mmc2 PDF

Article
Systematically Improving Espresso: Insights

from Mathematical Modeling and Experiment
Michael I. Cameron, Dechen
Morisco, Daniel Hofstetter, ...,
William T. Lee, Christopher H.
Hendon, Jamie M. Foster
[email protected] (C.H.H.)
[email protected] (J.M.F.)
HIGHLIGHTS
Development of a model for
extraction of espresso
Experimental observation of
inhomogeneous extraction
Optimization of espresso
parameters to minimize coffee
waste
Implementation to yield monetary

savings in a cafe setting
Coffee extraction performed in an espresso machine is a process that depends on

a delicate interplay between grind setting, coffee mass, water pressure and
temperature, and beverage volume. Using a mathematical model based on the
transport of solubles through a granular bed, paired with cafe-setting
experiments, we elucidate the origin of inconsistencies in espresso production.
Informed by the model, a protocol is proposed to systematically reduce coffee
waste and beverage variation, resulting in highly reproducible shots.
Cameron et al., Matter 2, 1–18

March 4, 2020 ª 2019 Elsevier Inc.
https://doi.org/10.1016/j.matt.2019.12.019
Please cite this article in press as: Cameron et al., Systematically Improving Espresso: Insights from Mathematical Modeling and Experiment,
Matter (2019), https://doi.org/10.1016/j.matt.2019.12.019
Article
Systematically Improving
Espresso: Insights from Mathematical
Modeling and Experiment
Michael I. Cameron,1,10 Dechen Morisco,1,10 Daniel Hofstetter,2 Erol Uman,3 Justin Wilkinson,4
Zachary C. Kennedy,5 Sean A. Fontenot,8 William T. Lee,6,7 Christopher H. Hendon,8,11,*
and Jamie M. Foster9,*
SUMMARY Progress and Potential

Espresso is a beverage brewed using hot, high-pressure water forced through a The modern coffee market aims to
bed of roasted coffee. Despite being one of the most widely consumed coffee provide products which are both
formats, it is also the most susceptible to variation. We report a novel model, consistent and have desirable
complimented by experiment, that is able to isolate the contributions of several flavour characteristics. Espresso,
brewing variables, thereby disentangling some of the sources of variation in one of the most widely consumed
espresso extraction. Under the key assumption of homogeneous flow through coffee beverage formats, is also
the coffee bed, a monotonic decrease in extraction yield with increasingly the most susceptible to variation
coarse grind settings is predicted. However, experimental measurements in quality. Yet, the origin of this
show a peak in the extraction yield versus grind setting relationship, with lower inconsistency has traditionally,
extraction yields at both very coarse and fine settings. This result strongly sug- and incorrectly, been attributed to
gests that inhomogeneous flow is operative at fine grind settings, resulting in human variations. This study’s
poor reproducibility and wasted raw material. With instruction from our model, mathematical model, paired with
we outline a procedure to eliminate these shortcomings. experiment, has elucidated that
the grinder and water pressure
INTRODUCTION play pivotal roles in achieving
The past century has seen an increase in the prevalence of coffee consumption, beverage reproducibility. We
as consumers have gained an appreciation for its complex and exciting flavors, suggest novel brewing protocols
and obvious psychological effects.1–7 As a result, the coffee industry contributes that not only reduce beverage
significantly to the economic stability of numerous producing and consuming variation but also decrease the
countries. For example, in 2015, the American coffee industry provided over mass of coffee used per espresso
1.5 million jobs, accounting for $225.2 billion (1.6% gross domestic product), by up to 25%. If widely
and resulting in ca. $30 billion in tax revenue.8 However, coffee-producing coun- implemented, this protocol will
tries now face new challenges owing to changing climate9–12 and shifts in con- have significant economic impact
sumer preferences. These challenges highlight the need to maximize the quality and create a more sustainable
and reproducibility of the beverage while minimizing the mass of coffee used to coffee-consuming future.
produce it.
Of all of the coffee formats, espresso is by far the most complicated and susceptible
to fluctuations in beverage quality. As historically defined by the Specialty Coffee
Association, an espresso is a 25–35 mL (ca. 20–30 g) beverage prepared from
7–9 g of ground coffee made with water heated to 92 C–95 C, forced through the
granular bed under 9–10 bar of static water pressure and a total flow time of
20–30 s. These metrics have been grandfathered into the industry and are signifi-
cantly detached from the recipes used in most cafes today. Coffee shops routinely
favor higher dry coffee mass (15–22 g), resulting in larger volume beverages
(30–60 g beverage mass), produced on machines that dynamically control both
Matter 2, 1–18, March 4, 2020 ª 2019 Elsevier Inc. 1

water pressure and temperature. The variables of tamp force, flow rate or time, dry
mass of coffee, and beverage volume are all determined by the machine’s operator.
There are other variables that have an impact on the beverage quality prior to the
ground coffee being exposed to water. The grind setting determines the particle
size distribution of the coffee grounds (and therefore the surface area).13 Once com-
pacted into a granular bed, the particle size distribution plays a role in controlling the
permeability of the bed and consequently the flow rate. A decreased flow rate can
be achieved in a number of ways: by decreasing the water pressure, grinding finer,
packing the bed more tightly, using more coffee, or some combination of these. A
further source of variability is that roasted coffee ages through off-gassing, losing
roast-generated volatiles thereby altering the resultant beverage density and
flavor.14,15
In principle, it is preferable to make objective statements about the flavor of food-

stuffs from knowledge of their molecular components. This poses problems for cof-
fee because there are 2,000 different compounds extracted from the grounds dur-
ing brewing.16,17 In practice, we are limited to more easily measurable descriptors.
The coffee industry uses extraction yield (EY), a ratio of solvated coffee mass to the
mass of dry coffee used to produce the beverage, to assess extraction. EY is calcu-
lated by first measuring the refractive index, a property that depends on tempera-
ture. While a refractive index measurement cannot be used to characterize the
beverage composition (i.e., it cannot be used to make qualitative statements about
chemical composition; the refractive response is highly molecule specific),18 it has
been shown to accurately correlate with extracted mass.19 This turn may be related
to flavor for a narrow range of brew parameters; we discuss this further in subsequent
sections. Accordingly, the Specialty Coffee Association advises that coffee most
frequently tastes best when the proportion of extracted dry mass is in the range
17%–23%. Coffee beverages with EYs exceeding 23% typically taste bitter, while
those below 17% are often sour. Furthermore, concentration (often referred to as
beverage strength) plays another key role in coffee beverage production. Here, 1FriskyGoat Espresso, 171 George St., Brisbane
we consider this a secondary problem and chose to monitor EY because it is still a City, QLD 4000, Australia
2Daniel Hofstetter Performance,
descriptor of flavor but also has significant economic implications (i.e., it tells us
Laenggenstrasse 18, Bachenbuelach 8184,
something about how efficiently we are using our coffee mass). In contrast, one Switzerland
could argue that beverage concentration is related to the consumer’s preference. 3Meritics Ltd., 1 Kensworth Gate, Dunstable LU6
3HS, UK
4Faculty
of Mathematics, University of
This paper reports the development of a multi-scale mathematical model for extrac-
Cambridge, Cambridge CB3 0WA, UK
tion from a granular bed. Here, multi-scale is used to emphasize the fact that the de-
5National Security Directorate, Pacific Northwest
scriptions of the physics spans different length scales (i.e., the size of the coffee National Laboratory, Richland, WA 99352, USA
grain, which is much smaller than the size of the espresso bed).20 We apply the 6Department of Computer Science, University of
model to espresso-style coffee extraction but note that it is readily generalizable Huddersfield, Huddersfield HD1 3DH, UK
to any liquid/granular biphasic system. The model offers scope to independently 7MACSI, Department of Mathematics and
Statistics, University of Limerick, Limerick, Ireland
alter familiar variables such as grind setting, water pressure, flow rate, coffee
8MaterialsScience Institute and Department of
dose, extraction kinetics, and so forth; these culminate in a prediction of EY. The
Chemistry and Biochemistry, University of
model’s ability to individually change each brewing parameter is crucial to devel- Oregon, Eugene, OR 97403, USA
oping enhanced understanding of brewing because altering parameters truly inde- 9Schoolof Mathematics & Physics, University of
pendently is difficult in an experimental setting. Portsmouth, Portsmouth PO1 2UP, UK
10Present address: ST. ALi Coffee Roasters, 12-18
Yarra Place, South Melbourne, VIC 3205, Australia
The model enables us to understand the origin of irreproducibility in espresso 11Lead Contact
(namely non-uniform flow), and it also informs us in proposing a novel strategy for
*Correspondence:
minimizing drink variation as well as dry coffee waste. We identify a critical minimum [email protected] (C.H.H.),
grind size that allows for homogeneous extraction. Below this setting, a counterintu- [email protected] (J.M.F.)
itive reduction in EY and increase in variability is observed. In a departure from the https://doi.org/10.1016/j.matt.2019.12.019
2 Matter 2, 1–18, March 4, 2020

Pore space / water Pump Coffee grains

P|z=0 = Ptot
Inlet
r
z
Outlet
P|z=L = 0
R0
Fines Boulders
Figure 1. A Schematic of the Espresso Basket Geometry

