial dispersion model

The axial dispersion (or axially dispersed plug flow) model is widely used in practice. This flow is
the superimposition of convection (bulk movement of the fluid as a plug) and some amount of dispersion.
In one dimension, and provided that dispersion can be expressed by a Fickian law, tracer concentration C
is given by the following balance equation:

C C  2C
U D 2 (1)
t x x

where U and D are fluid velocity and axial dispersion coefficient respectively.

This equation is rigorously applicable to flows in long pipes (i.e. with a very large length to
diameter ratio) or in one-dimensional columns filled with a porous medium. It is a proper approximation
also for quite a variety of quasi one-dimensional situations (river flow, underground water flow, packed
columns).

When the above equation is expressed in non-dimensional form, two parameters appear:

 A characteristic time constant τ = L/U, where L is the length of the system, and
 Non-dimensional Péclet number Pe = (U.L)/D, that represents the ratio of the convective to dispersive
effects. In other words, dispersion is predominant when Pe is low and negligible when it is large.

Fickian equation (Eq.1) can be solved for different initial and boundary conditions. The initial
condition is obviously C= 0 (zero concentration before tracer injection). There are many possibilities for
the boundary conditions, depending on whether dispersion is allowed or not at the boundaries of the
system.

A common analytical solution of the Fickian law describes the tracer concentration field as a
function of time and distance, when N moles of tracer are injected as a Dirac pulse at (t = 0, x= 0) and
dispersion is allowed at both ends of the domain (so-called “open-open” boundary conditions):

 x  Ut 2 
C t , x  
N
exp   
 (2)
4 Dt  4 Dt 

The RTD at the output of the system (at x = L) is easily deduced from this solution:
 1
1  Pe  2  Pe  t 2 
E t     exp    (3)
2   t   4 t 


The axial dispersed plug flow models have two parameters i.e.  and Pe. The mean residence time
(first moment) is equal to , and the variance (second moment) 2 = (2.2)/Pe.

The effect of varying the Pe is illustrated below using Eq. 3. The curves get sharper and sharper
when Pe is increased (Fig. 35). They always have one single peak and the peak height and tail length are
correlated (tail is short when peak is sharp and vice versa).

FIG.35. Axially dispersed plug flow model as a function of the Péclet number (Pe).

2. Perfect mixers in series.

As indicated by its name, the “perfect mixers in series” model is composed of perfect mixing cells
connected in series (Fig. 36).

FIG.36. Perfect mixers in series model

V is the total volume of the system, Q the flow rate. The number of mixers is J. Writing balance
equations for tracer concentration in each mixer shows that model parameters can be reduced to time
constant τ = V/Q and J. Some mathematical manipulation leads to the RTD function in time domain:

(4)

This expression behaves in much the same way as the one for the axially dispersed plug flow
model, J playing the same role as Pe, as shown in Fig. 37.
 FIG.37. Perfect mixers in series model as a function of J, number of perfect mixers.

As J gets large, impulse response gets closer and closer to the axial dispersion flow model.
Differences are insignificant beyond J ≈ 50. The following equivalence relationship is often quoted for
large values of J:
Pe ≈ 2. (J-1) (5)

The MRT =  and the variance 2= 2/J are calculated for the perfect mixers in series model.

One last question is the choice between the axially dispersed plug flow model and the mixer in
series model, since both can be used to represent experimental curves with one peak and “moderate”
tailing. This question holds only for low to medium values of J or Pe. On the one hand the axially
dispersed plug flow model can be thought better in a continuous system, like a pipe or a column. On the
other hand the physical relevance of this model can be held to suspicion at low Peclet numbers.
Experience has proved that the easiest to manipulate model is the perfect mixers in series, thus general
recommendation is to try this model for simulation the experimental data at the beginning.

