Chemical Engineering Science 66 (2011) 5045–5054

Contents lists available at ScienceDirect

Chemical Engineering Science

journal homepage: www.elsevier.com/locate/ces

Model-based scale-up of vacuum contact drying of

pharmaceutical compounds
Marcello Murru a, Giovanni Giorgio b,n, Sara Montomoli c, Francois Ricard b, Frantisek Stepanek d
a
Cargill R&D Europe, Havenstraat 84, 1800 Vilvoorde, Belgium
b
GlaxoSmithKline R&D, Gunnels Wood Road, Stevenage SG1 2NY, UK
c
ENEL Italia, Viale Delle Industrie, 6100 Roma, Italy
d
Department of Chemical Engineering, Imperial College London, South Kensington Campus, London SW7 2AZ, UK

a r t i c l e i n f o abstract

Article history: An integrated methodology for the scale-up of vacuum contact drying with intermittent agitation is
Received 10 May 2010 described in this work. The methodology combines a mathematical model of vacuum contact drying,
Received in revised form based on differential transient heat and energy balances, and a small-scale experimental apparatus for
7 April 2011
model validation and parameter estimation. The validated model was used for the estimation of drying
Accepted 28 June 2011
Available online 5 July 2011
times of six different pharmaceutical compounds at the pilot and manufacturing scale over a range of
drying conditions – pressure 15–200 mbar, temperature 45–70 1C, solvents: acetone, water, methanol,
Keywords: n-propanol, and isopropyl acetate. The mean difference between predicted and actual drying times for
Drying the six compounds was less than 9%, which is considered a significant improvement over current semi-
Heat transfer
empirical approaches to vacuum contact drying scale-up.
Mass transfer
Crown Copyright & 2011 Published by Elsevier Ltd. All rights reserved.
Mathematical modelling
Scale-up
Pharmaceutical processing

1. Introduction is the Nutsche filter-dryer, in which two unit operations (filtration

and drying) can be carried out sequentially without the need to
Drying is a common unit operation used to reduce the levels of manipulate wet filter cake. A Nutsche filter-dryer has a cylindrical
water or organic solvent in pharmaceutical materials to accep- geometry; typical diameters of a production-scale filter dryer can
table levels. Drying is often the bottle-neck in the Active Pharma- be of the order of two metres and handle cake depths of the order
ceutical Ingredient (API) manufacture—while other unit opera- of centimetres up to a few tens of centimetres (depending on cake
tions in the isolation stage such as extraction, crystallisation or permeability). Heat is supplied through the jacketed walls and
filtration typically require residence times of the order of tens of vapours are removed from an evacuated head-space above the
minutes, residence times in drying may be up to an order of filter cake. The process can be intensified by continuous or
magnitude longer (tens of hours) in some cases. Thus, there is intermittent agitation and/or the use of nitrogen bleed.
often economic pressure to select aggressive drying conditions in In vacuum contact dryers, several physical phenomena occur
an effort to minimise cycle times and increase throughput. simultaneously (Whitaker, 1977). The first is heat transfer by
However, the drying conditions can impact on the properties, conduction from an external source (heating jacket) through the
functionality, and quality of the material being dried thus affect- cake to the water or organic solvent (see Fig. 1). The heat supplied
ing further downstream operations (particle attrition, unwanted causes the solvent to evaporate and the resulting vapour has to
agglomeration, etc.). A balance between drying time and quality permeate through the filter cake and out of the drying equipment.
and performance is required in order to preserve the quality of Liquid solvent also permeates through the cake, by capillary flow.
the product. In the absence of mechanical agitation, both vapour removal and
Drying can be carried out in a large variety of equipments. In liquid motion inside the material are dictated by their respective
the pharmaceutical industry, vacuum contact drying is widely pressure gradients (Mahadevan et al., 2006; Shahidzadeh-Bonn
used to dry products that are sensitive to oxygen and temperature et al., 2007).
and are often toxic or explosive. One particularly popular design Of the physical phenomena involved in vacuum contact dry-
ing, heat conduction through a dry powder bed tends to be the
rate limiting one. Drying rate can be increased by increasing the
n
Corresponding author. Tel.: þ44 1438762944; fax: þ44 1438764414. jacket temperature and lowering the head-space pressure in order
E-mail address: [email protected] (G. Giorgio). to increase the driving force for heat transfer, as well as

0009-2509/$ - see front matter Crown Copyright & 2011 Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.ces.2011.06.059
5046 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054

aim of this work was to validate the model for several real
pharmaceutical compounds under a range of operating conditions
(pressure, temperature) and for different solvents, and to demon-
strate its use in an integrated methodology for process scale-up
that involves parameter identification at the laboratory scale and
the prediction of drying times at the pilot or manufacturing scale.
The model is also shown to be easily extensible to vacuum drying
with intermittent agitation of the filter cake.