The coffee grounds are shown in gray (Us ), and the pore space, which is filled with water during
extraction, is shown in blue (Ul ). The macroscopic spatial coordinate measuring depth through the
bed, z, the microscopic spatial coordinate measuring radial position within the spherical coffee
particles, r, as well as the basket radius, R0 , are also indicated.
Specialty Coffee Association recommendations, the model suggests that we should

ignore brew time and navigate the EY landscape using only mass of coffee and mass
of water as independent variables. We demonstrate that we are able to systemati-
cally reduce coffee waste and dramatically reducing shot variation, while also saving
the cafe both time and money in their production of espresso-based beverages. Our
approach is then implemented into a real cafe setting where the economic benefits
were monitored. Using these data, paired with those previously reported,8,21 we
estimate that a 25% reduction in coffee mass per coffee beverage will result in
approximately ca. $3.1 million per day.
DEVELOPMENT OF A RATIONAL MODEL FOR ESPRESSO EXTRACTION

Espresso is brewed in a cylindrical container denoted by z˛ð0; LÞ and R˛ ð0; R0 Þ (Fig-
ure 1). The solid coffee grounds occupy part of the cylinder, Us and contain a concen-
tration of soluble coffee, cs . The cylinder also contains inter-granular pore space, Ul ,
which is occupied by liquid during extraction, which itself contains a concentration of
coffee solubles cl . We use the term coffee solubles to denote the sum of the concen-
trations of all compounds in coffee; this is in line with the EY measurement. We note
that the model could readily be generalized to explicitly track any number of chem-
icals. However, the utility of doing so is questionable because one would need to
also provide or fit kinetic parameters for each individual compound, rendering the
model susceptible to overfitting. We would also require knowledge of initial concen-
trations, and these are difficult to measure for many species. The liquid flow between
the inlet ðz = 0Þ and outlet ðz = LÞ is driven by an overpressure (the pressure excess
relative to atmospheric pressure, Ptot ), applied by a pump. The model equations
take the form of a system of partial differential equations that describe (1) the trans-
port of coffee solubles from the interior of the grounds to their surface, (2) the ex-
change or dissolution of the solubles from the grounds into the liquid, and (3) the
migration of the solubles in the liquid by a combination of diffusion and advection.
The solubles in the liquid phase are transported by a combination of diffusion and
convection due to the flow of the liquid through the bed. The concentration of sol-
vated coffee is therefore governed by an advection-diffusion equation:
vcl
= V,ðDl Vcl ucl Þ in Ul ; (Equation 1)
vt
Matter 2, 1–18, March 4, 2020 3

where t, Dl , and u are time, the diffusivity of solubles within the liquid, and the veloc-
ity of the liquid, respectively. The liquid flow is solved for via the Navier-Stokes
equations
vu 1 m
+ ðu , VÞu = VP + V2 u; (Equation 2)
vt r r
V,u=0 in Ul ; (Equation 3)
where m, r, and P are liquid viscosity, density, and overpressure, respectively.
Work by Spiro and colleagues demonstrated that the transport of coffee solubles
through the interior of the grounds can be described by a diffusion process.22–28
Hence,
vcs
= V,ðDs Vcs Þ in Us ; (Equation 4)
vt
where Ds is the diffusivity of solubles within the grains. Here, we treat coffee grounds as
spherical dense particles, but we note that the coffee grains themselves may be irregu-
larly shaped and feature intragranular macropores, as previously observed in scanning
electron micrographs.29 As discussed in the next section, most particles in ground cof-
fee are smaller the macropore diameter observed in the micrographs. Nitrogen physi-
sorption was used to assess the microporosity of the coffee grounds; the data suggest
that ground coffee does not feature microporosity (see Supplemental Information).
Thus, we expect our description to hold for most espresso grind settings.
Boundary conditions at the inlet, z = 0, include a specified fluid overpressure, the

requirement that the water enters the basket with a purely normal velocity, and
that the normal flux of dissolved species should be zero:
Pjz = 0 = Ptot ; (Equation 5)
u,bt jz = 0 = 0; (Equation 6)
ð Dl Vcl + ucl Þ,b

n jz = 0 = 0; (Equation 7)
where bt and b n are the unit vectors tangent and normal to the surface z = 0, respec-
tively. At the exit we apply conditions of zero overpressure, zero tangential velocity,
and a condition that there is zero diffusive flux of coffee. In summary,
Pjz = L = 0; (Equation 8)
u,bt jz = L = 0; (Equation 9)
ð Dl Vcl Þ,b
n jz = L = 0: (Equation 10)
At the vertical edges of the cylinder, R = R0 , no flux conditions are applied to the
liquid coffee concentration, because the liquid cannot exit in these directions:
ð Dl Vcl + ucl Þ,b

n jR = R0 = 0; (Equation 11)
ujR = R0 = 0: (Equation 12)
On the boundaries between the grains and inter-granular pore space, Gint , there is a
flux of solubles per unit area, which we denote by G. Appropriate boundary condi-
tions are
ð Ds Vcs Þ , b
n = G; (Equation 13)
4 Matter 2, 1–18, March 4, 2020

ð Dl Vcl + ucl Þ , b
n = G; (Equation 14)
u = 0 on Gint ; (Equation 15)

where the former two capture mass transfer and the latter imposes that the liquid
should be stationary on the grain/pore space interface.
Determining the form of the reaction rate, G, is non-trivial, and it is something that is
not readily measured experimentally. However, it can be reasonably assumed that
the rate of transfer of solubles between the phases should depend on the local con-
centrations of solubles near the interface. Furthermore, the rate of extraction is zero
when (1) the liquid immediately outside the grain is saturated (i.e., at a concentration
csat ) or (2) when the liquid outside the grain is at the same concentration as the grain
(i.e., in equilibrium) or (3) when the grain is depleted of solubles (the experimental
upper limit of extraction is approximately 30% by mass). We therefore postulate a
rate that satisfies all of the above conditions, namely
G = kcs ðcs cl Þðcsat cl Þ on Gint ; (Equation 16)

where k is a rate constant. We note that the quantity csat likely depends on the local
temperature. One could readily incorporate a thermal model into the description,
but here we assume that the espresso basket is isothermal. This is justified on the ba-
sis that the heat capacity of water is relatively high and that espresso basket temper-
atures are actively controlled in most machines.
Coffee particulates remain dry until they are connected to the extraction apparatus,
at which point water is rapidly introduced to the bed, serving to wet the entire puck
and stabilize the particle temperature. Modeling this initial wetting (i.e., pre-infu-
sion) stage poses another series of interesting problems; the model presented
here is only valid once liquid infiltration has taken place, and we refer the interested
reader to a discussion on pre-infusion.30 We avoid explicitly modeling this stage by
assuming that at t = 0, when extraction begins, the bed is filled with liquid water that
is free from solubles. We therefore have
cl jt = 0 = 0; cs jt = 0 = cs0 (Equation 17)

and note that the errors engendered in making this approximation can be ex-
pected to be small because the intrusion stage represents only a small portion
of the overall extraction time. In Equation 17, cs0 is the concentration of solubles
in the grains initially. Concurrent with the wetting stage is the potential for the
grains in the bed to be rearranged by the invading fluid.28 Rearrangement that
may occur during the initial wetting stage will be accounted for later after the
equations have been homogenized. One of the results of this procedure is that
the geometry is encapsulated in the macroscopic quantity of permeability, and
by making this material property inhomogeneous, the model can mimic a non-uni-
form distributions of grounds.
Particle-Size Distribution of Ground Coffee

The model requires knowledge of the distribution of coffee particle sizes produced
by the grinder. The population, surface area, and volume fraction of the particles are
used to estimate the permeability of the bed, and this is crucial in determining the
liquid flow. Moreover, the particle size controls the extraction dynamics, because
it determines the typical distance (and in turn the typical time) over which solubles
must be transported within the grains before they reach the interface where they
can be dissolved into the liquid.
Matter 2, 1–18, March 4, 2020 5

A
contribution
total
Relative
surface area
counts
B
fines boulders
20 GS
2.5
boulders large particles are
exit grinder further ground
Volume percent (%)
2.0
15
1.5
10 1.0
fines exit
grinder
5
0
1 10 100 1000
Particle diameter (µm)
Figure 2. Particle Size Distributions Collected Using the Method Described in the Experimental
Procedures
(A) Surface area and number of coffee particulates produced with a grind setting G S = 2.5. Here,
99% of the particles are <100 mm in diameter and account for 80% of the surface area.
(B) The volume percent particle size distribution at G S = 2.5, 2.0, 1.5, and 1.0. Grinding finer reduces
the average boulder size and increases the number of fines. Intruders are boulders that are larger
than the aperture of the burr set and hence further fractured until they can exit the burrs.
Particle size distributions were measured using our described experimental proced-
ure; these data are shown in Figure 2A. We observe that, similar to the formation of
two families of particle sizes found in exploding volcanic rock,31 there are two groups
of particle sizes in ground coffee. Namely, boulders (which we define as larger than
100 mm) and fines (smaller than 100 mm in diameter). This bimodal distribution is
caused by large particles fracturing until they are sufficiently small to exit through
the grinder burr aperture.13 The size of the boulders are determined by the burr sep-
aration, whereas the fines (much smaller than the burr aperture) are thought to be
produced at the fracture interface. One piece of evidence supporting this idea is
that as the grind setting, GS, is reduced, the relative proportion of fines increases,
but their size remains constant (Figure 2B).
Multi-scale Homogenization and One-Dimensional Reduction