3. Dispersion and exchange models

Dispersion models account only for moderate tailing in impulse response curves. Many
experimental curves unfortunately do not fall within that category. This is especially the case with
processes involving exchanges between a main flow and a stagnant fluid or a porous solid phase. Special
models have been developed for such cases, on the basis of either axially dispersed plug flow concept or
perfect mixers in series model. Once again these two approaches can be shown to be equivalent.

The “perfect mixers in series with exchange” model is commonly used in RTD applications. The
conceptual representation of this model is given in Fig. 38.
 FIG.38. Perfect mixers in series with exchange.

Main flow rate Q goes through a series of J perfect mixers in series of volume V1; each perfect
mixer exchanges flow rate α.Q with another mixer of volume V2. This model has four independent
parameters that can be combined in many ways; one way is to consider parameters:

JV1 V
 , J, t m  2 and k=V2/V1
Q Q

τ is the mean residence time for the main flow; t m is the time constant for the exchange between
main flow and stagnant zone, or the inverse of a “transfer coefficient” between these two perfect mixing
cells (the larger t m , the smaller the exchange), k represents the relative volume of the stagnant zone with
respect to the whole volume.

There is no simple analytical expression for E(t) in time domain. First (MRT) and second moments
(variance) of the perfect mixers in series with exchange model are:
t = τ.(1+k) and 2 = [τ2 (1+k)2 /J + 2.k.τ. tm

This model has practical value in cases where the stagnant zone is expected. An example of a
radiotracer test in a wastewater treatment unit is presented in figure 39.

FIG.39. Comparison of mixers in series and mixers in series with exchange model.
 The experimental data has been attempted successively to fit with the mixers in series and the
mixers in series with exchange models. As seen the model of mixers in series with exchange fits better
and practically it has better sense from the water flow dynamics point of view. This model is suitable for a
number of processes (river flows, flow of chemically active substances in porous media, flow in trickle
bed reactors or packed columns, etc.).

5.2.4. Rules for combining simple models

The models reviewed above are obviously not able to represent all possible tracer experiments (for
instance, no one of them is able to describe multiple peaks as can be observed in systems with recycling).
It is therefore necessary to have a set of rules for combining these models, in order to accommodate any
shape of RTD or impulse response function. Basically, models can be associated in three ways: parallel,
series and with recycling.

(a) Models in parallel

The pattern is the following (Fig. 40):

FIG.40. Models in parallel.

System i has RTD Ei  t  , MRT t i and variance  i2 . The rule for determining the RTD function
E(t) of the global system is, as follows:
Q1 Q
Et   E1  t   2 E 2  t 
Q Q

First and second moments for models in parallel are:

First moment t Second moment  2
Q1 Q2 Q1 2 Q2 2 Q1Q2
t  t1  t 2  1   2  t1  t 2 2
Q Q 2 Q Q Q

(b) Models in series

The pattern is described in Fig. 41.

System 1 System 2

Flow rate Q
 FIG.41. Models in series.

The mean residence time (first moment) for the whole system is: t  t1  t 2 , while the variation
(second moment) is given by:  2  12   22 .

The RTD function E(t) for the global system in series is calculated by convoluting the RTD
functions of sub-systems:
E  t   E1  t  * E2  t 
where
* is the convolution product (the ordinary product in the Laplace domain).

(c) Models with partial recirculation

The pattern is shown in Fig. 42. Part of flow rate (αQ) is recirculating from output into input of the
system.

Flow rate (1+)Q

Node A
System 1
Flow rate Q Flow rate Q

System 2

Flow rate Q

FIG.42. Models with recycling

t 1 and t 2 are respectively the MRTs in sub-systems 1 and 2, and 1 and 2 their respective
variances. First (MRT) and second (variance) moments for model of the whole system with partial
recirculation are:
t  1    t1  t 2
 2  1   12   22   1   t1  t 2 
2

The rule for associating the RTD is not simple in the time domain. This equation can be solved for
E(t) in the Laplace domain only.