2. The drying model

2.1. Model assumptions

The model is based on differential mass and energy balances in

the wet cake of specified geometry, whereby the wet cake is
regarded as a stationary porous solid with constant porosity but
Fig. 1. Schematic of physical phenomena taking place in a wet cake during static spatially and temporally variable moisture content (pore satura-
vacuum contact drying in a Nutsche filter dryer with axial symmetry. tion) and temperature. The principal model assumptions are:

– Vapour flow is not constrained by the cake’s pores so that the

by mechanical agitation of the powder bed. However, many
vapour transfer from the evaporation location to the head-
pharmaceutical compounds are thermo-sensitive, which poses a
space is not the rate-limiting step.
limit on the maximum temperature that can be used. A mini-
– The liquid-phase transfer through the cake is described by the
misation of undesired effects of agitation – such as agglomeration,
Richards equation with a linear approximation for the depen-
balling, attrition, and breakage – is also crucial for the quality of
dence of the apparent moisture diffusion coefficient on the
the product. For this reason optimum conditions often involve
moisture content (see Eq. (10)).
static drying coupled with intermittent agitation, or starting
– The heat transfer is by conduction only and the heat dissipa-
agitation at a specific level of Loss On Drying (LOD—fraction of
tion is by evaporation only.
solvent to wet cake mass).
– The local cake properties (porosity, transport properties) are
Therefore, in design and optimisation of drying processes,
spatially homogeneous and scale-independent.
there is a need for stable and reliable models to quantify and
– The adsorption/desorption of the solvent on the solid material,
predict drying rates and drying times with a satisfactory accuracy
as well as any other specific solvent–solid interactions are
(Kemp, 2007). The uses of a model for drying curves predictions
neglected.
can include: the prediction of end-point of drying and suggestion
– The liquid evaporation is described by a kinetic model assum-
of time for sampling, cycle time estimation at the pilot or
ing that the liquid evaporates proportionally to a difference
manufacturing scale from laboratory-scale experiments, predic-
between the equilibrium and actual local vapour pressure; this
tion of the start point of agitation and comparison of different
kinetic step is not the rate controlling one—to make the model
agitation regimes, comparison of different operating conditions
insensitive to this parameter the kinetic constant is deliber-
(pressure, temperature, cake volume), comparison of different
ately chosen very high (see Eq. (3)).
solvents, comparison of dryer designs, real-time model-predictive
– The physical properties of all materials are constant during
process control that can be useful when operating conditions
drying and spatially homogeneous.
deviate from the set values.
Drying models of different complexity can be found in the
literature. In concentrated-parameter models, spatially averaged Some of these assumptions of course limit the model applic-
state variables (usually temperature and moisture content) are ability. The model is not suitable in cases where large variations
considered and the model has the form of Ordinary Differential of physical properties of the solid material occur during drying.
Equations (see e.g. Michaud et al., 2008a). In distributed-parameter Corrections could potentially be applied to the parameters that
models, the state variables are functions of both time and space, define the physical properties of the cake, but this would require a
and the model therefore has the form of partial differential law for the variation of the parameters with time or with the
equations (PDE’s). Distributed-parameter models can further be other variables of the system (e.g. temperature, solvent content,
classified as either continuum or discrete. In continuum models, particle size, etc.). If the cake structure (e.g. porosity) were scale-
the heat and mass transfer phenomena are described by macro- dependent – which might be the case for highly compressible
scopic laws (Fourier’s law for heat conduction, Darcy’s law for fluid cakes – this dependence would need to be reflected in the values
flow) and the wet porous solid is regarded as a continuum with of the transport parameters (thermal conductivity and apparent
constant or smoothly varying values of effective transport proper- moisture diffusivity). As both quantities also scale with the cake
ties (Kohout et al., 2006). In discrete models, the heterogeneous permeability, the necessary information could be derived from
microstructure of the porous medium is explicitly represented, the filtration scale-up data.
either by a pore network model (e.g. Wei et al., 1987; Metzger The model is also not suitable for very low permeability cakes
et al., 2007) or by the reconstructed porous medium approach that hinder the vapour flow towards the top of the cake. This
(Kosek et al., 2005). The discrete, continuum, and concentrated- limitation could be rectified by adding a further mass balance for
parameter models can be related to one another by hierarchically the vapour phase produced by evaporation (see Kohout et al.,
applying the method of spatial averaging (Whitaker, 1999; Kohout 2006). While this is not principally a problem, from practical
and Stepanek, 2007). experience it is known that most filter cakes fall into the category
In the present work, a dynamic distributed-parameter model where vapour diffusion is not rate limiting. Finally, if other drying
derived from that developed by Kohout et al. (2006) is used. The or moisture transfer mechanisms such as adsorption/desorption,
 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054 5047