Direct solution of Equations 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, and 17 on
a realistic packed bed geometry comprising many millions of individual grains is
intractable, even using modern high-performance computing. Therefore, rather
than tackling the problem directly, we make use of the vast disparity in the length
scales between that of a coffee grain (10 mm, referred to as the microscopic scale)
and that of the whole bed (1 cm, referred to as the macroscopic scale) to
6 Matter 2, 1–18, March 4, 2020

systematically reduce the system using the asymptotic technique of multiple scales
homogenization. Such techniques have been applied to problems with a similar
structure in electrochemistry,32 and rather than present this very involved calculation
in full here, we provide a summary in the Supplemental Information and refer the
interested reader to Richardson and co-workers33 where the details of an analogous
calculation are presented.
The macroscopic system of equations, valid on the larger macroscopic length scale
of the entire bed, systematically follow from the microscopic Equations 1, 2, 3, 4, 5,
6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, and 17. The application of the multiple scales
technique significantly reduces the model complexity, but the requisite information
about the microscale variations is retained. For example, because the dissolution
rates depend on the concentration of solubles on the microscopic particle surfaces,
the multi-scale system contains a series of microscale transport problems that must
be solved inside representative grains. It is crucial that this microscopic information
is preserved in the multi-scale model, because it will allow us to study the effects of
different grind settings on the overall macroscopic behavior of the extraction.
Motivated by the bimodal distribution of particle sizes in the model, it may be assumed
that the bed is composed of two families of spherical particles with radii a1 (fines) and a2
(boulders). Further, we denote the Brunauer-Emmett-Teller (BET) surface area of the
different families of particles by bet1 and bet2 , respectively. The BET surface area char-
acterizes the amount of interfacial surface area between two intermingled phases per
unit volume of the mixture, and therefore has units of 1=m. We also introduce cs1 and
cs2 to denote the concentrations of solubles in the two particle families. The resulting
macroscopic equation for the concentration of solubles in the liquid is

vc v vc
ð1 fs Þ l = Deff l qc l + bet;1 G1 + bet;2 G2 : (Equation 18)
vt vz vz
Here, the quantity c l is the concentration of solubles in the liquid as predicted by the
multi-scale modeling approach; whereas cl appearing in the equations in the Sup-
plemental Information is the concentration of solubles in the liquid as predicted
by the original microscopic model. The upscaled and reduced versions of Equations
7 and 10 are

vc
Deff l + qc l = 0; (Equation 19)
vz z=0

vc l
Deff = 0: (Equation 20)
vz z = L
These assert that there should be no flux of solubles across the inlet and no diffusive
contribution to the flux at the outlet. In the next section, we show that parameter es-
timates indicate that diffusive fluxes are negligible compared with those due to
advection in typical espresso brewing conditions. Hence, it is the flow of the liquid
through the pores that is primarily responsible for moving solubles through the
bed once they have been dissolved. Hence, even though the physical relevance of
the latter condition in Equation 19 is not completely clear, it has negligible impact
on the model solution.
The microscopic equations to be solved are

vcsi 1 v vcsi
= 2 r 2 Ds ; for i = 1; 2; (Equation 21)
vt r vr vr
Matter 2, 1–18, March 4, 2020 7

and the symmetry and dissolution rate boundary conditions, which arise from Equa-
tion 13, and which act to couple the micro- and macroscale transport problems, are
9
vcsi
Ds =0 > >
vr r = 0 =
for i = 1; 2: (Equation 22)
vcsi >
>
;
Ds = G
vr r = ai
i
The problem is closed by supplying the initial conditions Equation 17 and

the reaction rates, Gi . This has precisely the same form as Equation 16
but with additional subscripts to differentiate the boulders from the fines, i.e.,
Gi = kcsi ðcsi c l Þðcsat c l Þ.
A formula for EY in terms of the model variables can be derived by first noting that it
follows from Equations 18 and 20 that an expression for the rate at which soluble
mass enters the cup is given by
dMcup
= pR 20 qcl z = L : (Equation 23)
dt
On integrating this equation along with the initial condition that there is no solvated
mass at t = 0 and dividing by the dry mass of coffee initially placed in the basket, Min ,
we obtain
Rt
pR 20 q 0shot cl jz = L dt
Extraction yield ðEYÞ = ; (Equation 24)
Min
where EY is described as the fraction of solvated mass compared with the total mass
of available coffee. Here, tshot is the flow time. Equation 24 is used in the following
sections as a means to compare model predictions of EY with experimental
measurements.
Tuning the Model to Espresso Extraction Data

Initially, simulations of espresso extraction were run using a cafe-relevant recipe of
20 g of dry grounds used to produce a 40 g beverage under 6 bar of static water
pressure. Values for the radius of an espresso basket (R0 ), the viscosity of heated wa-
ter (m), the saturation concentration of heated water (csat ), and the concentration of
solubles initially in the grounds (cs0 ) are readily available in the literature (see the
Supplemental Information for a summary of values and their sources). Moroney
and colleagues30 estimate that the volume fraction of grounds in a packed bed is
fs = 0:8272 and this, along with the density of grounds and the bed radius, allows
us to derive a value for the bed depth via the relationship
Min
pR 20 L = f: (Equation 25)
rgrounds s
While it is likely that bed depth varies slightly across the range of grind settings (as
the volume fraction changes), we assume that the bed depth is constant for a given
dry mass of coffee, Min . Values for both the radii and BET surface area for the two
differently sizes families of particles in the grounds can be derived from the data
shown in Figure 3 provided that both families are distributed homogeneously
throughout the bed. The Darcy flux, q, determines the flow rate of the liquid through
the bed and varies with the grind setting. They are estimated using the shot times
presented in Figure 4 from the equation
Mout
q= ; (Equation 26)
pR 20 rout tshot
8 Matter 2, 1–18, March 4, 2020

A 28
impact of dose on EY
27 6 bar / 40 g out
26
16 g
Extraction yield (mass %)
25
18 g
24
23 20 g
22 22 g
21 24 g
20
19
18
1.1 1.3 1.5 1.7 1.9 2.1 2.3
EK 43 grind setting, GS
B
26
impact of pressure on EY
25 20 g in / 40 g out
4 bar
24
6 bar
23
22 8 bar
21 10 bar
20
19
18
1.1 1.3 1.5 1.7 1.9 2.1 2.3
Figure 3. Extraction Yield as a Function of Grind Size, with Varying Coffee Dose and Water
Pressure
(A) The effect of changing the coffee dose Min with constant water pressure shows that reducing the
initial coffee mass but keeping the beverage volume constant results in higher extractions.
(B) The effect of changing the pump overpressure, P, with a constant brew ratio shows an increase in
extraction yield with decrease in water pressure.
where Mout is the mass of the beverage (40 g), and we make the assumption that the
density of the drink, rout , is the same as that of water, but we note that this is an area
that could be improved in future model developments. We emphasize the difference
between Mout and Mcup ; the former is the total mass of the beverage, whereas Mcup
(used in Equation 23) is the total mass of solubles in the beverage. The parameter
values discussed above are tabulated in the tables presented in the Supplemental
Information.
Matter 2, 1–18, March 4, 2020 9

A 45
R2 = 0.995
Shot time (tshot, s)
40
35
30
25
P = 6 bar
20
B
Regime 2: Partially clogged flow
23
22
21
Regime 1: Standard flow

20
1.1 1.3 1.5 1.7 1.9 2.1 2.3
Figure 4. Espresso Extraction Yield as a Function of Grind Setting

(A) PW = 6 bar, t F = 98 N shot times are inversely proportional to G S .
(B) Extracted mass percent can be described by two regimes. Regime 1: a standard flow system
where an expected increase in extraction percent is observed with reducing G S . Regime 2: partially
clogged flow is operative when there are too many fines (ca. G S = 1.7), forming aggregates and/or
inhomogeneous bed density, effectively reducing the surface area of the granular bed.
Three parameters, namely Deff , Ds , and k, remain to be specified. The

effective macroscopic diffusivity in the liquid is often related to the diffusivity Dl
via Deff = BDl where B is the permeability factor. This accounts for the reduction in
the diffusive fluxes due to the obstacles provided by the intermingled phase, in
this case the coffee grains. This permeability factor can either be computed via a
homogenization calculation34 or can be estimated using the Bruggemann approxi-
3=2
mation, which asserts that B = εl .35 Unfortunately, neither Dl nor Deff have been
experimentally characterized. However, if we adopt a value for the diffusivity of a
typical compound in water and then compare the expected size of the flux of solu-
bles due to diffusion versus convection, it seems clear that a safe conclusion is
that the former is significantly smaller than the latter:
Deff csat
qcsat : (Equation 27)
L
We note that this same conclusion was also reached previously.30 Henceforth, we
assign a small value to the macroscopic diffusivity of solubles in the liquid so that
diffusive transport is negligible compared with convection due to liquid flow. The
final two parameters, Ds and k, are fitted to the experiment. The results of this fitting
are shown in Figure 4 and lead to values of
10 Matter 2, 1–18, March 4, 2020

