Article

Braess’s paradox and programmable

behaviour in microfluidic networks

https://doi.org/10.1038/s41586-019-1701-6 Daniel J. Case1, Yifan Liu2, István Z. Kiss2, Jean-Régis Angilella3 & Adilson E. Motter1,4*

Received: 4 September 2018

Accepted: 1 August 2019 Microfluidic systems are now being designed with precision as miniaturized fluid
Published online: 23 October 2019 manipulation devices that can execute increasingly complex tasks. However, their
operation often requires numerous external control devices owing to the typically
linear nature of microscale flows, which has hampered the development of integrated
control mechanisms. Here we address this difficulty by designing microfluidic
networks that exhibit a nonlinear relation between the applied pressure and the flow
rate, which can be harnessed to switch the direction of internal flows solely by
manipulating the input and/or output pressures. We show that these
networks— implemented using rigid polymer channels carrying water—exhibit an
experimentally supported fluid analogue of Braess’s paradox, in which closing an
intermediate channel results in a higher, rather than lower, total flow rate. The
harnessed behaviour is scalable and can be used to implement flow routing with
multiple switches. These findings have the potential to advance the development of
built-in control mechanisms in microfluidic networks, thereby facilitating the creation
of portable systems and enabling novel applications in areas ranging from wearable
healthcare technologies to deployable space systems.

Fulfilment of the promise of microfluidics to operate as autonomous In this Article, we explore new physics that emerges by combining
microscale networks in which fluids can be transported, mixed, reacted, network theory and fluid mechanics to induce nonlinear behaviour
separated and processed is no longer limited by experimental fabrica- in microfluidics and effectively create a passive two-terminal flow-
tion challenges, but rather by difficulties in creating built-in controls1–3. switch device that is entirely operated on-chip, directly by the work-
The importance of this limitation can be appreciated by noting that the ing fluid. Previous work that has achieved built-in control capabilities
development of the modern microelectronics that form the basis of (often externally actuated), including oscillatory flows15–18 and flow rate
computer microprocessors was ultimately determined by the creation regulation19,20, generally relied on flexible membranes and surfaces.
of integrated circuits, with all components fabricated on the same sub- Microfluidics with such flexible components require flows with very
strate. Microfluidics have already reached a level of integration in which low Reynolds numbers—a regime in which fluid inertia, and thus the
networks with thousands of components, including control devices, only nonlinear term of the Navier–Stokes equations for incompressible
are built on a single compact chip. However, in contrast with electronic fluids, becomes negligible. This has led researchers to often discount the
integrated circuits, existing on-chip fluid control devices still need to potential effects of fluid inertia on the flows (as reviewed, for example,
be actuated externally. For example, microfluidic circuits fabricated in refs. 21,22). Recent work has shown, however, that inertial forces can
from flexible polydimethylsiloxane (PDMS) can now incorporate a large serve as a powerful on-chip tool to manipulate microfluidic dynam-
number of control valves, which nevertheless have to be operated using ics locally23,24, including shaping streamlines25,26, mixing fluids27 and
control fluids through a control layer that lies on top of the working fluid directing particles28,29. Here, we present networks designed to amplify
network4,5. As a result, microfluidics are still predominantly controlled inertial effects by incorporating properties of porous media that can be
by external hardware, despite substantial efforts over the past 20 years used for non-local fluid routing and manipulation of output patterns.
to develop systems with new control schemes6–10. The construction of Figure 1a shows a schematic representation of a microfluidic system
systems that forgo the current reliance on external hardware is crucial to with the fundamental network structure we consider. It consists of
further the development of portable microfluidic systems for pressing five segments arranged as two parallel channels connected by a link-
applications, ranging from point-of-care diagnostics and health monitor- ing channel, where the inlets are kept at a common pressure Pin and
ing wearables to analysis kits for field research11–14. This requires devel- the outlets are held at a common, lower, pressure, Pout. One of the
oping next-generation integrated circuits in which not only the control outlet channels is modified to generate a nonlinear pressure–flow
devices but also the operation of those devices is integrated on-chip. The relationship, which is achieved by introducing an array of cylindrical
development of such a level of integration has been fundamentally lim- obstacles. Our principal results are supported by theory, simulations
ited by the fact that, at the microscale, fluid flows tend to respond linearly and experiments, and they show that we can: (i) induce a flow direc-
to pressure changes and thus cannot be easily amplified or switched. tion switch through the linking channel solely by varying the pressure

Department of Physics and Astronomy, Northwestern University, Evanston, IL, USA. 2Department of Chemistry, Saint Louis University, St Louis, MO, USA. 3Normandie Université, UNICAEN,
1

UNIROUEN, ABTE, Caen, France. 4Northwestern Institute on Complex Systems, Northwestern University, Evanston, IL, USA. *e-mail: [email protected]

Nature | Vol 574 | 31 October 2019 | 647

Article
a Simulation Velocity
L2 L4 Low High Re = 1
a
Pin Q2 P2 Q4 Pout
High Linking Low
L3 Q3
pressure L1 channel L5 pressure b Re = 220

Pin Q1 P1 Q5 Pout

Direction of flow
b Segment with Experiment
obstacles c Small recirculation d Large recirculation
regions regions
Re = 9 Re = 143