surface diffusion or desolvation are important, the model would Since WS is assumed to be constant and WL is a state variable
have to be extended to take these phenomena into account. and therefore known in each time and location, WG can be
calculated as a complement to one. The local dependence of the
2.2. Governing equations effective thermal conductivity on the solvent content is calculated
from the thermal conductivity of the dry and fully liquid-satu-
Based on the above assumptions, the governing system of rated porous medium (filter cake) using an expression proposed
partial differential equations is composed of a mass balance in by Kohout et al. (2004)
the liquid phase and an energy balance on the cake (liquid and WL
solid), which are coupled by a kinetic expression for the evapo- leff ¼ ldry þ ðlwet ldry Þ ð7Þ
1WS
ration rate.
Liquid-phase mass balance: Let us now turn to the description of liquid-phase mass
transfer. Assuming capillary flow to be the dominant mechanism
@WL _ LG
m
¼ rUðDL rWL Þ ð1Þ of liquid phase migration in the filter cake, the mass flux of the
@t rL
liquid phase is given by Darcy’s law
Energy balance: rL
jL ¼  kðWL ÞrpL ð8Þ
@T mL
/rCp S _ LG DHvap
¼ rUðleff rTÞm ð2Þ
@t
where k(WL) (m2) is the relative permeability, mL (Pa s) is the
In the above equations, WL (dimensionless) is the liquid phase liquid viscosity, and pL (Pa) is the liquid-phase pressure, related to
volume fraction, T (K) is temperature, t (s) is time, DL (m2 s  1) is the gas-phase pressure through the capillary pressure, pC:
the apparent liquid phase diffusion coefficient (more on this
pC ðWL Þ ¼ pG pL ð9Þ
coefficient below), leff (W m  1 K  1) is the effective thermal 
conductivity, rL (kg m  3) is the liquid phase density, rCp The capillary pressure is generally a non-linear, decreasing
3 1 1
(J m K ) is the mean volumetric heat capacity, DHvap (J kg ) function of the moisture content. After substitution into Eq. (8)
is the latent heat of vaporisation, and m _ LG (kg m  3 s  1) is the and differentiation, we obtain the Fick-like moisture transport
local evaporation rate. We postulate that the evaporation rate is term occurring in Eq. (1), which is also known as the Richards
proportional to the difference between the equilibrium and the equation in the porous media literature (e.g. Daian and Saliba,
current vapour pressure, i.e., 1993).
8
< kvap rL ðp pG Þ=pG if WL 40
n
> jL ¼ rL DL ðWL ÞrWL ð10Þ
m_ LG ¼ 0 if WL r0 ð3Þ
>
:0 The apparent moisture diffusivity in the unsaturated porous
if pn r pG
medium
where kvap (s  1) is a rate constant, pG (Pa) is the actual local  
kðWL Þ @pc ðWL Þ
vapour pressure, and pn(T) is the vapour pressure that would be in DL ð W L Þ ¼  ð11Þ
mL @ WL
equilibrium with the local temperature. pn(T) can be evaluated
from the Antoine equation is an a priori unknown function that combines the product of
permeability and the derivative of capillary pressure with respect
B
log10 pn ¼ A ð4Þ to moisture content. Various expressions for the dependence of
C þT
the relative permeability and the capillary pressure on solvent
The constants A, B and C can be found for common solvents e.g. content in an unsaturated porous medium have been proposed in
in Smallwood (2006). Due to the assumption of zero resistance to the literature (see e.g. Mahadevan et al., 2006); their parameters
vapour phase mass transfer in the present model, the vapour have to be determined empirically from experimental data. In this
phase pressure pG is constant and equal to the head-space work, we use a simple two-parameter linear relationship of the
pressure in all points of the filter cake. form
The kinetic expression for the evaporation rate m _ LG has been (
introduced mainly for numerical reasons; if the value of kvap is aðWL WnL Þ if WL Z WnL
DL ¼ ð12Þ
if WL o WL
n
chosen sufficiently large then the local vapour–liquid equilibrium 0
is essentially instantaneous and the solution becomes indepen- n
where WL is the residual saturation (for which the capillary
dent of kvap as vaporisation is not the rate-limiting step. However,
pressure diverges) and a is a proportionality constant. Both
the governing set of PDE’s (Eqs. (1)–(2)) can still be integrated as a
adjustable parameters have to be determined by fitting labora-
differential rather than a differential-algebraic system. It should
tory-scale experimental data; however, as material properties
be noted though, that in a general case – e.g. when drying porous
they should be preserved upon scale-up.
particles, hygroscopic materials, solvates or other systems with
intra-particle mass-transfer – the rate-limiting step may well be
the local evaporation rate and in that case the rate constant kvap 2.3. Boundary conditions
would have to be treated as an adjustable parameter with an
influence on the overall drying rate. The governing PDE’s (Eqs. (1) and (2)) were integrated
To complete the set of governing equations, the mean volu- numerically on a spatially 2D domain in cylindrical coordinates,
metric heat capacity is calculated from the pure components and taking advantage of the axial symmetry of the Nutsche filter dryer
the phase volume fractions according to (cf. Fig. 1). The COMSOL Multiphysics finite element software
  package (http://www.comsol.com) was used for the numerical
rCp ¼ WL rL Cp,L þ WS rS Cp,S þ WG rG Cp,G ð5Þ
integration of the PDE’s. The following boundary conditions were
where ri and Cp,i are the density and the specific heat capacity, applied:
respectively, of each phase. The phase volume fractions must at r ¼0 (axis):
satisfy
@WL @T
¼ ¼0 ð13Þ
WL þ WS þ WG ¼ 1 ð6Þ @r @r
5048 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054

at r ¼R and z ¼0 (side and bottom walls): according to the following equations:

R
@WL @WL /rCp STdV
¼ 0 or ¼0 ð14aÞ T~ ¼ RV ð18Þ
@r @z V /rCp SdV

@T @T and
leff ¼ hq ðTh TÞ or leff ¼ hq ðTh TÞ ð14bÞ Z
@r @z 1
W~ L ¼ WL dV ð19Þ
at z¼H (top of filter cake): V V