2
Ds = 6:25 3 1010 m2 s; k = 6 3 107 m7 kg s1 : (Equation 28)
The Effect of Altering the Brew Ratio and Water Pressure

Under the assumption that the bed geometry depends only on the grind setting, it is
straightforward to explore the role of altering both the dry coffee mass (i.e., the brew
ratio) and the static water pressure. When the coffee mass is altered, the only param-
eter that needs to be altered is the bed depth, L. We continue to operate under the
assumption that the bed has a fixed volume fraction of coffee grounds, and hence
the depth of the bed is directly proportional to the coffee dose (see Equation 25
and the Supplemental Information). When the pump overpressure (Ptot ) is increased,
the Darcy flux (q) is increased in direct proportion,34 whereas the shot time is
decreased in an inverse proportional manner (Equation 26 and the Supplemental In-
formation). We monitor EY (the mass fraction of grains that are dissolved), enabling
the isolation of both coffee mass and overpressure. These results are shown in
Figure 3.
BREWING ESPRESSO
The model predicts that EY can be increased by grinding finer, using lower
pressure water, and/or using less coffee. The model was then compared with
experimental coffee brewing, performed in a cafe setting. Using an espresso ma-
chine set to P = 9 bar of water pressure resulted in clogging at fine grind settings
(see the Supplemental Information). To circumvent this difficulty, the pressure was
reduced to P = 6 bar. Coffee must be tamped to level the granular bed. We explored
a range of tamp pressures but did not observe an appreciable variation in shot time
or EY, and so we standardized our tamping procedure using an automated device
that pressed the bed at 98 N (see the Supplemental Information). Flavor differences,
however, were noted but not quantified. The barista needs to taste the coffee in the
cafe setting to ensure the beverage has the qualities that they desire. In summary,
however, lower water pressure and tamp force allowed for access to a wider range
of grind settings, thereby allowing for systematic sampling of shot time and EY
over all relevant espresso grind sizes (Figure 4). These results indicate that the rela-
tionship between shot time is linear (Figure 4A); with a coarser grind setting resulting
in shorter shot times.
Examination of the extracted mass of coffee as a function of grind setting reveals a

more puzzling outcome (Figure 4B). From our model, it was anticipated that
decreasing the grind setting should increase the extracted mass because the grinder
(1) produces more fines, yielding higher surface area, (2) produces smaller boulders,
reducing the length of transport pathways for solubles from their interior to their sur-
face, and (3) increases shot times and, in turn, contact time and allows more time for
dissolution of coffee compounds. This counterintuitive decrease in EY with grind set-
tings less than 1.7 indicates that regions of the bed are not being evenly extracted
(i.e., flow is no longer homogeneous). The EY measurements are made based on a
sample of the beverage from the brewed cup of coffee and are hence indicative
of an ‘‘average’’ EY of the grains throughout the bed. Thus, the onset of nonhomo-
geneous flow should be accompanied by a perceived mixture of both under- and
overextracted flavors; this experience is very familiar in specialty coffee. Consumers
may describe the same espresso as tasting both bitter and acidic (orthogonal flavors
originating from dissimilar chemical motifs, particularly detectable as the coffee
cools).36,37 The industry often uses grind settings less than 1.7, in part to hit the
time targets described in the definition of espresso.
Matter 2, 1–18, March 4, 2020 11

25
experimental mean
±1 std. deviation
Standard flow
24
Partially clogged flow
23
22
21
1.1 1.3 1.5 1.7 1.9 2.1 2.3

Grind setting (GS)
Figure 5. Simulated and Experimental Espresso Extraction
A comparison of the model and experimental espresso shots collected using a standard recipe of
20 g of coffee dry mass, 40 g of beverage mass, produced using 6 bar water pressure. The partially
clogged and standard flow regimes are highlighted in orange and blue, respectively.
The non-monotonic trend in EY with GS can be attributed to two competing extrac-

tion regimes, namely (1) the expected flow conditions where extraction increases as
the coffee is ground finer, and (2) aggregation and/or inhomogeneous density in the
bed causing partially clogged flow and reduced average extraction. This highlights a
fundamental problem in correlating a coffee beverage with refractive index mea-
surements because there are countless ways to achieve a given EY. For example,
examining our data reveals that one can obtain two 22% EY shots by keeping the
brew ratio fixed and setting the GS to either 1.3 or 2.0. The chemical composition
of the faster shot cannot be the same as the slower shot owing to molecular differ-
ences in solubility, dissolution rate, and resultant molecule-dependent impact on
the refractive index.38 This result does not undermine our use of EY, but rather illus-
trates that the barista indeed needs to taste the coffee, rather than measure its sol-
vated mass.
Our results provide an avenue to tackle three highly relevant issues in the coffee in-
dustry: first, how can one improve espresso reproducibility given the non-linear
dependence of EY on the grind setting? Second, what should one do to reduce
shot time or EY variability? And third, can we systematically improve espresso repro-
ducibility while minimizing coffee waste? We address these questions later.
ACCOUNTING FOR THE PARTIALLY CLOGGED FLOW REGIME

The discrepancy between the model and experiment for finely ground coffee (GS<
1.7; Figure 5) allows us to estimate the amount of the bed that is effectively inacces-
sible due to clogging. For extraction in the partially clogged regime, the flow is non-
uniform, and this causes some regions of the granular bed to be more extracted than
others. Without a precise map of the bed geometry, it is difficult to characterize the
partially clogged flow pattern precisely. However, one can imagine an extreme case
12 Matter 2, 1–18, March 4, 2020

of partial clogging in which some regions of the coffee bed have zero flow (and are
therefore non-extracted), while the remainder is subject to homogeneous flow.
Based on a comparison between the model prediction and the experimental mea-
surement, and assuming this extreme case of partial clogging (i.e., there are regions
in the coffee bed that are entirely dry), we find that there is a 13.1%, 6.1%, and 2.6%
difference in predicted EY versus experimental values for grind settings of 1.1, 1.3,
and 1.5, respectively. Of course, it is likely that the clogging results in a spread of
EYs, with a portion of the mass extracted to 0%, 1%, 2%, etc. Thus, estimating the
level of inefficiency in a given extraction is not possible using the refractive index
measurement alone, and we hope that follow-up studies will provide molecular han-
dles on the disparities in flavor comparing homogeneous and partially clogged ex-
tractions with identical EYs.
We are able to adapt the model to recover this downward trend after reaching the
critical grind setting. In essence, this is achieved by reducing the accessible surface
area of a fraction of the granular bed, simply modeled by reducing the surface area
of the planar faces of the cylindrical bed that is exposed to incident water. This pro-
cedure was performed empirically until our predicted EY matched the experimental
data. More important, however, are the implications of the need to reduce the pro-
portion of coffee that is accessible in the model. First, this suggests that in many cir-
cumstances where flow is inhomogeneous, there are regions of the granular bed that
have been extracted far higher than measured with an average EY. Second, the grind
setting plays a major role in determining how much dry coffee mass is wasted in the
brewing process. Finally, marriage of the model and the experiments provide us with
one clear avenue to optimize espresso extraction.
SYSTEMATICALLY IMPROVING ESPRESSO

Since the extent of the clogging is determined by the size distribution in the grind,
and the variation in how different coffees grind are negligible,13 it is reasonable to
expect all coffees to exhibit this inhomogeneous ‘‘peaked’’ relationship between
EY and grind setting for a fixed pump pressure and brew ratio. We therefore use
the insights gleaned by comparing the model with experiments above to make
generic recommendations on how to address the important economic aims of maxi-
mizing EYs, while simultaneously producing enjoyable espresso and reducing dry
coffee waste. The following sections discuss several approaches to achieving these
goals.
Maximizing Extraction Yield by Altering GS

The grind setting that gives rise to the maximum extraction yield, EYmax, for a given
set of brewing parameters should correspond to the finest grind that maintains ho-
mogeneous extraction from the coffee bed. Since, by definition, EYmax is greater
than or equal to the EY obtained for the extraction parameters used in cafes, the ba-
rista can always find their targeted EY by first extracting at EYmax, then changing the
brew ratio by increasing the volume of water used to produce the shot. In practice,
this is achieved by first locating a so-called tasty point (i.e., an espresso shot that
tastes good to the barista), then locating the grind setting corresponding to EYmax
using either the refractive technique detailed here or by simply tasting the coffee.
Following Figure 6A, the operator must only adjust the grind setting to find EYmax
(shown in green), followed by reducing the volume of water used in the extraction
to result in the same EY as measured using the refractive index device (shown in
red). This process results in reduction in the total cup volume but does increase
the cup concentration and reproducibility.
Matter 2, 1–18, March 4, 2020 13

A
Improved espresso reproducibility Locate maximum extraction for a Decrease shot volume (i.e. ratio)
Increased beverage concentration fixed brew ratio and water pressure to obtain similar tasty point EY
maximum extraction maximum extraction
Extraction yield
Extraction yield
Barista determined
tasty point
Iterate between both regimes

Extraction yield
Extraction yield Grind setting Grind setting
Extraction yield
15 g in / 40 g out 15 g in / 40 g out
Grind setting
20 g in / 40 g out
Tasty point Optimized tasty point 20 g in / 40 g out
B
Decreased beverage concentration Reduce dry coffee mass Arrive at similar tasty point EY
Reduction in coffee dry mass used in extraction by grinding much finer
Figure 6. Schematic Illustrating Two Strategies to Improve Espresso Reproducibility

After determining a tasty point (yellow), the barista can obtain the same EY by (A) grinding coarser to find the maximum EY (green) and then reducing
the water mass, or (B) downdosing and grinding much coarser. The former results in smaller beverages with higher coffee concentration, and the
latter results in less dry mass coffee being used at a lower concentration. Here ‘‘in’’ and ‘‘out’’ refer to dry coffee mass and beverage volume,
respectively.
Systematic Reduction of Coffee Mass by Downdosing and Grinding Coarse