Pin 100 μm 100 μm

Pout e f
20 obstacles No obstacles
775 200

–ΔP/Re (Pa)
675 100

575 0
40 60 80 100 120 140 40 60 80 100 120 140
Fig. 1 | System schematics. a, Microfluidic network consisting of two parallel Reynolds number, Re Reynolds number, Re
channels, joined by a linking channel, that connect high- and low-pressure fluid Fig. 2 | Development of nonlinear flow. a, b, Simulated flow in a channel with
reservoirs. Grey filled circles represent stationary cylindrical obstacles. The obstacles (open circles), showing no recirculation for low Re (a) and noticeable
labels denote pressures (P), channel lengths (L) and flow rates (Q), with arrows recirculation near the obstacles for larger Re (b). c, d, Experimentally observed
indicating the positive flow direction. b, Generic multiswitch microfluidic flows around the obstacles (filled circles), visualized using pictures of
network consisting of an array of parallel channels interconnected by multiple fluorescent particles (marked in pink). The particle tracks trace the underlying
linking channels. A subset of channel segments contain cylindrical obstacles. flow structure, confirming the development of recirculation regions (white
Flow is driven through the network by a single pressure difference (Pin − Pout). areas) as Re is increased from low (c) to moderate values (d). e, f, Experimentally
measured relation between pressure loss and Re for a channel with (e; red curve)
and without (f; blue curve) obstacles. The dashed line in e is a reference to guide
difference between the inlets and outlets; and (ii) identify a pressure the eye and indicates an approximately quadratic relation between pressure loss
difference above which the total flow rate between the inlets and out- and flow rate.
lets increases on closing the linking channel. We also predict negative
conductance transitions when the linking channel is equipped with
an offset fluidic diode, which are transitions associated with non-
monotonic pressure–flow relations analogous to those previously where μ is the dynamic viscosity of the fluid. To induce deviations from
realized using flexible diaphragm valves30. The counter-intuitive this linear regime, we consider the effect of introducing multiple sta-
behaviour described in (ii) is formally equivalent to the so-called tionary obstacles in the channel. Figure 2a, b shows simulations of the
Braess’s paradox originally established for traffic networks31,32, where Navier–Stokes equations for a channel with ten cylindrical obstacles
closing a shortcut road has the possible effect of increasing net traffic of radius r = w/5 (Methods). We observe recirculation regions forming
flow. We demonstrate integration of the flow switch described in (i) near the obstacles for sufficiently large Reynolds number Re ≡ 2ρQ/μ,
by considering larger microfluidic networks, as illustrated in Fig. 1b, where ρ is the fluid density. The recirculation regions first appear for
which incorporate multiple linking channels and are thus capable Re of the order of 10, and their number and size depend on Re. These
of exhibiting multiple flow switches. Flows through these networks localized structures are hallmarks of fluid inertia effects (and thereby
are driven by a single pressure difference and yet can be designed to of nonlinearity). We investigate how fluid inertia effects compound to
exhibit a variety of flow states by programming the pressure at which impact the total flow rate by performing simulations across moderate
each flow switch occurs. values of Re when different numbers of obstacles are present. We find
that a nonlinear relation between the pressure drop ΔP and flow rate
System design and nonlinearity Q = μRe/2ρ emerges as soon as obstacles are introduced, and that the
We consider conditions under which all channel segments have the same nonlinearity becomes more pronounced as the number of obstacles is
width w, the working fluid is water, and all surfaces (including obstacles) increased (Supplementary Information section S3.1 and Supplementary
have no-slip boundaries. We assume, without loss of generality, that the Fig. 3).
pressure Pout at the outlets is zero, and consider scenarios in which either The nonlinearity we observe in the relation between ΔP and Q con-
the static or the total pressure is controlled at the inlets (Methods). We forms to the Forchheimer effect in porous media, which characterizes
examine two network configurations of the system in Fig. 1a: the con- flow through many interconnected microchannels when local inertial
nected configuration, in which the two parallel channels are allowed effects at the points of interconnection are non-negligible, even for
to exchange fluid through the linking channel; and the disconnected laminar flow33–35. We use the Forchheimer equation to derive a relation
configuration, in which the linking channel is closed or removed. In between ΔP and Re for the channel with obstacles, given by
our theoretical analysis and simulations, the flows are assumed to be
two-dimensional, yet the main results carry over to three dimensions, αμ 2 L βμ 2 L 2
−ΔP = Re + Re (2)
as verified in our experiments. 2ρw 4ρw 2
For a straight microfluidic channel of length L ≫ w without obstacles,
an approximate steady-state solution of the Navier–Stokes equations where α is the reciprocal permeability and β is the non-Darcy flow coef-
in two dimensions yields a linear relation between the total volumetric ficient, both depending solely on the system geometry (Methods).
flow rate per unit depth Q and the pressure drop ΔP along the channel: The physical mechanism giving rise to this nonlinearity is the increase
12μL in flow recirculation and velocity gradients for larger Re, as evidenced
−ΔP = Q (1)
w3 in Fig. 2a, b for Re = 1 and 220. To test the impact of the inertial effects

648 | Nature | Vol 574 | 31 October 2019

 6 difference in the total flow rate exists primarily in the difference in Q5,
Difference in total flow rate, ΔQ and Q3 acts as a controlling variable of Q5.
Percentage of connected system flow rate