@WL @T This approximation is only valid if the agitation time can be

¼ ¼0 ð15Þ neglected and the cake is properly mixed at the end of it. A second
@z @z
underlying assumption is that the entire volume of the cake is
The boundary conditions for the mass balance (Eq. (1)) reflect mixed, i.e. that there are no dead zones.
the absence of mass flux at the walls, axial symmetry, and zero
flux of the liquid phase through the top of the cake. It should be
2.6. Model parameters
noticed that the mass balance is carried out only on the liquid
phase, which is consumed in the evaporation term (Eq. (3)), hence
The system of governing equations and auxiliary relationships
justifying the zero flux condition on the top boundary. Similarly
defined in the previous sections requires several parameters that
the heat balance has an isolation condition at the top of the cake
can be summarised in four main groups:
and axial symmetry.
The heat transfer coefficient, hq, reflects the combined heat
(a) Equipment-specific parameters:
transfer from the heating fluid to the wall and the conduction
– Vessel diameter, R.
through the dryer walls. Th is the temperature of the heating fluid.
– Wall heat-transfer coefficient, hq.
Depending on the dryer design (laboratory vs. pilot or manufac-
(b) Operation parameters:
turing scale), the bottom filter plate cannot always be heated in
– Cake depth, H.
the same way as the side walls, hence Th and/or hq of the bottom
– Initial moisture content, wL0.
plate can have a different value from the side wall, or be given as
– Initial temperature, T0.
a function of the radial position rather than fixed.
– Head-space pressure, pG.
– Jacket temperature, Th.
2.4. Initial conditions – Period between intermittent agitation (if applicable), tint.
(c) Properties of the solvent and cake:
The initial conditions for the two state variables are based on – Latent heat of vaporisation, DHvap.
the assumption of uniform distribution of solvent in the cake and – Coefficients A, B, C in Antoine equation.
n
uniform initial temperature in the cake. Inhomogeneous distribu- – Parameters a and WL for the apparent diffusion coefficient,
tion of moisture content could in principle also be considered, e.g. DL .
based on a mode of cake drainage and deliquoring, however, this – Density of solid and liquid, rS, rL.
was not done in this work. The values of the initial solvent – Specific heat capacity of solid and liquid, CpS, CpL.
content and temperature were in each case based on the experi- – Thermal conductivity of the liquid, lL.
mentally measured LOD (loss on drying) and the temperature of – Effective thermal conductivity of dry material, ldry.
the cake after the blowback of the last wash. The initial conditions – Solid phase volume fraction, WS.
can be summarised as (d) Numerical parameter:
– Evaporation rate constant, kvap.
rS wL0
WL ðt ¼ 0Þ ¼ WL0 ¼ WS 8r,z ð16Þ
rL 1wL0
From the above parameters, those related to the solvent can be
found in the open literature (e.g. Smallwood, 2006). However, the
Tðt ¼ 0Þ ¼ T0 8r,z ð17Þ
solid- and cake structure-related parameters have to be usually
where wL0 ¼ (mwet  mdry)/mwet (dimensionless) is the initial LOD determined experimentally either by direct measurement or by
of the cake and T0 (K) is the initial average cake temperature. fitting of small-scale drying experimental data. The methodology
used in this work for unknown parameter identification is
described in the following section.
2.5. Intermittent agitation

As was mentioned in the Introduction, although mechanical 3. Experimental methodology

agitation can enhance heat transfer (and therefore drying rate) by
bringing wet material closer to the heated walls once a dry layer 3.1. Physical properties measurements and collection
with a poor thermal conductivity has formed, it is often not
desired to agitate continuously due to adverse effects such as Most of the organic compounds studied in pharmaceutical
‘‘snowballing’’ and particle attrition. Intermittent agitation con- R&D are new chemical entities (NCE’s) and have not been fully
sists of longer periods of static drying interleaved by short periods characterised for physical properties, therefore some measure-
when agitation is on. The on/off ratio is typically less than 1/10, ments are required for the missing parameters of the model.
e.g. a few minutes every hour. The model presented in Section 2.2 These parameters can be solid specific or cake specific and
above does not contain momentum and mass balances of the therefore depend on the packing of particles in the filter cake
solid phase therefore it cannot model mechanical agitation and not only on the crystal structure. The underlying assumption,
explicitly. However, due to the small on/off ratio, the effects of when using the model for scale-up predictions is that all the
intermittent agitation can be approximated by periodically (with physical properties are maintained across scales.
a period tint) re-setting the initial conditions, i.e. by spatial The true density for a perfect crystal can be determined by
averaging of the temperature and moisture content profiles crystallographic methods, either by determining a crystal structure
 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054 5049