As we demonstrated in Figure 3, our model informs us that a reduction in dry
coffee mass results in an increased EYmax (shown schematically in blue in Figure 6).
Thus, a barista is able to achieve highly reproducible espresso with the same EY
as the 20 g espresso by reducing the coffee mass to 15 g and counter-intuitively
grinding much coarser (as shown in red, Figure 6B). This modification may result in
very fast shots (<15 s), a reduction in espresso concentration, and a different flavor
profile.
The Specialty Coffee Association espresso parameters mandate that the extraction
should take 20–30 s; we speculate that this might be partially responsible for the pre-
vailing empirical truth that most coffee is brewed using grind settings that cause
partially clogged/inhomogeneous flow. Remembering that the initial tasty point
may lie in the clogged flow regime, some of the bed is extracted much more than
the refractive index measurement suggests. By lowering the dry coffee mass and
grinding to maximize EY, the operator may notice that they are able to push their ex-
tractions much higher than before, while achieving highly reproducible espresso.
Indeed, the two approaches presented in Figure 6 are complimentary, because
the former increases the shot concentration and the latter decreases the dry mass.
There are circumstances where businesses make decisions on the minimum concen-
tration and volume of espresso that is acceptable to present to customers. When
iterated, these approaches result in optimization of beverage volume, concentra-
tion, and other economic implications.
14 Matter 2, 1–18, March 4, 2020

25
combined espresso shots to obtain
complexity and high reproducibility
24 expected EY 15 g in / 40 g out
high EY, low TDS
23
partially clogged
flow tasty point
20 g in / 40 g out
22 low EY, high TDS
21
+ to obtain more consistently
1.1 1.3 1.5 1.7 1.9 2.1 2.3

Grind setting (GS)
Figure 7. A Procedure to Approximate Flavors Obtained in the Partially Clogged Flow Regime
Blending two espresso shots provides an avenue to obtain the flavor complexity of the partially
clogged flow tasty point (yellow) with less shot variation. One high dry mass, low EY shot
(purple) combined with one low dry mass, high EY shot (green) provides an approximation to the
tasty point.
Blending Shots
Beyond sensory science studies, a persistent difficulty is that there is no rapid route
to assessing the quality of two identical EYs made with different grind settings or
brew parameters. It is clear that espresso made at 22% EY in the partially clogged
regime tastes more ‘‘complex’’ than a fast 22% EY obtained using the optimization
routine presented in Figure 6. In an attempt to recover the same flavor profile as
the partially clogged flow regime, a shot must contain a mixture of higher and lower
extractions. Consider the tasty point in Figure 7: One can approximate its flavor pro-
file by blending two shots: (1) a low extraction/high dose (purple point) and (2) a high
extraction/low dose (green point). This procedure can more economically yield a
shot with a flavor profile that should approximate that which was previously only
obtainable in an economically inefficient partially clogged shot. Blending shots
does double the total volume of the beverage, and the procedure comes with the
added combinatorial complexity associated with calibrating two shots that, when
mixed together, yield superior flavor. We expect only the most enthusiastic practi-
tioners would consider this approach, but it may well be actionable in an industrial
setting where extraction is carried out in bulk.
IMPLEMENTATION AND OUTLOOK

In 2017, we implemented the waste reduction protocol into a local specialty cafe in
Eugene, Oregon. Examination of their sales data between September 2018 and
September 2019 revealed that the cafe produced 27,850 espresso-containing bev-
erages. Previously, each beverage would have contained 20 g of coffee dry mass.
This specialty-grade coffee is valued at $0.53 per 20 g. By systematically reducing
this mass by 25%, the cafe was able to save $0.13 per drink, amounting to a revenue
Matter 2, 1–18, March 4, 2020 15

increase of $3,620 per year. In addition to the monetary saving, the shot times were
routinely reduced to 14 s, significantly reducing the order-to-delivery time. From this
proof of concept, we can speculate on the larger economic benefit gained by the
procedure detailed herein. Encompassing both specialty- and commodity-grade
coffee beverage products, we estimate that the average coffee beverage is pro-
duced using ca. $0.10 of coffee. Considering a 25% reduction in coffee mass (i.e.,
$0.025 saving), and considering the daily coffee consumption in the United States
(124,000,000 espresso-based beverages per day),21 our protocol yields $3.1 million
savings per day, or $1.1 billion per year. Of course, this poses significant problems
for the entire supply chain, because being more efficient with ground coffee does
yield less revenue for roasters, importers, and producers and further highlights
that there is still much work to be done to improve efficiency in the industry, while
also uniformly increasing profits.
While we do not present solutions to all of these interesting problems, we have

described the formulation of a novel model for extraction (i.e., mass transfer) from
a granular bed composed of mixed particle sizes to a liquid that flows through this
bed. Multiple-scale homogenization, which exploited the disparity between the
length scales associated with individual grains and those of the whole bed, has
been used to reduce the model from its original intricate geometry to a multi-scale
model that has markedly less geometric complexity and is therefore usable and
more readily diversified. The model is able to faithfully reproduce experimental
measurements in regimes where flow is homogeneous and accurately predicts EYs
as a function of meaningful parameters, such as coffee mass, water mass, and extrac-
tion rates. While the EY is not directly indicative of quality, it does allow for economic
arguments to be made. Furthermore, such measurements, paired with model pre-
dictions, have led to novel insight, which suggests several strategies for systemati-
cally improving espresso reproducibility as well as reducing coffee waste, leading
to more sustainable production of high-quality beverages. Ultimately, we conclude
by presenting a route to obtain complex flavor profiles while maintaining economic
savings through blending of shots. While the latter is not necessarily practical, it
does highlight that partially clogged flow may impart complexity, albeit with large
variation.
EXPERIMENTAL PROCEDURES
Espresso was prepared using standard equipment at Frisky Goat Espresso. Twenty
gram ridge-less baskets were fitted into the porta-filters of a San Remo Opera three
group espresso machine. The Opera allows for precise control of shot time, water
pressure (PW), and temperature. Coffee was ground on a Mahlkönig EK 43 grinder
fitted with coffee burrs. Espresso is typically ground at a grind setting (GS, a.u.) =
1.3–2.3, depending on the coffee. Tamp force (tF) was controlled using the Barista
Technology BV Puqpress, an automated tamper accurate to within G3 N. We elec-
ted to use an espresso profile specialty coffee, roasted by Supreme Roasters (Bris-
bane, Australia). The mass of coffee in the basket and the mass of the outgoing liquid
coffee were measured on an Acaia Lunar espresso balance.
Although an exhaustive characterization of the chemistry in each shot is necessary for

the absolute description of shot composition (and therefore quality), we can use the
total extracted mass as a first approximation to gauge reproducibility. The concen-
tration of coffee is phenomenologically related to the refractive index of the
beverage (which is temperature dependent) and can be recovered using a previ-
ously presented methodology.19
16 Matter 2, 1–18, March 4, 2020

We elected to use a representative modern espresso recipe (i.e., 20.0(5) g of dry

ground coffee in, 40.0(5) g of beverage out). Temperature was kept constant at
92 C. Espresso shots were discarded if, due to human error, the shot mass was
outside a G1 g tolerance. Exact beverage masses were included in each calculation
of EY. Calibrating measurements were made using a larger sample size, n = 20, and
subsequent data were collected in pentaplicate.
Laser diffraction particle size analysis was performed on a Beckman Coulter LS13 320
MW. The instrument has a built-in dark-field reticule, which is used to ensure correct
optical alignment. An alignment check was carried out prior to every run to ensure
the optimum accuracy of the particle size distribution.
Nitrogen physisorption isotherm data were acquired at 196 C on a Quadrasorb SI

(Quantachrome Instruments). Prior to measurement, each ground coffee sample was
degassed twice at 200 C under vacuum for 16 h. As a consequence of the minimal
gas adsorption (due to low surface area and a lack of small pores) at partial pressures
<0.3, it was not possible to obtain a linear fit to the BET equation. Therefore, a spe-
cific BET surface area value is not reported.
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.matt.
2019.12.019.
ACKNOWLEDGMENTS
This work was enabled by charitable donations of equipment from Barista Technol-
ogy BV (Puqpress), Acaia Corp. (balance), Frisky Goat Espresso (personnel and cof-
fee), and our continued collaboration with Meritics Ltd (particle size analyses). The
authors are grateful to M. and L. Colonna-Dashwood, A. Thomas Murray, and R.
Woodcock for insightful discussions. We thank Tailored Coffee Roasters, Eugene,
OR, for their implementation of the procedures detailed herein, and B. Sung and
M. Pierson for providing their revenue data. This work used the Extreme Science
and Engineering Discovery Environment (XSEDE), which is supported by National
Science Foundation grant number ACI-1053575.
AUTHOR CONTRIBUTIONS
The study was conceived by C.H.H., M.I.C., J.M.F, and W.T.L. M.I.C. and D.M.
brewed the espresso. D.H. and E.U. performed the grinding experiments. Z.C.K
and S.A.F. performed the gas sorption measurements. J.W., W.T.L, and J.M.F.
formulated the mathematical model. J.M.F carried out the model analysis and wrote
the code to solve the reduced model. C.H.H., W.T.L., and J.M.F. wrote the manu-
script, and all authors contributed to the final version.
DECLARATION OF INTERESTS
The authors declare no competing financial interests.
Received: July 31, 2019

Revised: November 5, 2019
Accepted: December 16, 2019
Published: January 22, 2020
Matter 2, 1–18, March 4, 2020 17