4 Link flow rate, Q3 The observation of a lower total flow rate for the connected configu-
Difference in Q4 ration compared to the disconnected configuration for fixed Pin is a
2
manifestation of a fluid analogue of Braess’s paradox. Indeed, if we
0 consider the disconnected system driven by an inlet pressure Pin > Pin⁎ ,
the addition of the linking channel can result in a decrease in the total
–2
steady-state flow rate (as large as 10% in our simulations). The value of
–4 the critical pressure Pin⁎ depends, of course, on the dimensions of the
channels, but we find that the onset of Braess’s paradox and the flow
–6 switch always occur at the same pressure for the range of parameters
investigated. We obtain similar results for Braess’s paradox and flow
–8 switching when instead the total pressure is controlled at the inlets
Pin* (Supplementary Information section S3.4). Our observation of Braess’s
–10
0 500 1,000 1,500 2,000 paradox and flow switching also has the potential to lead to additional
Inlet static pressure, Pin (Pa) control features when existing microfluidic components are integrated
into our system. For example, by incorporating an offset fluidic diode36
Fig. 3 | Braess’s paradox and flow switching. Simulation results for the in the linking channel, the system can undergo negative (and positive)
connected and disconnected configurations of the system for a range of inlet conductance transitions, where an increase in Pin leads to an abrupt
pressures Pin. The flow rates are presented as a percentage of the total flow rate decrease in the total flow rate (Supplementary Information section S4).
through the connected system, QC, where we adopt the sign convention for the
flow directions as defined in Fig. 1a. The flow through the linking channel Experimental results
switches direction at the critical pressure Pin = Pin⁎ , which coincides with the onset
We performed experiments to validate our predictions of flow switch-
of negative ΔQ that marks the occurrence of Braess’s paradox.
ing and Braess’s paradox in a network with dimensions typical of micro-
fluidics. A schematic of the experimental apparatus is presented in
Fig. 4a, where an open/close valve is used to implement the addition/
in realistic systems, we perform experiments using microchannels removal of the linking channel (Methods). With the valve open, a flow
fabricated from stiff PDMS (hardened by curing). Figure 2c, d shows switch is observed at a critical driving pressure Pin⁎ in the range 5–10 kPa,
experimental evidence of the increase in the number and size of as demonstrated in Fig. 4a by images of the flows through the channel
the recirculation regions with Re, in agreement with our simula- junctions at the end points of this pressure range. (The switching behav-
tions. An approximately linear relation between −ΔP/Re and Re (and iour has no reliance on the valve, as explicitly shown in Supplementary
thus an approximately quadratic relation between −ΔP and Q) for Fig. 11.)
a channel containing 20 obstacles is shown in Fig. 2e, which contrasts A confirmation of Braess’s paradox in this system is shown in Fig. 4b
with the constant relation measured for a channel without obstacles for driving pressures above Pin⁎ , as observed in our simulations. The
in Fig. 2f. measured total flow rate is higher when the linking channel valve is
closed than when it is open, thus demonstrating the paradox, and the
Switching and Braess’s paradox magnitude of the paradox is observed to be larger for higher driving
We incorporate the channel segment with obstacles characterized above pressures. A breakdown of how the flow rate changes in channel seg-
into a network by considering the microfluidic system presented in Fig. 1a. ments 4 and 5 individually is shown in Fig. 4c, d. Closing the valve causes
We take the common static pressure Pin at the inlets to be the controlled the flow rates through both channels to increase, which is in agreement
variable in the system. The total flow rate through the network is now with direct simulations and is yet another striking aspect of Braess’s
simply the sum of the flows at the outlets, (Q4 + Q5). In Fig. 3, we present paradox in this system; it would be, at first, intuitive to expect that Q5
results for this system from direct simulations of the steady-state solu- would decrease when the in-flow from the linking channel is switched
tions of the Navier–Stokes equations. As Pin is increased from zero, the off. Time series of the flow rates measured as the linking channel is
flow rate through the linking channel Q3 is initially positive before chang- sequentially opened and closed further illustrate the transitions under-
ing direction and becoming negative once a critical pressure, defined as lying the paradox (as shown in Supplementary Fig. 12).
Pin⁎, is reached (Fig. 3). This flow switch results from the nonlinear change In our experiments, the total pressure is controlled at the inlets and
in pressure loss along the channel segment containing obstacles, which the experimental results are in full qualitative agreement with simula-
causes a switch in the sign of the pressure difference along the linking tions performed under the same pressure boundary conditions (Sup-
channel ΔP21 (approximately P2 − P1) as the flow rate through the system plementary Information section S3.4). This illustrates the robustness of
increases with Pin. We define QC to be the total flow rate for the connected the phenomenon, given that our simulations are in two dimensions and
system configuration and QD to be the total flow rate for the disconnected three-dimensional effects are expected to be present in the experiments.
system configuration, where both are regarded as functions of Pin. We note that different aspects of the paradox have been considered in
Figure 3 shows ΔQ ≡ Q C − QD for a range of applied pressures Pin. fluid networks, but only for macroscopic (that is, non-microfluidic)
Intuition may suggest that ΔQ is positive for all values of Pin because the systems and while modelled by ad hoc flow equations37–39. Analogues
linking channel in the disconnected system can be considered to have of the paradox have also been studied in several other areas, includ-
an infinite fluidic resistance, while for the connected system configura- ing electrical, mechanical, biological, and contemporary traffic net-
tion the resistance of the linking channel is finite. Hence, reducing the works40–44. These examples show that Braess’s paradox is a potentially
resistance of any component of the system may seem to imply that the general network phenomenon, which has remained unexplored in
total flow rate should increase for fixed Pin. We observe, however, that microfluidic networks.
ΔQ becomes negative for Pin above the critical pressure that marks the
flow switch, Pin⁎ , meaning that an open linking channel between the par- Network model
allel channels results in a lower total flow rate. Figure 3 also shows that To characterize the microfluidic system in Fig. 1a, we construct an ana-
the flow rate through the channel segment with obstacles, Q4, remains lytic model that captures the flow properties observed in our simula-
largely unchanged between the two configurations. Therefore, the tions and experiments. The model consists of pressure–flow relations