by single crystal X-ray diffraction or by determining the unit cell measuring cylinder and calculated from its definition:
and space group symmetry from a powder X-ray diffraction mS
pattern. Density is related to the unit cell volume by the following rb ¼ ð21Þ
Vb
equation:
where mS is the mass of the powder sample and Vb the bulk
Mw Z volume occupied by it.
rS ¼ ð20Þ Thermal conductivity and heat capacity of both loose and
NA V0
compact powder were measured by the transient hot wire
where rS is the true density of the material, Mw is the molecular method. The measurement was easily carried out using the KD2
mass, Z is the number of molecules in the unit cell, NA is Avogadro’s Pro thermal properties analyser (Decagon Devices, Pullman, USA)
number, and V0 is the volume of the unit cell. The value of density equipped with a 3 cm dual needle sensor. The Arctic silver paste
calculated with this method was used in the simulation and for the was applied on the sensor to allow for a better thermal contact
calculation of the cake void and solid volume fractions. between the probe and the powder.
The bulk density of the compressed powder (filter cake) has to
be measured in situ during the drying experiments after blow- 3.2. Laboratory-scale experiments
down. This value should be used for static drying case as the cake
remains compressed during drying without mechanical agitation. Laboratory-scale static drying experiments are a crucial com-
However, in case the cake is agitated, a bulk density of the loose ponent of the overall scale-up methodology as they allow the
powder is appropriate due to ‘‘aeration’’ effect of the agitation. For fitting of the liquid effective diffusivity parameters. The experi-
the loose powder the bulk density was obtained simply by ments were carried out in a 500 ml glass filter dryer whose set-up
measuring the weight and the volume of the powder in a configuration is shown schematically in Fig. 2. In order to obtain a
drying curve, the vapour leaving the jacketed vessel (the dryer)
was condensed using a condenser inserted in the vacuum line
before the vacuum pump equipped with a vacuum controller. The
mass of the collected condensate as function of time was recorded.
The temperature of the cake was measured by thermocouples and
all the equipment was connected to a computer with data
acquisition software for all the experiments. An overhead stirrer
was embedded in the filter cake to facilitate agitation when
required.
Before the vacuum drying was started, the cake depth was
measured in order to know the value of H for boundary condi-
tions, and the initial moisture content of the cake – the loss on
drying (LOD) – was determined for the initial conditions.
The actual vacuum drying experiment then involves setting
the desired heating jacket temperature and the vacuum level, and
starting the agitator if required (intermittent or continuous). The
mass of condensate as function of time is recorded by means of
computer data logging Apart from the mass of condensate, other
parameters that are logged are: jacket temperature, head-space
pressure, cake temperature (noting the exact position of the
Fig. 2. Set-up of a 500 ml glass filter dryer used in the lab scale experiments. thermocouples in the cake). The cake depth before and after

Table 1
Summary of scale-up predictions and actual drying times at pilot plant or manufacturing scale.

Compound, cake volume, Predicted drying Actual drying a (m2 s  1), WnL (dimensionless), pG, solvent, Th
agitation mode time (h) (model) time (h) (experiment) ldry (W m  1 K  1)

GSKC#4 2.7 2.4 1  10–6 100 mbar

31 L 0.02 Acetone
Intermittent agitation 0.057 50 1C

GSKC#2 12 12 1  10–6 15 mbar

130 L 0.05 n-Propanol
static 0.058 60 1C

GSKC#5 stage 8 Bx1 10.5 11.6 1  10–7 120 mbar

230 L 0.12 Methanol
Intermittent agitation 0.13 45 1C

GSKC#1 stage 3 Bx1 25 22 21  10–7 50 mbar

400 L 0.001 Water
Continuous agitation 0.088 70 1C

GSKC#5 stage 8 Bx2 4.6 5 1  10–7 200–45 mbar

230 L 0.12 Methanol
Intermittent agitation 0.13 45 1C

GSKC#3 stage 5 Bx1 4.5 4 1  10–7 100 mbar

230 L 0.08 Isopropyl acetate
Intermittent agitation 0.117 55 1C
5050 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054

drying, and the mass of solid charged and discharged are also 60
manually recorded.
55

3.3. Pilot-plant and manufacturing-scale experiments 50

Drying time [h]

45
The model output was compared with pilot plant and manu-
facturing-scale experiments. The scale of the equipment (in terms 40
of volume processed) was from 100 times to 1000times bigger
than the lab experiments. 35
At this scale it was not possible to perform experiments in
30
replicates but the experimental measurements have a good accuracy:
25
– Pressure 0–1000 mbar710 mbar.
20
– Jacket temperature:  20–15072 1C. 0 0.05 0.1 0.15 0.2 0.25
Thermal conductivity [W m-1 K-1]
In order to measure the moisture level a sample of the cake
was extracted from the drier and the loss on drying was Fig. 4. Effect of thermal conductivity on drying time, simulation results. Cake
determined with a validated and accurate method: radius 40 cm, cake height 10 cm. Solid apparent diffusion coefficient 0.1 m2/s,
solid heat capacity 2060 J kg  1 K  1. Solvent isopropyl acetate and initial solvent
content 15% w/w.
– LOD content 0–10070.1%

Due to the complexity, risks and the cost of the sampling

procedure the samples were taken at regular interval and only the 60
final sample (final point) was measured twice before the drier
55
discharge (cf. Table 1).
50
Drying time [h]

4. Results and discussion 45

40
4.1. Model-based parametric sensitivity study
35
A preliminary study on the effect of the cake parameters on
30
drying time was conducted in order to identify the most sensitive
parameters and the required accuracy of their measurements. The 25
parametric sensitivity study was conducted on static drying cases
varying the cake properties in the following ranges: 20
1000 1250 1500 1750 2000 2250 2500 2750 3000
Thermal conductivity of the cake 0–10–6 W m  1 K  1 Cp [J Kg-1 K-1]
Apparent diffusion coefficient 0.02–0.2 m2 s  1
Solid heat capacity 1000–3000 J kg  1 K  1 Fig. 5. Effect of solid heat capacity on drying time, simulation results. Cake radius
40 cm, cake height 10 cm. Solid apparent diffusion coefficient 0.1 m2/s, thermal
conductivity 0.168 W m  1 K  1. Solvent isopropyl acetate and initial solvent
The values used for the study are within a typical range for content 15% w/w.