REFERENCES
1. Sunarharum, W.B., Williams, D.J., and Smyth, S., Miller, B., Butler, K.T., Melot, B.C., Speirs, 27. Spiro, M., and Chong, Y.Y. (1997). The kinetics
H.E. (2014). Complexity of coffee flavor: a R.W., and Hendon, C.H. (2016). The effect of and mechanism of caffeine infusion from
compositional and sensory perspective. Food bean origin and temperature on grinding coffee: the temperature variation of the
Res. Int. 62, 315–325. roasted coffee. Sci. Rep. 6, 24483. hindrance factor. J. Sci. Food Agric. 74,
416–420.
2. Spencer, M., Sage, E., Velez, M., and Guinard, 14. Glöss, A.N., Schönbächler, B., Rast, M.,
J.-X. (2016). Using single free sorting and Deuber, L., and Yeretzian, C. (2014). Freshness 28. Corrochano, B.R., Melrose, J.R., Bentley, A.C.,
multivariate exploratory methods to design a indices of roasted coffee: monitoring the loss Fryer, P.J., and Bakalis, S. (2015). A new
new coffee taster’s flavor wheel: design of of freshness for single serve capsules and methodology to estimate the steady-state
coffee taster’s flavor wheel. J. Food Sci. 81, roasted whole beans in different packaging. permeability of roast and ground coffee in
S2997–S3005. Chimia (Aarau) 68, 179–182. packed beds. J. Food Eng. 150, 106–116.
3. Schenker, S., Heinemann, C., Huber, M., 15. Ross, C.F., Pecka, K., and Weller, K. (2006). 29. Coffee beans. www.optics.rochester.edu/
Pompizzi, R., Perren, R., and Escher, R. (2002). Effect of storage conditions on the sensory workgroups/cml/opt307/spr16/.
Impact of roasting conditions on the formation quality of ground Arabica coffee. J. Food Qual.
of aroma compounds in coffee beans. J. Food 29, 596–606. 30. Moroney, K.M. (2016). Heat and mass transfer
Sci. 67, 60–66. in dispersed two-phase flows, PhD thesis
16. Farah, A. (2012). Coffee constituents. In Coffee, (University of Limerick).
4. Baggenstoss, J., Poisson, L., Kaegi, R., Perren, Y.-F. Chu, ed. (John Wiley & Sons), pp. 21–58.
R., and Escher, F. (2008). Coffee roasting and 31. Fowler, A.C., and Scheu, B. (2016). A theoretical
aroma formation: application of different time- 17. Folmer, B. (2016). The Craft and Science of
Coffee (Academic Press). explanation of grain size distributions in
temperature conditions. J. Agric. Food Chem. explosive rock fragmentation. Proc. R. Soc. A
56, 5836–5846. 472, 20150843.
18. Tan, C.-Y., and Huang, Y.-X. (2015).
5. Rao, S. (2014). The Roaster’s Companion (Scott Dependence of refractive index on
Rao). concentration and temperature in electrolyte 32. Foster, J.M., Chapman, S.J., Richardson, G.,
solution, polar solution, nonpolar solution, and and Protas, B. (2017). A mathematical model for
6. Rao, S. (2008). The Professional Barista’s protein solution. J. Chem. Eng. Data 60, 2827– mechanically-induced deterioration of the
Handbook (Scott Rao). 2833. binder in lithium-ion electrodes. SIAM J. Appl.
Math. 77, 2172–2198.
7. Rao, S. (2010). Everything but Espresso: 19. V. Fedele, Universal refractometer apparatus
Professional Coffee Brewing Techniques (Scott and method, (2012), US Patent 8239144, filed 33. Richardson, G., Denuault, G., and Please, C.P.
Rao). Mar 31, 2010, and published Aug 7, 2012. (2012). Multiscale modelling and analysis of
lithium-ion battery charge and discharge.
8. National Coffee Association of USA. 20. Kanouté, P., Boso, D.P., Chaboche, J.L., and J. Eng. Math. 72, 41–72.
Understanding the economic impact of the Schrefler, B.A. (2009). Multiscale methods for
United States coffee industry. http://www. composites: a review. Arch. Comput. Methods 34. Foster, J.M., Gully, A., Liu, H., Krachkovskiy, S.,
ncausa.org/Industry-Resources/Economic- Eng. 16, 31–75. Wu, Y., Schougaard, S.B., Jiang, M., Goward,
Impact/Economic-Impact-Infographic. G., Botton, G.A., and Protas, B. (2015).
21. E-Imports. Coffee statistics. http://www.e- Homogenization study of the effects of cycling
9. Gay, C., Estrada, F., Conde, C., Eakin, H., and importz.com/coffee-statistics.php. on the electronic conductivity of commercial
Villers, L. (2006). Potential impacts of climate lithium-ion battery cathodes. J. Phys. Chem. C
change on agriculture: a case of study of coffee 22. Spiro, M., and Selwood, R.M. (1984). The 119, 12199–12208.
production in Veracruz, Mexico. Clim. Change kinetics and mechanism of caffeine infusion
79, 259–288. from coffee: the effect of particle size. J. Sci. 35. von Bruggeman, D.A.G. (1935). Berechnung
Food Agric. 35, 915–924. verschiedener physikalischer konstanten von
10. Jaramillo, J., Muchugu, E., Vega, F.E., Davis, A.,
23. Spiro, M., and Page, C.M. (1984). The kinetics heterogenen substanzen. i.
Borgemeister, C., and Chabi-Olaye, A. (2011).
and mechanism of caffeine infusion from dielektrizitätskonstanten und leitfähigkeiten
Some like it hot: the influence and implications
coffee: hydrodynamic aspects. J. Sci. Food der mischkörper aus isotropen substanzen.
of climate change on coffee berry borer
Agric. 35, 925–930. Ann. Phys. 416, 636–664.
(Hypothenemus hampei) and coffee
production in East Africa. PLoS One 6, e24528.
24. Spiro, M., and Hunter, J.E. (1985). The kinetics 36. Robinson, J.O. (1970). The misuse of taste
11. Davis, A.P., Gole, T.W., Baena, S., and Moat, J. and mechanism of caffeine infusion from names by untrained observers. Br. J. Psychiatry
(2012). The impact of climate change on coffee: the effect of roasting. J. Sci. Food Agric. 61, 375–378.
indigenous Arabica coffee (Coffea arabica): 36, 871–876.
predicting future trends and identifying 37. Cruz, A., and Green, B.G. (2000). Thermal
priorities. PLoS One 7, e47981. 25. Spiro, M., Toumi, R., and Kandiah, M. (1989). stimulation of taste. Nature 403, 889–892.
The kinetics and mechanism of caffeine
12. Bunn, C., Läderach, P., Rivera, O.O., and infusion from coffee: the hindrance factor in 38. Clará, R.A., Marigliano, A.C.G., and Sólimo,
Kirschke, D. (2015). A bitter cup: climate intra-bean diffusion. J. Sci. Food Agric. 46, H.N. (2009). Density, viscosity, and refractive
change profile of global production of Arabica 349–356. index in the range 283.15 to 353.15 k and vapor
and Robusta coffee. Clim. Change 129, 89–101. pressure of a-pinene, d-limonene, G-linalool,
26. Spiro, M. (1993). Modelling the aqueous and citral over the pressure range 1.0 kPa
13. Uman, E., Colonna-Dashwood, M., Colonna- extraction of soluble substances from ground atmospheric pressure. J. Chem. Eng. Data 54,
Dashwood, L., Perger, M., Klatt, C., Leighton, roast coffee. J. Sci. Food Agric. 61, 371–372. 1087–1090.
18 Matter 2, 1–18, March 4, 2020