Nature | Vol 574 | 31 October 2019 | 649

Article
a Pin = 5 kPa Pin = 10 kPa b Pin = 20 kPa Pin = 100 kPa
2,125 7,050

Q4 + Q5 (μl min–1)
Channel 5 flow rate, Channel 4 flow rate, Total flow rate,
2,105 6,900
200 μm 200 μm

Q2 Q4 2,085 6,750
Open Closed Open Closed
Channel 4 c Open valve
495 1,425
Open/close

Q4 (μl min–1)
valve
Pressure 485 1,325
pump Q3
Closed valve
Q1 Q5 475 1,225
Open Closed Open Closed
Channel 5 d 1,630 5,650
200 μm 200 μm

Q5 (μl min–1)
1,620 5,550

1,610 5,450
Pin = 5 kPa Pin = 10 kPa Open Closed Open Closed

Fig. 4 | Experimental observation of flow switch and Braess’s paradox. right insets) the flow switching pressure Pin⁎ , where the flow directions are
a, Experimental setup of the system presented in Fig. 1a, with flow tracking indicated by the arrows. b, Total flow rate (Q4 + Q5) when the linking channel valve
images (insets) at the junctions. An air-pressure pump is used to equally is ‘open’ or ‘closed’ (see diagrams at right) for two different driving pressures
pressurize two vials containing red and blue dyed water, where each vial is above Pin⁎ . c, d, Breakdown of the total flow rate into Q4 (c) and Q5 (d) for the two
connected to one of the system inlets. The linking channel is equipped with an states of the valve. The plotted flow rates are averages derived from time series
open/close valve and channel 4 contains 20 obstacles. Images of the dyed flows data, and the error bars indicate one standard deviation. The observed increase in
through the junctions are shown for Pin below (5 kPa; left insets) and above (10 kPa; the total flow rate when the valve is closed is direct evidence of Braess’s paradox.

for each channel segment and, crucially, includes the most dominant where a and c are positive parameters and prime denotes derivative. If
term resulting from minor pressure losses at the channel junctions45,46 total pressure is controlled and dynamic pressure terms are included,
(Methods). We model the contribution of the latter as an additive term the paradox is also predicted for δPin > 0 provided that a relation similar
K(Q3/Q1)f(Q5) to the pressure–flow equation for channel segment 5, to equation (4) is satisfied (details for both cases are presented in Sup-
where the scaling factor f and the coefficient K are increasing func- plementary Information section S2). The dependence of condition (4)
tions for Pin ≥ 0 such that f(0) = K(0) = 0. Several results are obtained on β and K′(0) underlines the crucial roles of nonlinearity and minor
from this model for Pin > 0, as assumed throughout. First, if β = 0 (that losses in giving rise to Braess’s paradox in our experiments, and shows
is, the quadratic term is zero in equation (2)) when the static pressure in particular that minor losses have to be sufficiently large. Indeed, if the
is controlled or the dynamic pressure is negligible, then flow switching effect of minor losses is neglected, a manifestation of Braess’s paradox
does not occur, in agreement with direct simulations (Supplementary is still predicted to occur, but with much smaller magnitude and only
Information section S3.2). Second, when β > 0, a steady-state solution for δPin < 0, which is inconsistent with our simulations and experiments
can be found satisfying Q3 = 0 provided that the following geometric (Supplementary Information section S2.3).
condition is satisfied: The result in equation (4) also highlights a fundamental difference
between microfluidic and electronic circuits, namely that minor losses
12L2 L5
L1 < = L⁎ (3) (that is, energy losses associated with interactions between circuit com-
αw 2L4 ponents) do not have direct analogues in common electronics. Given
the central role played by such losses in equation (4), we posit that this
This solution identifies the critical pressure Pin⁎ . Third, for flow rates in difference might be the reason why no equivalent of the Braess paradox
the linking channel, the model predicts that a variation δQ3 is negatively effect we present has been observed in electrical networks, even though
related to a variation δPin around Pin⁎. This indicates that Pin above (below) aspects of it have40. We further investigated the impact of interactions
Pin⁎ results in a negative (positive) flow rate through the linking channel. between channel segments by varying the junction angles to show that
The first result implies that, in our experiments, the Forchheimer effect the paradox can be further enhanced by manipulating the minor losses
is necessary to achieve a flow switch. The second and third results, which (Supplementary Information section S3.3).
hold even for when dynamic pressure is non-negligible, show that this
model captures the flow switching behaviour observed in the simulations Networks with multiple programmed switches
and experiments. Importantly, we validate the flow-switching condition The system considered thus far can be generalized to create larger
in equation (3) by demonstrating quantitative agreement between the microfluidic networks with multiple flow switches—that is, networks
model and simulations both when the static pressure and when the total with multiple disjoint channel segments in which the flow initially in one
pressure is controlled (Supplementary Information section S3.2). direction can be individually ‘switched’ to move in the opposite direction
The model also predicts Braess’s paradox as observed in our experi- through the manipulation of one driving pressure alone. In our design,
ments and simulations. Specifically, under the condition that equation (3) the linking channel plays the role of a switch (and can be referred to as
is satisfied and dynamic pressure is small (or static pressure is controlled), such). Figure 1b shows the multiswitch generalization of the network
the model predicts the paradox to occur for δPin > 0 if and only if in Fig. 1a, which incorporates multiple linking channels and a subset
of channel segments with obstacles. We experimentally demonstrate
a  an instance of a six-switch network that exhibits flow switching in all
K ′(0)βf   > c (4)
β linking channels (as presented in Supplementary Information section