cakes composed of pharmaceutical products and common sol-

vents. The scale chosen is that of a typical industrial dryer (40 cm

60 60
55
50
50
Drying time [h]
Drying time [h]

45 40

40 30
35
20
30

25 10

20 0
0 2 4 6 8 10 12 0 50 100 150 200 250
10-7 * Apparent diff. coeff. [m2/s] Headspace pressure [mbar]

Fig. 3. Effect of diffusion coefficient on drying time, simulation results. Cake radius Fig. 6. Effect of head-space pressure on drying time, simulation results. Cake
40 cm, cake height 10 cm. Solid effective thermal conductivity 0.168 W m  1 K  1, radius 40 cm, cake height 10 cm. Solid apparent diffusion coefficient 0.1 m2/s,
solid heat capacity 2060 J kg  1 K  1. Solvent isopropyl acetate and initial solvent thermal conductivity 0.168 W m  1, solid heat capacity 2060 J kg  1 K  1. Solvent
content 15% w/w. isopropyl acetate and initial solvent content 15% w/w.
 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054 5051

diameter) and the solvent is isopropyl acetate which is also 0.35

commonly used in pharmaceutical processes. Parameters such experimental
as solid density and volume fraction affect the calculation 0.3 slope 5e-6, thetar 0.01
indirectly by varying the apparent diffusion coefficient therefore slope 1e-6, thetar 0.01
were not considered in the parametric sensitivity study. However, slope 5e-7, thetar 0.001

Solvent content w/w

0.25
an incorrect value of the solid volume fraction can result in an
incorrect value of the initial liquid volume fraction.
Figs. 3–5 show the relative importance of the various para- 0.2
meters suggesting that the two main parameters which affect the
drying time are the effective thermal conductivity of the porous 0.15
medium and the apparent diffusivity of the solvent in the cake.
The effect of the solid heat capacity is much less significant in 0.1
the case of static drying, however it can play a more important
role when the cake is agitated. Figs. 6 and 7 show the effect of
0.05
head-space pressure and initial LOD. As it was already noted by
Kohut et al. (2006), the head-space pressure is crucial for the
calculation of the drying time since it dictates the driving force of 0
0 2000 4000 6000 8000 10000 12000 14000 16000
the heat trasfer. Even a small variation of few mbar of the head-
Time [sec]
space pressure can significantly affect drying time.
n
Fig. 8. Effect of moisture migration parameters a (slope) and WL on drying curves
for the system GSKC#1 in 1-propanol, dried at 40 mbar with a jacket temperature
4.2. Parameter identification of 50 1C, without agitation. Experimental data and simulation results as indicated
in the legend.
As it was demonstrated in the preliminary sensitivity study, the
moisture migration coefficient DL heavily affects the drying curve,
n
therefore the parameters a and WL that define it were obtained 0.4
from matching the experimental drying curves. Note that from all experimental
the parameters appearing in the model equations (cf. Section 2.6 0.35 calculated
n
above), only a and WL need to be regarded as adjustable para-
meters; all other parameters can be determined independently 0.3
without the knowledge of an experimental drying curve (at any
Solvent content w/w

n
scale). The procedure to regress the parameters a and WL can be 0.25
shown for one of the compounds (GSKC#1) with 1-propanol
solvent, studied as an example. The drying curve resulting from a
0.2
static drying experiment at lab scale (10 cm diameter, 4 cm cake
depth) at Th ¼50 1C and pG ¼ 40 mbar is shown in Fig. 8. The drying
model was run while systematically varying the values of a and WL
n 0.15
in order to minimise the integral difference between the experi-
mental and calculated data points. 0.1
n
The optimised parameters a and WL were then used for drying
simulation at different scales and conditions provided that the 0.05
cakes had the same physical properties. Before scale-up, the data
were verified using another laboratory-scale experiment with the 0
same compound and solvent, but dried under different conditions 0 1000 2000 3000 4000 5000
(Th ¼80 1C and pG ¼30 mbar). The experimental and the calculated Time [sec]

45 Fig. 9. Validation of model using experimental data for GSKC#1 at 30 mbar

pressure and 80 1C wall temperature, with optimised parameters a ¼5  10  7
n
and WL ¼ 0:001.
40
curves are shown in Fig. 9. Although the agreement is good, the
Drying time [h]

small difference could in this particular case be attributed to the

35 fact that the same material as in the previous experiment was
used, re-wetted with the wash solvent after the first drying.
Therefore, small changes in the transport properties could have
30
occurred despite the fact that the bulk density of the two cakes
(measured) was the same.
25
4.3. Scale-up