Matter, Volume 2
Supplemental Information
Systematically Improving
Espresso: Insights from Mathematical
Modeling and Experiment
Michael I. Cameron, Dechen Morisco, Daniel Hofstetter, Erol Uman, Justin
Wilkinson, Zachary C. Kennedy, Sean A. Fontenot, William T. Lee, Christopher H.
Hendon, and Jamie M. Foster
Supplemental Information
Systematically improving espresso: insights from mathematical modeling and
experiment
Michael I. Cameron and Dechen Morisco
Frisky Goat Espresso, 171 George St., Brisbane City, QLD, 4000, AUS and
Current affiliation: ST. ALi Coffee Roasters, 12-18 Yarra Place, South Melbourne, VIC, 3205, AUS
Daniel Hofstetter
Daniel Hofstetter Performance, Laenggenstrasse 18, CH-8184 Bachenbuelach, CHE
Erol Uman
Meritics Ltd., 1 Kensworth Gate, Dunstable, LU6 3HS, UK
Justin Wilkinson
Faculty of Mathematics, University of Cambridge, Cambridge, CB3 0WA, UK
Zachary C. Kennedy
National Security Directorate, Pacific Northwest National Laboratory, Richland, WA, 99352, USA
Sean A. Fontenot
Department of Chemistry and Biochemistry, University of Oregon, Eugene, OR, 97403, USA.
William T. Lee
Department of Computer Science, University of Huddersfield, Huddersfield, HD1 3DH, UK and
MACSI, Department of Mathematics and Statistics, University of Limerick, Limerick, Ireland
Christopher H. Hendon
Materials Science Institute and Department of Chemistry and Biochemistry,
University of Oregon, Eugene, OR, 97403, USA
Jamie M. Foster
School of Mathematics & Physics, University of Portsmouth, Portsmouth, PO1 2UP, UK
(Dated: January 7, 2020)
I. SUPPLEMENTAL EXPERIMENTAL lutions into/out of the liquid arising from the boundary
PROCEDURE conditions (Equation 14) results in a reaction-advection-
diffusion equation. The calculation can be adapted
A. Multiple scales homogenization appropriately1 to show that
κ
In this section we carry out upscaling (or homogeniza- q = − ∇P, (1)
µ
tion) of the system of equations formulated in the Supple-
∂c∗
mental Information. Rather than resorting to a lengthy (1 − φs ) l
multiple scales analysis we use the previous results from ∂t (2)
the rigorous analysis1 to formulate the multi-scale sys- = ∇ · (Def f ∇c∗l − qc∗l ) + bet,1 G1 + bet,2 G2 ,
tem of equations. Before proceeding we note the abuse
of notation used in this section and henceforth; namely, where the star has been appended to cl to emphasize that
that the dependent variables here, and in the multi-scale it has undergone the homogenization procedure. Here, κ
homogenization (main text) are not strictly the same as is the permeability of the packed bed and q is the Darcy
those appearing in model development (main text) but flux (i.e., the discharge per unit area with units of m/s)
are instead their homogenized counterparts. Despite this, which is related to the average fluid velocity within the
in the interests of brevity we opt not to embed this dis- pore space, ν, via q = (1 − φs )ν. One might conjec-
tinction within the notation. ture that in espresso making applications pressures are
Upscaling the Navier-Stokes equations at suitably sufficiently high, and the pores sufficiently small, that
small Reynolds numbers (so that flow is laminar) on a turbulent flow could be present. However, the experi-
porous media results in Darcy’s law2,3 , whilst upscaling mental evidence strongly indicates that a model based
(Equation 1) and accounting for the source/sink of so- on Darcy’s law (rather than Ergun or Forchheimer4,5 )
2
is suitable6 . In Supplemental Equation 2, bet,i are the equation reduces to

Brunauer-Emmett-Teller surface areas (defined to be the
surface area, of grain species i, per unit volume of puck), κ ∂P
q=− , (4)
and Def f is the effective macroscopic diffusivity. The µ ∂z
quantity φs appearing in Supplemental Equation 2 is the where the scalar q is the z-component of q.
local volume fraction of coffee grains which, under our Since the primary focus of the model is to capture the
assumption that the grains are spheres, is related to the extraction process after the initial wetting phase we as-
BET surface areas and radii by sume that in the regime of interest the flow is steady.
Despite this simplification we will provide scope for the
bet,i 4 3
φs = φs,1 + φs,2 , where φs,i = πa . (3) model to capture consolidation of the bed that may occur
4πa2i 3 i in the wetting phase where the finer grains may be swept
towards the bottom of the basket by the intruding liquid
The latter equation is a product of: (i) the surface area of — in the coffee industry this process is known as fines
particles per volume of bed divided by the surface area migration. Importantly for our model such a phenomena
of a single particle—which is therefore the number of would lead to spatial variations in the bed permeability.
particles of a particular type per volume of bed, and; (ii) The solution to Supplemental Equation 4, with spatial
the volume of a single particle of a particular type. The dependence in κ, supplemented by the boundary condi-
product of these two factors is therefore the local volume tions in Equations 5 and 8 is
fraction of the bed of particles of a particular size, and the
1 L
Z
sum over the terms gives the local solid volume fraction κef f Ptot
q= , κef f = κ(s)ds. (5)
φs . µL L 0
The form of the boundary and initial conditions to
close Supplemental Equation 2 remain unchanged, and Inserting the result of Supplemental Equation 5 into Sup-
are Equations 7, 10, and 17, with the exception that u is plemental Equation 1 and reducing to dependence only
interchanged with q. An important consequence of up- on z gives equations the system Equation 18–Equation
scaling the Navier-Stokes equations to give Darcy’s law is 19 for c∗l . On assuming radial symmetry within each cof-
that the order of the system is reduced. We are therefore fee grain the resulting problems to be solved for cs1 and
not at liberty impose all the boundary conditions laid out cs2 are Equation 21–Equation 22. The problem is closed
in the main text. To close Supplmental Equation 1 we by supplying the initial conditions in Equation 17.
retain Equations 5 and 8 and the component of 9 normal
to the boundary, i.e., q· n̂|R=R0 = 0. The omission of the
remaining conditions is justified on the basis that, had we C. Tuning the model to espresso extraction data
carried out the homogenization explicitly, we would have
imposed it when solving on the microscopic length scale. Table S1 summarizes the parameter values harvested
The governing equations for the coffee concentration from the literature. The values shown in Table S2 were
in the solid phase remain unchanged by the homogeniza- derived by choosing a demarcation between fines and
tion procedure; csi are still to be determined by solving boulders of 100 µm and integrating the grind data shown
Equation 4 with their associated boundary and initial in Figure 2 leading to the volume contributions of the two
conditions, Equation 13 and the latter equation in Equa- families of particles. This was then converted into BET
tion 17. In the context of the multiple scales approach surface area contributions using the relationship in Sup-
one can think of the retention of the full system for csi plemental Equation 3. Intermediate grind settings were
as being necessitated by the need to evaluate the reac- linearly interpolated from these data.
tion rates, Gi , which can only be done by returning to
the “microscopic” scale. One must therefore solve two
equations of the form shown in Equation 4 at each sta- D. Numerical approach
tion in macroscopic coordinate system, one with i = 1
the other with i = 2, to predict the coffee concentration Here, an approach to numerically solve Equation 17–
profiles within a representative coffee grain of radius ai Equation 22 is described. The method centers around:
(for i = 1, 2). (i) finite difference approximations of spatial derivatives
in z; (ii) the use of a control volume method10 for treat-
ment of the spatial dependence in r; (iii) and, MATLAB’s
ODE suite for temporal integration. The code is hosted
B. One-dimensional reduction on a GitHub repository and is available.11
Finite differences are the method of choice for the
It follows from the form of the upscaled equations derivatives in z primarily for their ease of implementa-
and boundary conditions that the marcoscopic solution tion. The particular control volume method chosen is
should be one-dimensional, i.e., that it should depend particularly apt to treat the transport of coffee in the
only on z, depth through the basket. In this case Darcy’s solid phase for two reasons. Firstly, because it exhibits
3
Description Sym. Val. Ref.

Radius of bed R0 29.2 × 10−3 m —
Viscosity of water at 90◦ C µ 3.15 × 10−4 Pa·s 7
Density of water 90◦ C ρout 997 kg/m3 —

Roasted coffee bulk density ρgrounds 330 kg/m3 8
Saturation concentration of water csat 212.4 kg/m3 9
Initial concentration of solubles in grounds cs0 118.0 kg/m3 9

9
Volume fraction of grounds in packed bed φs 0.8272
Dose of grounds “in” Min 20 g —
Mass of beverage “out” Mout 40 g —
Pump overpressure Ptot 5 bar —
TABLE S1. Parameter values above the horizontal score were taken from the literature, and those below were chosen to mimic
the experimental extraction protocol used here.
Grind setting, Gs 1.0 1.5 2.0 2.5

Vol. fraction of fines, φs1 0.1689 0.1343 0.1200 0.0780
Vol. fraction of boulders, φs1 0.6583 0.6929 0.7072 0.7492
Radius of boulders, a2 (µm) 273.86 335.41 335.41 410.79
BET surface area of fines, bet1 (1/m) 10.187×103 8.099×103 7.239×103 4.703×103
BET surface area of boulders, bet2 (1/m) 7.211×103 6.197×103 6.325×103 5.472×103
TABLE S2. Parameter values above the horizontal score were extracted directly from the experimental data shown in Figure
2 and those below were subsequently inferred using Supplemental Equation 3. Throughout, we take the radius of fines to be
fixed with a1 = 12µm.
perfect conservation of coffee mass, which can be hard in r and at these locations we denote the value of the cof-
to ensure using standard finite differences owing to the fee concentration in small and large particles using the
singularity at the origin of the radial coordinate in the following shorthands cs1 (z, r, t)|z=zj ,r=rk = cs1,j,k (t) and
spherical diffusion equation. Secondly, in contrast to cs2 (z, r, t)|z=zj ,r=rk = cs2,j,k (t), respectively. Thus, the
many other control volume methods, it provides direct index j indicates the representative particle’s position in
access to the concentration on the surface of the parti- z whereas k labels the radial position within a particu-
cles thereby avoiding the need for extrapolation which lar representative particle. The (2M + 1) × N unknown
would inevitably introduce additional errors. This sur- functions of time were converted into one large column
face concentration determines the reaction rate across the vector u(t) as follows
solid grain boundaries and so accurate evaluation of this
quantity is crucial for reliable simulation. After applying u(t) = [cl (t)∗T cs1,1 (t)T cs1,2 (t)T
these treatments for the spatial dependencies, the system (6)
of PDEs (Equation 17–Equation 22) are reduced to sys- · · · cs1,N (t)T cs2,1 (t)T cs2,2 (t)T · · · cs2,N (t)T ]T .
tem of coupled ODEs. We select the MATLAB routine
ode15s to integrate this system forward in time because:
We now rewrite the problem in the form
(i) it is able to cope with solving a system of differential-
algebraic equations; (ii) it offers adaptive time-stepping,
and; (iii) has relatively modest computational cost. du
M = f (u), (7)
We introduce N equally spaced grid points, zj for dt
j ∈ [1, N ], thereby dividing the spatial z-domain into
N − 1 equally spaced subdomains. The grid spacing in where M is the mass matrix and f (u) is a nonlinear func-
z is therefore given by hz = 1/(N − 1). Henceforth we tion which arises from the application of finite difference
adopt the shorthand notation c∗l (z, t)|z=zj = c∗l,j (t). At approximations (to the equations in z) and the control
each station in z we must solve for the coffee concentra- volume method (for the equations in r). The system of
tion within a representative grain, i.e. , at each zj we ODEs (Equation 7) is written in the standard form ac-
must solve two equations of the form shown in Equa- cepted by MATLABs ode15s.
tion 21; one with i = 1 for cs1 and another with i = 2 Below we present the first N entries of the nonlinear
for cs2 . Each of the 2N copies of Equation 21 are dis- function f (u) arising from the discretisation of Equation
cretized by introducing M equally spaced grid points, rk 18 and its boundary conditions (Equation 19). The re-
for k ∈ [1, M ], which subdivide each r-domain into M −1 maining 2M N equations arising from applying control
subdomains. The grid spacing in r is therefore given by volumes to Equation 20 are previously detailed10 , and so
hr = 1/(M −1). In total we have N ×M different stations in the interests of brevity we do not repeat them here.
4
Grind setting, Gs 1.1 1.3 1.5 1.7 1.9 2.1 2.3