5.2.5. Optimization procedure - Curve fitting method

Model is a time function with unknown parameters. Modeling means to match to the experimental
RTD curve a parametric functional. The evaluation of the model parameters is performed by means of the
optimization (curve fitting) of the experimental RTD Eexp(t) with the model (or theoretical RTD) Em(t,pi),
where pi-are the model parameters (which represent the process parameters). Simple models (i = 1-2) are
preferred as more reliable and practicable.
 Fitting the model RTD function with the experimental RTD curve is performed by the least square
curve fitting method. The quality of the fit is judged by choosing the model parameters to minimize the
sum of the squares of the differences between the data and model. The values of the model parameters
corresponding to the minimum value of the squares of the differences are chosen as the best.

   Eexp. (t )  Em (t , pi)2 dt  Minimum
0

There are commercial and homemade RTD software for modeling experimental RTD curve.
Progepi RTD software package distributed to many tracer groups allows the user to simulate the response
to any input of any network of elementary interconnected basic flow patterns (Fig. 43).

The RTD method still remains a global approach. RTD systemic analysis requires the choice of a
model, which is often semi-empirical and rather idealized (combination of perfect mixers, dead volumes,
etc.). It happens that different models gives appropriate fitting with the experimental RTD curve. The
Progepi RTD software may also be used to determine the parameters of the different models giving the
same response and, the subsequent physical soundness of these parameters leads to the choice of realistic
model. In addition, local measurements may be validated via the simulated local response within the
model and to optimize the corresponding parameters optimized.

It is also possible to obtain a preliminary treatment of the tracer curves including background
correction, exponential extrapolation, and normalization. The software automatically estimates moments
of both experimental and theoretical curves. This software is mainly useful in determining compartmental
models on the basis of the hydrodynamic flow behaviour of complex reactors and in simulating and
estimating the mass balance in processes with multiple recirculations.

In fact, the results of RTD modeling are not depending on the performance of particular software,
but different software facilitates extraction of information and interpretation, in particular for complex
process analysis.
 FIG.43. RTD software for modelling a process from experimental RTD curve

5.2.6. Example of simple modeling: mixer cascade with three compartments

A simple industrial mixer cascade was selected for illustrating the RTD modeling (Fig. 44). The
mixer cascade consisting of three successive compartments separated from each other by perforated walls.
From flow point of view the process can be seen as three ideal mixers with back mixing. Such mixers are
used in chemical wood pulping industry for causticizing (recovering of NaOH). The volume of the vessel
was 228 m3 and the throughput 172 m3/h at the time of test. The volumes of the compartments were
almost equal. About 10 mCi of 82Br as KBr-solution was used as tracer. The tracer was introduced to the
process input by pouring the solution (about 5 L) instantly. 2” NaI (Tl)- scintillation detector was used to
monitor tracer concentration both at inlet and outlet of reactor.

FIG.44. An industrial mixer consisting of cascade of three compartments.

The structure of this reactor suggests the simple model of three identical ideal mixers in series.
Experimental RTD, first three-ideal mixers model and the back-mixing model are presented in Fig.45.
 FIG.45. Tentative of modelling the experimental RTD curve.

The best-fitted model was three ideal mixers with backmixing. Backmixing ratio fitting better the
experimental RTD data was found to be 50%. This value was considerably high and suggested that the
reactor performance could be improved by decreasing the cross-sectional area of perforation.

There are some guidelines for selecting a model. Table IV gives some recommendations of best
models for different processes.

TABLE IV RELATIONSHIPS BETWEEN PROCESSES AND EXISTING MODELS DESCRIBING THESE

PROCESSES.