20 Using the methodology described in the previous sections, the

0 10 20 30 40 50 model was used to predict the drying curves at pilot plant and
Initial LOD %w/w manufacturing scale for different compounds at different condi-
tions of pressure, wall temperature and agitation. The crystals
Fig. 7. Effect of initial solvent content on drying time, simulation results. Cake
radius 40 cm, cake height 10 cm. Solid apparent diffusion coefficient 0.1 m2/s,
forming the cakes vary widely in both size and particle shape.
thermal conductivity 0.168 W m  1, solid heat capacity 2060 J kg  1 K  1. Solvent This results in cakes with different packing, porosity, and pore
isopropyl acetate and pressure 120 mbar. size distribution.
5052 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054

The model was run at the pilot or manufacturing scale using ldry for the uncompressed (loose) powder was used. Table 1
the same parameters used for the lab scale runs, just by changing summarises the results obtained from the simulations detailing
the geometry (radius and depth) of the cake. In case of inter- the conditions (pressure and temperature), the scale (cake
mittent agitation the value of the effective thermal conductivity volume) and the values of the moisture transport parameters a
n
and WL used in the simulations.
As can be seen in Table 1, although the predicted and
0.35 measured drying times do deviate to some extent, these differ-
ences are very small compared to prevailing practice where semi-
0.3 empirical methods of drying end-point estimation often lead to
Solvent content w/w

inaccuracies of the order of 100% (e.g., vacuum is broken and a

sample taken with the expectation of being dry at a certain time,
0.25
whereas the actual time eventually required to reach the desired
solvent content can be up to twice as long). The mean percentage
0.2 difference between simulations and experiments for the runs
reported in Table 1 was 9% and no greater than 12%.
0.15

0.1 4.3.1. Static case

0 2 4 6 8 10 12 14 The best agreement was achieved in the static drying case
Time [hours] (compound GSKC#2), where the drying time required for reaching
a 20% w/w solvent content matches the actual value measured in
Fig. 10. Simulation output for static drying at pilot plant scale (130 L cake) of the plant. The drying curve output of the simulation for GSKC#2
GSKC#2.
at pilot plant scale is shown in Fig. 10, and the simulated solvent

Fig. 11. Solvent content and temperature profiles for GSKC#1, static drying— Fig. 12. Apparent moisture diffusion coefficient and local evaporation rate profiles
simulation results. for GSKC#1, static drying—simulation results.
 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054 5053

0.35 305

0.3 300

0.25 295

Solvent content w/w

Temperature [K]
0.2 290

0.15 285

0.1 280

0.05 275

0 270
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [hour]

Fig. 13. Simulation results for intermittent agitated drying at pilot plant scale (31 L cake) of GSKC#4.

0.049 the dry layer, and moisture migration by capillary flow from the
10 min wet to the dry region (cf. Lehmann et al., 2008).
0.042 30 min As seen in Fig. 10, the drying rate in a static drying case
1 hour decreases significantly with time, resulting in very long drying
Solvent content w/w

0.035 static time at large scale. This behaviour is more evident at larger scale
because of the relatively larger proportion of the dry section of
0.028 the cake in contact with the wall that is offering high heat transfer
resistance (lower surface to volume ratio after scale-up). Agitation
0.021 is crucial at large scale in order to reduce drying time by
effectively contacting the wet powder with the wall.
0.014

0.007 4.3.2. Intermittent agitation

Fig. 13 shows a simulated drying curve and the average
0 temperature profile for a case of intermittent agitation where the
0 5 10 15 20
static drying periods were 1 h followed by 10 min of agitation. The
time elapsed sudden decrease of temperature is also detected experimentally
Fig. 14. Example of model application for intermittent agitation drying time
and it is due to the fast evaporation rate resulting from contacting
estimation under different intermittent agitation regimes (simulation results for cold wet and hot dry powders during mixing. The curves become
GSKC#5 stage 8, methanol solvent, cake volume 230 L, Th ¼45 1C, pG ¼ 45 mbar). very steep in the point of the homogenisation because of the
approximation that mixing is instantaneous.
However, in some cases agitation can promote agglomeration
of the material if the solvent content is beyond a certain value
content and temperature profiles in the cake at t ¼4300 s from the (often called ‘‘sticky point’’). In such cases, agitation should be
start of drying are shown in Fig. 11. avoided until the solvent concentration is below the sticky point.
The concentration and temperature profiles also affect the Agitation in the sticky regime should be avoided, and the model
evaporation rate and the apparent moisture diffusion coefficient presented in this work was used to test alternative intermittent
profiles shown in Fig. 12. The evaporation rate is practically zero agitation regimes and their effect on drying rate. Fig. 14 shows the
everywhere except in the region at the boundary of the wet part of effect of different intermittent agitation regimes at the plant scale
the cake (the drying front). There is a clear temperature profile in in the case of GSKC#3 in comparison with the static drying case. It
the dry layer of the cake, whereas the temperature in the wet core is evident that even occasional intermittent agitation can drama-
of the cake is constant and equal to the boiling point at pG, meaning tically improve drying time. Increasing the frequency of agitation
the driving force for evaporation is zero. In the zone of the cake only marginally improves drying time for GSKC#3. The heat
next to the wall, the evaporation rate is also zero due to the lack of provided by the wall to the bed during static drying heats up
solvent (WL ¼0). With increasing time, the drying front progresses the powder next to the wall generating a temperature profile that
towards the centre of the cake and the wet core decreases in size. gradually drops to the boiling point of the solvent where the wet
The apparent moisture diffusion coefficient DL acts in the opposite core of the cake is reached. The heat accumulated in the dry
way from the drying front and allows the permeation of the powder is suddenly released during the agitation cycle resulting
solvent towards the wall. Therefore, high moisture mobility will in a much faster evaporation rate and consequent decrease of
promote faster drying. The position of the drying front is the result temperature (cf. also Michaud et al., 2008b). All these phenomena
of a dynamic balance between heat transfer from the wall through are also observed experimentally at small and large scale.
5054 M. Murru et al. / Chemical Engineering Science 66 (2011) 5045–5054