Shot time, tshot (s) 37.6 35.7 33.9 30.5 28.5 26.6 24.0
Darcy flux, q (m/s) 3.98×10−4 4.20×10−4 4.42×10−4 4.91×10−4 5.26×10−4 5.63×10−4 6.24×10−4
TABLE S3. Parameter values above the horizontal score are experimental values shown in Figure 4 whilst the values of the
Darcy flux below the horizontal score were computed via Equation 26.
Dose of grounds “in”, Min (g) 16 18 20 22 24

Bed depth, L (m) 15.0×10−3 16.8×10−3 18.7×10−3 20.6×10−3 22.5×10−3
TABLE S4. Parameter values that were adjusted to explore the effects of altering the dose of grounds “in”.
We have
M1,1 = 0, (8)

Def f 3 1
f1 = − − u1 + 2u2 − u3 + u1 , (9)
hz 2 2
Mi,i = 1 − φs , (10)
Def f
fi = (ui+1 − 2ui + ui−1 )
h2z
(11)
ui+1 − ui−1
− + bet,1 K,
2hz
MN,N = 0, (12)

Def f 1 3
fN =− uN −2 − 2uN −1 + uN . (13)
hz 2 2
5
Grind setting, Gs 1.1 1.3 1.5 1.7 1.9 2.1 2.3

Pump overpressure,
Ptot = 3 bar Darcy flux, q (m/s) 2.39×10−4 2.53×10−4 2.65×10−4 2.95×10−4 3.16×10−4 3.38×10−4 3.74×10−4
Pump overpressure,
Pump overpressure,
Pump overpressure,
TABLE S5. Parameter values that were adjusted to explore the role of altering the pump overpressure. The values of the
Darcy flux for a 5 bar overpressure are identical to those on Table S3 and the others have been computed from those via the
relationship in Equation 26.
6
E. Gas uptake and experimental tamp force
The BET sorption data and effect of tamp force are

presented.
20
15
N2 adsorbed (cm3/g)
10
5 desorption
adsorption
0
0.0 0.2 0.4 0.6 0.8 1.0
Partial pressure (P/P0)
FIG. S1. Brunauer-Emmett-Teller gas sorption measurement

of coffee ground at 2.5.
55
50
9 bar, G
S = 1.3
45
Shot time (s)
40
6 bar, G =
S 1.3
35
6 bar, GS = 1.5
30 9 bar, GS = 1.
5
25
100 200 300
Tamp force, tF (N)
FIG. S2. Monitoring shot time as a function of tamp force and

static water pressure. Fine grind settings and high pressure
resulted in choking the espresso machine.
7
1
G. Richardson, G. Denuault, and C. P. Please (2012), the steady-state permeability of roast and ground coffee in
“Multiscale modelling and analysis of lithium-ion battery packed beds,” J. Food Eng. 150, 106 – 116.
7
charge and discharge,” J. Eng. Math. 72, 41–72. J. Kestin, M. Sokolov, and W. A. Wakeham (1978), “Vis-
2
G. Allaire (1989), “Homogenization of the stokes flow in a cosity of liquid water in the range - 8 c to 150 c,” J. Phys.
connected porous medium,” Asymp. Anal. 2, 203–222. Chem. Ref. Data 7, 941–948.
3 8
D. Polisevsky (1986), “On the homogenization of fluid A. S. Franca, L. S. Oliveira, J. C. F. Mendonça, and X. A.
flows through porous media,” Rend. Sem. Mat. Univ. Po- Silva (2005), “Physical and chemical attributes of defective
litecn. Torino 44, 383–393. crude and roasted coffee beans,” Food Chem. 90, 89–94.
4 9
N. Dukhan, Ö. Bağcı, and M. Özdemir (2014), “Experi- K. M. Moroney, Heat and mass transfer in dispersed two-
mental flow in various porous media and reconciliation of phase flows, Ph.D. thesis, University of Limerick (2016).
10
forchheimer and ergun relations,” Exp. Therm. Fluid Sci. Y. Zeng, P. Albertus, R. Klein, N. Chaturvedi, A. Ko-
57, 425–433. jic, M. Z. Bazant, and J. Christensen (2013), “Efficient
5
A. C. Fowler, B. Scheu, W. T. Lee, and M. J. McGuinness conservative numerical schemes for 1d nonlinear spherical
(2010), “A theoretical model of the explosive fragmenta- diffusion equations with applications in battery modeling,”
tion of vesicular magma,” Proc. Royal Soc. Lon. A 466, J. Electrochem. Soc. 160, A1565–A1571.
11
731–752. J. M. Foster, “Espresso simulation code,” https://
6
B. R. Corrochano, J. R. Melrose, A. C. Bentley, P. J. Fryer, github.com/jamiemfoster/Espresso (2019), accessed 07
and S. Bakalis (2015), “A new methodology to estimate December 2019.

mmc2 PDF

Uploaded by

Copyright:

Available Formats

mmc2 PDF

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

mmc2 PDF

Uploaded by

Copyright:

Available Formats

Article

Systematically Improving Espresso: Insights

Implementation to yield monetary

Coffee extraction performed in an espresso machine is a process that depends on

Cameron et al., Matter 2, 1–18

SUMMARY Progress and Potential

Matter 2, 1–18, March 4, 2020 ª 2019 Elsevier Inc. 1

In principle, it is preferable to make objective statements about the flavor of food-

2 Matter 2, 1–18, March 4, 2020

Pore space / water Pump Coffee grains

Figure 1. A Schematic of the Espresso Basket Geometry

Specialty Coffee Association recommendations, the model suggests that we should

DEVELOPMENT OF A RATIONAL MODEL FOR ESPRESSO EXTRACTION

Matter 2, 1–18, March 4, 2020 3

Boundary conditions at the inlet, z = 0, include a specified fluid overpressure, the

Pjz = 0 = Ptot ; (Equation 5)

ð Dl Vcl + ucl Þ,b

ð Dl Vcl + ucl Þ,b

ujR = R0 = 0: (Equation 12)

4 Matter 2, 1–18, March 4, 2020

u = 0 on Gint ; (Equation 15)

G = kcs ðcs cl Þðcsat cl Þ on Gint ; (Equation 16)

cl jt = 0 = 0; cs jt = 0 = cs0 (Equation 17)

Particle-Size Distribution of Ground Coffee

Matter 2, 1–18, March 4, 2020 5

Multi-scale Homogenization and One-Dimensional Reduction

6 Matter 2, 1–18, March 4, 2020

The microscopic equations to be solved are

Matter 2, 1–18, March 4, 2020 7

The problem is closed by supplying the initial conditions Equation 17 and

Tuning the Model to Espresso Extraction Data

8 Matter 2, 1–18, March 4, 2020

Matter 2, 1–18, March 4, 2020 9

Regime 1: Standard flow

Figure 4. Espresso Extraction Yield as a Function of Grind Setting

Three parameters, namely Deff , Ds , and k, remain to be specified. The

10 Matter 2, 1–18, March 4, 2020

The Effect of Altering the Brew Ratio and Water Pressure

Examination of the extracted mass of coffee as a function of grind setting reveals a

Matter 2, 1–18, March 4, 2020 11

Partially clogged flow

1.1 1.3 1.5 1.7 1.9 2.1 2.3

The non-monotonic trend in EY with GS can be attributed to two competing extrac-

ACCOUNTING FOR THE PARTIALLY CLOGGED FLOW REGIME

12 Matter 2, 1–18, March 4, 2020

SYSTEMATICALLY IMPROVING ESPRESSO

Maximizing Extraction Yield by Altering GS

Matter 2, 1–18, March 4, 2020 13

maximum extraction maximum extraction

Iterate between both regimes

Extraction yield Grind setting Grind setting

Figure 6. Schematic Illustrating Two Strategies to Improve Espresso Reproducibility

Systematic Reduction of Coffee Mass by Downdosing and Grinding Coarse

14 Matter 2, 1–18, March 4, 2020

high EY, low TDS

1.1 1.3 1.5 1.7 1.9 2.1 2.3

IMPLEMENTATION AND OUTLOOK

Matter 2, 1–18, March 4, 2020 15