650 | Nature | Vol 574 | 31 October 2019

 a a1 f b
d
Pin
O1
1 e 2 Switching Initial
sequence [ state
b1
O2 2 4 6 8 10 1 3 5 7 9
3 c1
a2 d 4 O1
O3
Pin 5 6 O2
b2

Outlets
c2 O4 O3
a3 7 8
O4
O5
9 10 O5
b3
O6 O6
d c3

c 0.25 d 0.25
Flow rate (μl s–1 per mm depth)

Flow rate (μl s–1 per mm depth)

0.00 0.00

2 4 6 8 10 2 4 6 8 10
–0.25 –0.25
0 100 200 300 400 500 600 0 100 200 300 400 500 600
3 3

0 0

1 3 5 7 9 1 3 5 7 9
–3 –3
0 200 400 600 800 1,000 1,200 1,400 1,600 0 200 400 600 800 1,000 1,200 1,400 1,600
Inlet static pressure (Pa) Inlet static pressure (Pa)

Fig. 5 | Flow patterns in a multiswitch network. a, Schematic of ten-switch fluids is assumed to occur when passing through the same channel segment.
network. Fluids of different colours are driven to each inlet by a common static c, d, Model predictions (c) and simulation results of the Navier–Stokes equations
pressure source, Pin. The outlets are labelled by O1–O6 and the linking channels (d) for the flow rate through each linking channel for a network designed to
by 1–10. The arrows indicate the flow direction through each linking channel and exhibit the switching sequence in b. The flow rates are labelled according to the
multicoloured circles schematically indicate the fluid composition at each channels in a and are divided into two sets (top and bottom panels) for clarity.
outlet for an initially low Pin. The segment lengths are denoted by ai, bi, ci, d, e and Positive flow rates correspond to flow in the upward direction in a, and each flow
f, where a common length is assumed for all linking channels and the segments switch occurs when the corresponding curve crosses the horizontal axis. The
with obstacles are marked with filled grey circles. b, Patterns of outlet flows for segment dimensions that give rise to the particular switching order in b–d are
the network programmed with a chosen switching sequence as Pin is increased. reported in Supplementary Table 1. All 21 possible outlet flow colour
Each column of coloured circles denotes the outlet flows after the combinations are realized between the switching sequence presented here and
corresponding flow switch occurs, where mixing between different coloured those in Supplementary Fig. 13.

S6.2). Multiswitch networks can be designed by extending the network More generally, for a multiswitch network with nc horizontal channels
model presented above. interconnected by nl linking channels, the number of possible internal
One such network with ten linking channels is presented in Fig. 5a. By flow states is nl + 1 if each linking channel exhibits a flow switch. In addition,
marking each inlet flow with a different colour, we show that a variety the possible number of unique colour combinations in the outlet flows is
of patterns can form in the outlet flows (coloured circles in Fig. 5). The nc(nc + 1)/2 if each inlet flow is marked with a different colour. All colour
specific pattern at an outlet depends on the order in which the flow combinations can be realized over the set of all switching sequences,
switches occur as Pin is varied. The network model for larger systems provided that there exist flow paths allowing mixing of every set of k adja-
is constructed by combining pressure–flow relations for each channel cent colours for k ranging from 1 to nc. The myriad states possible in such
segment with flow rate conservation equations for each junction. Using multiswitch networks underlie their ability to process inputs into multiple
this model, we can design a network for which each flow switch occurs outputs and thus to support various applications, including implement-
near a target value of Pin by optimizing the dimensions of the channel ing different mixing orders of chemical reagents and devising schemes
segments (Methods). for the parallel generation of mixtures with tunable concentrations.
As illustrated in Fig. 5, a set of 11 different internal flow states and 17
unique colour combinations at the outlets are possible for the switching Conclusions and outlook
sequence realized in Fig. 5b. Figure 5c, d shows the agreement between the The flow switch, conductance transitions and Braess’s paradox estab-
model predictions of these flow states and results from direct simulations lished in this study are all emergent behaviours of common origin result-
of the Navier–Stokes equations. This variety of states (and output patterns) ing from nonlinearity and interactions between different parts of the
is achieved with only three channel segments containing obstacles and system. The nonlinearity is directly determined by fluid inertia effects,
is parameterized by a single control variable—the driving pressure Pin. which can be enhanced and manipulated through the placement of
Moreover, the switching is implemented solely through the working fluid, obstacles and has the advantage of not being reliant on flexible com-
which differs from existing approaches that rely on flexible valves and ponents, fluid compressibility or dedicated control flows. The onset of
additional control flows15. Thus, multiswitch networks exhibit several prop- Braess’s paradox is marked by the flow-switching pressure, above which
erties exploitable in the design of new controllable microfluidic systems. the increased resistance of the nonlinear channel causes the flow to be