Industrial processes Recommended model

Aeration sludge channel reactor Perfect mixing cells in series (Number of mixing cells is a
function of both gas and liquid flow-rates).
Processes with endless screws (extruders, Perfect mixing cells in series exchanging with a dead
mixers, spiral classifiers…) volume (the number of cells is a function of both inlet
flow-rate and speed of rotation)
Multiphase fixed-bed reactors : RTD of Two perfect mixing cells in series model in parallel
liquid phase
Classified bed crystallizers Perfect mixing cells in series with back-mixing

The experience has shown that:

 tracer work is first of all an experimental one and all of efforts should be dedicated to the quality and
reliability of the measurements,
 experimental data may be interpreted at different level by tracer specialist or in collaboration with
chemical engineering specialist in complex cases ,
 not to use a unique dedicated software but to use different software,
 tracer experimental data remain of prime importance and the results of the experiments are not linked
to the output of particular software, but different software facilitates extraction of information and
interpretation, in particular for complex process analysis.
5.3. CONVOLUTION AND DECONVOLUTION PROCEDURES

Dirac pulse (instantaneous) injection is not always feasible in industrial and environmental systems.
There are situations where a single injection only is feasible in a system of many processing vessels in
series, and several detectors are installed along the vessel to measure the experimental response curves at
different locations of the system (Fig. 46).

FIG. 1. FIG.46. Radiotracer test: one injection several detection locations

5.3.1. Convolution

Let consider a radiotracer test in a system, in which tracer is injected at the entrance during a period
of time, and detected at the outlet. Let C0(t) and (C(t) be the tracer concentration histories (curves) at the
inlet and outlet respectively. Let E(t) be the RTD function of the system, that means the response at its
outlet to the injection of a very short unit pulse (Dirac pulse injection), δ(t), at the inlet. Convolution is the
mathematical transformation that allows to predict the response to any injection history C0(t) from the
known E(t).

Taking advantage of the linearity of the system, it is possible to express the response to the whole
injection sequence Co(t) as the sum of the responses to each individual injection step. In other words,
signal at the outlet can be written as:

C t    C0 u E t  u du
0
which expresses that C(t) is the convolution product of Co(t) and E(t), often noted as C0*E.
Decomposition of Co(t) into multiple steps, calculation of response to each step, summation of the
responses) are illustrated in Fig. 47.

Several mathematical methods can be used to resolve convolution or deconvolution problems:

 direct numerical calculation (time domain or Fourier transform)
 correlation techniques
 optimisation techniques

Experience has shown that simple methods always work for convolution but they usually fail for
deconvolution. There are many software solving the convolution or deconvolution operations, thus not
going into details of mathematical operations we recommend to the reader just to make use of these
software.

FIG.47. Principle of convolution method

Convolution is not much used in real tracer applications, because the main problem of tracer
method is to identify and quantify the residence time distribution (RTD) function that means to determine
the model of the process and to find its parameters. Normally, during a radiotracer test the experiental
response curves are recorded at the inlet and outlet of the system, thus the C0(t) and C(t) are known. The
most often problem is to find the E(t) from C0(t) and C(t). This is so called the deconvolution operation.

5.3.2. Deconvolution

Convolution consists in calculating C(t) when C0(t) and E(t) are known. Deconvolution is the
inverse procedure, i.e. calculation of E(t) when C0(t) and C(t) are known. This case, where responses at
two measurement stations are known, is being the most common one in tracer applications. The
deconvolution of C(t) and Co(t) is noted as: C/Co, keeping in mind that if convolution is commutative,
deconvolution is not.

There are several reasons why one needs to deconvolute signals. Any stationary and linear system
is unambiguously characterized by its impulse response function, or its response to a Dirac pulse
stimulus. This response can, in most practical cases, be considered as the RTD of the traced material.
Knowledge of that impulse response function is therefore desirable to be able to predict system response
to any stimulus, or to analyze system behaviour in terms of RTD. Unfortunately, injection of a true Dirac
pulse of tracer is not always achievable. Besides, it is quite common to have several detectors at different
stages of complex systems. Even if tracer is injected as a perfect Dirac pulse, only readings from the first
detector can be considered as the impulse response function of that part of the system; one has to
deconvolute the signals from the second and first detectors (i.e. calculate C2  C1 ) to get the impulse
response function of the second part of the system, and so forth with the next detectors.