5. Conclusions compositions of the solvent mixture, and an additional state variable

expressing the local solvent composition would have to be included.
A mathematical model of vacuum contact drying of wet powder
beds was implemented and used within a model-based scale-up
framework for the estimation of drying times at the pilot and
Acknowledgements
manufacturing scale. The model is physically based and it contains
only parameters that are scale-independent, hence no additional
The support of Matthew Johnson for the crystal density
parameter fitting or re-estimation is required during scale-up. A
measurement is gratefully acknowledged.
methodology for the measurement of all required parameters at the
laboratory scale has been described in this work and demonstrated
on the scale-up of a wide range of compounds, solvents, and
References
operating conditions (pressure, temperature). The model has also
been applied to systems operated in the intermittent agitation mode.
Daian, J.-F., Saliba, J., 1993. Transient moisture transport in a cracked porous
The model used in this work is ‘‘minimal’’ in the sense that many medium. Transp. Porous Media 13, 239–260.
phenomena that are also known to occur during vacuum contact Kemp, I.C., 2007. Drying software: past, present, and future. Drying Technol. 25,
drying of pharmaceutical compounds have not been included. As 1249–1263.
Kohout, M., Collier, A.P., Stepanek, F., 2004. Effective thermal conductivity of wet
discussed in Section 2.1 above, these phenomena could include
particle assemblies. Int. J. Heat Mass Transfer 47, 5565–5574.
specific solvent–solid interactions (desorption, desolvation), changes Kohout, M., Collier, A.P., Stepanek, F., 2006. Mathematical modelling of solvent
in the cake structure during drying (particle attrition or breakage), or drying from a static particle bed. Chem. Eng. Sci. 61, 3674–3685.
non-uniform initial moisture distribution in the cake due to imperfect Kohout, M., Stepanek, F., 2007. Multi-scale analysis of vacuum contact drying.
Drying Technol. 25, 1265–1273.
deliquoring. However, the structure of the balance equations allows Kosek, J., Stepanek, F., Marek, M., 2005. Modelling of transport and transforma-
for future extension of the model. For example, the model can be tion processes in porous and multi-phase bodies. Adv. Chem. Eng. 30,
integrated into a process-scale simulation that also takes into con- 137–203.
Lehmann, P., Assouline, S., Or, D., 2008. Characteristic lengths affecting evaporative
sideration phenomena such as multiple vessels connected to the drying of porous media. Phys. Rev. E 77, 056309.
same vacuum line, pressure fluctuations, etc. As can be seen in many Mahadevan, J., Sharma, M.M., Yortsos, Y.C., 2006. Flow-through drying of porous
simulation outputs (Fig. 10) the first part of the drying curve is very media. AIChE J. 52, 2367–2380.
Metzger, T., Irawan, A., Tsotsas, E., 2007. Influence of pore structure on drying
steep suggesting that the evaporation rate is very high. This is usually
kinetics: a pore network study. AIChE J. 53, 3029–3041.
due to the assumption that the pressure is constant throughout Michaud, A., Peczalski, R., Andrieu, J., 2008a. Modeling of vacuum contact drying of
drying and the set point is achieved from the beginning of the drying. crystalline powders packed beds. Chem. Eng. Process. 47, 722–730.
However, the pressure does not drop instantaneously to the set point Michaud, A., Peczalski, R., Andrieu, J., 2008b. Optimization of crystalline powders
vacuum contact drying with intermittent stirring. Chem. Eng. Res. Des. 86,
but it usually drops slowly to it depending on the pump efficiency 606–611.
and plant design. In order to model the pressure pathway, phenom- Shahidzadeh-Bonn, N., Azouni, A., Coussot, P., 2007. Effect of wetting properties on
ena beyond the dryer boundary have to be considered in the the kinetics of drying of porous media. J. Phys.: Condens. Matter 19, 112101.
Smallwood, I.M., 2006. Handbook of Organic Solvent Properties. Elsevier,
simulation. Amsterdam.
The model can also be extended to address the cases where Wei, C.K., Davis, H.T., Davis, E.A., Gordon, J., 1987. Evaporation from porous
mixed solvents are used. Thermodynamic equilibria have to be media: a capillary model of a penetrating front. Chem. Eng. Commun. 56,
solved in this case in order to determine the composition of the 269–284.
Whitaker, S., 1977. Simultaneous heat, mass and momentum transfer in porous
solvent at each point in time and space. In addition, the apparent media: a theory of drying. Adv. Heat Transfer 13, 119–203.
moisture diffusivity is likely to vary with time due to different Whitaker, S., 1999. The Method of Volume Averaging. Kluwer, Dordrecht.