Nature | Vol 574 | 31 October 2019 | 651

Article
routed in the negative direction through the linking channel. When 14. Sackmann, E. K., Fulton, A. L. & Beebe, D. J. The present and future role of microfluidics in
biomedical research. Nature 507, 181–189 (2014).
constrained by a diode, the switch in flow direction also enables nega- 15. Leslie, D. C. et al. Frequency-specific flow control in microfluidic circuits with passive
tive conductance transitions. Our results demonstrate an approach elastomeric features. Nat. Phys. 5, 231–235 (2009).
for routing and switching in microfluidic networks through control 16. Mosadegh, B. et al. Integrated elastomeric components for autonomous regulation of
sequential and oscillatory flow switching in microfluidic devices. Nat. Phys. 6, 433–437
mechanisms that are coded into the network structure, thus responding (2010).
to the call for design strategies that allow diverse microfluidic systems 17. Duncan, P. N., Nguyen, T. V. & Hui, E. E. Pneumatic oscillator circuits for timing and control
to be assembled from a small set of core components2,47. of integrated microfluidics. Proc. Natl Acad. Sci. USA 110, 18104–18109 (2013).
18. Duncan, P. N., Ahrar, S. & Hui, E. E. Scaling of pneumatic digital logic circuits. Lab Chip 15,
Here we considered the scenario in which the inlets and the outlets 1360–1365 (2015).
are (separately) held at the same pressure, rendering the network a two- 19. Doh, I. & Cho, Y.-H. Passive flow-rate regulators using pressure-dependent autonomous
terminal system in all cases, because this is the most stringent scenario deflection of parallel membrane valves. Lab Chip 9, 2070–2075 (2009).
20. Collino, R. R. et al. Flow switching in microfluidic networks using passive features and
for flow manipulation. If a multi-terminal system is configured, by allow- frequency tuning. Lab Chip 13, 3668–3674 (2013).
ing the pressures at each of the inlets (and/or outlets) to be varied inde- 21. Stroock, A. D. et al. Chaotic mixer for microchannels. Science 295, 647–651 (2002).
pendently, then the effects that we presented may be further enhanced. 22. Squires, T. M. & Quake, S. R. Microfluidics: fluid physics at the nanoliter scale. Rev. Mod.
Phys. 77, 977–1026 (2005).
Finally, although we focused on boundary conditions in which the inlet 23. Amini, H., Lee, W. & Di Carlo, D. Inertial microfluidic physics. Lab Chip 14, 2739–2761
pressures are controlled, it would be natural to explore in future research (2014).
the scenario in which the controlled variables are the inlet flow rates. We 24. Zhang, J. et al. Fundamentals and applications of inertial microfluidics: a review. Lab Chip
16, 10–34 (2016).
anticipate, for example, that the negative conductance transitions would 25. Tesař, V. & Bandalusena, H. C. H. Bistable diverter valve in microfluidics. Exp. Fluids 50,
then be converted into pressure amplification (pressure release) transi- 1225–1233 (2011).
tions in which the inlet–outlet pressure difference increases (decreases) 26. Amini, H. et al. Engineering fluid flow using sequenced microstructures. Nat. Commun. 4,
1826 (2013).
abruptly at the transition point. Accordingly, Braess’s paradox is also 27. Sudarsan, A. P. & Ugaz, V. M. Multivortex micromixing. Proc. Natl Acad. Sci. USA 103,
expected to take a complementary form in which closing the linking 7228–7233 (2006).
channel causes the inlet–outlet pressure difference to drop. Incidentally, 28. Di Carlo, D., Edd, J. F., Humphry, K. J., Stone, H. A. & Toner, M. Particle segregation and
dynamics in confined flows. Phys. Rev. Lett. 102, 094503 (2009).
it is this complementary form of Braess’s paradox that has been previ- 29. Wang, X. & Papautsky, I. Size-based microfluidic multimodal microparticle sorter. Lab
ously established for electrical circuits40, thus suggesting an additional Chip 15, 1350–1359 (2015).
30. Xia, H. M. et al. Analyzing the transition pressure and viscosity limit of a hydroelastic
correspondence between electronic and microfluidic circuits.
microfluidic oscillator. Appl. Phys. Lett. 104, 024101 (2014).
31. Braess, D. Über ein Paradoxon aus der Verkehrsplanung. Unternehmensforschung 12,
258–268 (1968).
Online content 32. Braess, D., Nagurney, A. & Wakolbinger, T. On a paradox of traffic planning. Transport. Sci.
39, 446–450 (2005).
Any methods, additional references, Nature Research reporting summa- 33. Rojas, S. & Koplik, J. Nonlinear flow in porous media. Phys. Rev. E 58, 4776–4782
ries, source data, extended data, supplementary information, acknowl- (1998).
34. Andrade, J. S. Jr, Costa, U. M. S., Almeida, M. P., Makse, H. A. & Stanley, H. E. Inertial effects
edgements, peer review information; details of author contributions
on fluid flow through disordered porous media. Phys. Rev. Lett. 82, 5249–5252 (1999).
and competing interests; and statements of data and code availability 35. Fourar, M., Radilla, G., Lenormand, R. & Moyne, C. On the non-linear behavior of a laminar
are available at https://doi.org/10.1038/s41586-019-1701-6. single-phase flow through two and three-dimensional porous media. Adv. Water Resour.
27, 669–677 (2004).
36. Adams, M. L., Johnston, M. L., Scherer, A. & Quake, S. R. Polydimethylsiloxane based
1. Pennathur, S. Flow control in microfluidics: are the workhorse flows adequate? Lab Chip microfluidic diode. J. Micromech. Microeng. 15, 1517–1521 (2005).
8, 383–387 (2008). 37. Calvert, B. & Keady, G. Braess’s paradox and power-law nonlinearities in networks. J. Aust.
2. Stone, H. A. Microfluidics: tuned-in flow control. Nat. Phys. 5, 178–179 (2009). Math. Soc. Ser. B 35, 1–22 (1993).
3. Perdigones, F., Luque, A. & Quero, J. M. Correspondence between electronics and fluids 38. Penchina, C. M. Braess’s paradox and power-law nonlinearities in five-arc and six-arc two-
in MEMS: designing microfluidic systems using electronics. IEEE Ind. Electron. Mag. 8, terminal networks. Open Transplant. J. 3, 8–14 (2009).
6–17 (2014). 39. Ayala, L. F. & Blumsack, S. The Braess paradox and its impact on natural-gas-network
4. Thorsen, T., Maerkl, S. J. & Quake, S. R. Microfluidic large-scale integration. Science 298, performance. Oil Gas Facilities 2, 52–64 (2013).
580–584 (2002). 40. Cohen, J. E. & Horowitz, P. Paradoxical behavior of mechanical and electrical networks.
5. Geertz, M., Shore, D. & Maerkl, S. J. Massively parallel measurements of molecular Nature 352, 699–701 (1991).
interaction kinetics on a microfluidic platform. Proc. Natl Acad. Sci. USA 109, 16540– 41. Youn, H., Gastner, M. T. & Jeong, H. Price of anarchy in transportation networks: efficiency
16545 (2012). and optimality control. Phys. Rev. Lett. 101, 128701 (2008).
6. Seker, E. et al. Nonlinear pressure-flow relationships for passive microfluidic valves. Lab 42. Nicolaou, Z. G. & Motter, A. E. Mechanical metamaterials with negative compressibility
Chip 9, 2691–2697 (2009). transitions. Nat. Mater. 11, 608–613 (2012).
7. Weaver, J. A., Melin, J., Stark, D., Quake, S. R. & Horowitz, M. A. Static control logic for 43. Pala, M. G. et al. Transport inefficiency in branched-out mesoscopic networks: an analog
microfluidic devices using pressure-gain valves. Nat. Phys. 6, 218–223 (2010). of the Braess paradox. Phys. Rev. Lett. 108, 076802 (2012).
8. Tanyeri, M., Ranka, M., Sittipolkul, N. & Schroeder, C. M. Microfluidic Wheatstone bridge 44. Motter, A. E. & Timme, M. Antagonistic phenomena in network dynamics. Annu. Rev.
for rapid sample analysis. Lab Chip 11, 4181–4186 (2011). Condens. Matter Phys. 9, 463–484 (2018).
9. Kim, S.-J., Lai, D., Park, J. Y., Yokokawa, R. & Takayama, S. Microfluidic automation using 45. Crane Co. Engineering Division. Flow of Fluids through Valves, Fittings, and Pipe.
elastomeric valves and droplets: reducing reliance on external controllers. Small 8, Technical paper no. 410 (Crane Co., 2010).
2925–2934 (2012). 46. Khodaparast, S., Borhani, N. & Thome, J. R. Sudden expansions in circular microchannels:
10. Li, L., Mo, J. & Li, Z. Nanofluidic diode for simple fluids without moving parts. Phys. Rev. flow dynamics and pressure drop. Microfluid. Nanofluidics 17, 561–572 (2014).
Lett. 115, 134503 (2015). 47. Bhargava, K. C., Thompson, B. & Malmstadt, N. Discrete elements for 3D microfluidics.
11. Chin, C. D., Linder, V. & Sia, S. K. Commercialization of microfluidic point-of-care Proc. Natl Acad. Sci. USA 111, 15013–15018 (2014).
diagnostic devices. Lab Chip 12, 2118–2134 (2012).
12. Araci, I. E., Su, B., Quake, S. R. & Mandel, Y. An implantable microfluidic device for self- Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
monitoring of intraocular pressure. Nat. Med. 20, 1074–1078 (2014). published maps and institutional affiliations.
13. Bhatia, S. N. & Ingber, D. E. Microfluidic organs-on-chips. Nat. Biotechnol. 32, 760–772
(2014). © The Author(s), under exclusive licence to Springer Nature Limited 2019

652 | Nature | Vol 574 | 31 October 2019

Methods nonlinearity53,54). Through additional experiments, we confirmed that
pressure–flow relations similar to those in Fig. 2e, f hold for channels
Navier–Stokes simulations constructed from materials with both higher rigidity (SU-8 photoresist)
The numerical simulations were performed using48 OpenFOAM version and lower rigidity (Flexdym) than the PDMS (Supplementary Informa-
4.1. We used meshes with an average cell area ranging from 10 μm2 to tion section S5 and Supplementary Fig. 10). We note that porous-like
340 μm2, where the finest meshing was applied near the obstacles. All structures have been previously used both to study non-inertial effects
meshes were generated using Gmsh49. The two-dimensional solutions in microfluidics, such as droplet formation55 and viscous fingering56,
were found using the simpleFoam solver within OpenFOAM, employ- and to study inertial effects in larger systems57. In our system, inertial
ing second-order numerical schemes, where a fixed static pressure of effects arise at the microfluidic scale even for a much smaller number
zero was set for the boundary conditions at the outlets. At the inlets, of obstacles than the typical number in porous-like materials.
the static (total) pressure was fixed for the static (total) pressure con-
trolled cases. For simulations of the multiswitch network in Fig. 5, the Network flow model construction
same geometry and dimensions were used as for the model predictions, The analytic model used to describe the system in Fig. 1a is constructed
provided in Supplementary Table 1, and equal driving pressures were as follows: (i) we consider the pressure at the inlets Pin to be in the vicin-
applied at each of the six inlets. ity of Pin⁎ ; (ii) we approximate the pressure–flow relation through the
linking channel as Q3 = κ(γP1 − P2), where κ is the channel conductivity
Reynolds numbers and γ is a free parameter allowing for an effective pressure difference;
The characteristic length scale used in defining the Reynolds number of (iii) the flow equation for each other channel segment without obstacles
the flow is the hydraulic diameter of the channels, defined as 4A/P, where is written as in equation (1), where −ΔP is the pressure drop along the
A is the area and P is the perimeter of the channel cross-section (common segment and L is the segment length; (iv) for the channel segment with
to all segments). The hydraulic diameter in two and three dimensions is obstacles, we take the flow equation to be in the form of equation (2)
2w and 2wh/(w + h), respectively, where h is the height of the channels in (with Re expressed as 2ρQ/μ); (v) we include the most dominant term
the three-dimensional case. The characteristic velocity used in two and resulting from minor pressure losses at the channel junctions. Therefore,
three dimensions is Q/w and Q/wh, respectively. Therefore, we define the model consists of five pressure–flow relations, in addition to two
Re = 2ρQ/μ for our simulations in two dimensions and Re = 2ρQ/μ(w + h) flow conservation equations at the junctions: Q3 + Q2 − Q4 = 0 and
for our experiments in three dimensions. The undeclared ranges of Re Q3 + Q5 − Q1 = 0. When the static pressure is controlled at the inlets, the
for the channel segment with obstacles considered in the presented data only nonlinearity that exists in the model comes from the Forchheimer
are: 21–385 (Fig. 3), 12–121 (Fig. 4), 1–220 (Fig. 5), 1–380 (Supplementary term due to the presence of obstacles and the minor loss term. The
Fig. 2), 4–111 (Supplementary Fig. 4), 40–385 (Supplementary Fig. 7), model can also be adapted for when total pressure is controlled by tak-
20–400 (Supplementary Fig. 8), 2–10 (Supplementary Fig. 11b), 75–85 ing the static pressure at each inlet to be Pin − kρQ2/2w2, where Pin now
(Supplementary Fig. 11c), 76–89 (Supplementary Fig. 12), 10–20 (Sup- denotes total pressure and the coefficient k is a constant of order unity
plementary Fig. 14b) and 110–120 (Supplementary Fig. 14c). that only depends on the shape of the inlet velocity profile (k ≈ 1 for a
uniform velocity profile at the inlet, as considered here). However, the
Pressure boundary conditions dynamic pressure term ρQ2/2w2 is often negligible in real microfluidic
We consider two different boundary conditions for the driving pressure systems because of the high pressures needed to drive fluid though the
Pin at the system inlets. Under one condition, total pressure is controlled channels. Indeed, in our experiments, the dynamic pressure near Pin⁎
and the inlets open directly into a high-pressure reservoir. Under the was smaller than the static pressure by two orders of magnitude and
other condition, static pressure is controlled and the inlets are connected smaller than the pressure loss due to the Forchheimer effect by one
to the reservoir by pressure regulators. Total pressure is the sum of order of magnitude. This can also be seen in Fig. 2f, where a constant
static pressure and dynamic pressure, where dynamic pressure is defined relation between Re and ΔP/Re is measured. Details of the model are
1
as 2 ρv 2 for a fluid with density ρ and velocity v. The distinction between presented in Supplementary Information section S1.
these boundary conditions is often neglected in the microfluidics lit-
erature when the Reynolds number is less than one50, but it can become Designing multiswitch networks
important for larger Reynolds numbers (even though the flow remains For a network with multiple switches and a given set of channel dimen-
laminar)51. sions, the value of Pin for which a specific flow switch occurs can be deter-
mined through the addition of a constraint to the model that enforces
Pressure–flow relations the flow through the corresponding linking channel to be zero. Then,
We use equation (1) to describe the pressure–flow relation for straight, the dimensions of a chosen subset of channel segments may be varied
obstacle-free channels; it is derived directly from the Navier–Stokes through an optimization procedure in order to design a network for
equations by assuming plane Poiseuille flow through a two-dimensional which each flow switch occurs near a target value of Pin. Depending on
channel. To describe the nonlinear pressure–flow relation observed which dimensions are allowed to be adjusted, the desired relative order
for the channel with obstacles, we refer to the Forchheimer equation: of the switches can be achieved exactly, and the final set of switching
−ΔP = αμLV + βρLV2, where V is the average fluid velocity. In two dimen- pressures can be very close to the target ones (often <5% difference),
sions, V = Q/w = μRe/2ρw and, thus, the Forchheimer equation can be where the former is expected to be more important in applications.
written in the form of equation (2). In agreement with equation (2), we Further details on the design of multiswitch networks are presented
find an excellent linear fit between −ΔP/Re and Re for a channel with ten in Supplementary Information section S6.1.
obstacles, and we validate the fit by predicting flows through the same
channel for a fluid with a different viscosity (Supplementary Information PDMS channel fabrication
section S3.1). We observe no unsteady flow through the channel with The flow channels were assembled by sealing a patterned PDMS chip
obstacles due to vortex shedding for Re of up to 400, as expected for against a glass slide. The PDMS chip was made by pouring a mixture of
systems with highly confined obstacles52, which permits the use of the PDMS oligomer and cross-linking curing agent (Sylgard 184) at a weight
steady-state relation in equation (2) over the range of Re considered here. ratio of 10:1 into a mould after being degassed under vacuum. The mixture
We experimentally verify the source of nonlinearity in PDMS channels was cured at 74 °C for 1 h and then peeled off from the mould to yield the
with obstacles, which were designed to have approximately square cross- microchannel design. The dimensions of the channels in Figs. 2 and 4 were
sections to minimize deformation (which could lead to other forms of 200 μm (width) × 185 μm (height), and the diameter of the obstacles was
Article
97 μm. After punching the holes for inlet and outlet connections, the PDMS rate at each outlet. Red (3 g l−1, FD&C Red #40, Flavors & Colours) and
chip was thermally aged at 200 °C for 12 h to reduce pressure-induced blue (1.5 g l−1, FD&C Blue #1, Flavors & Colours) dyes were added into DI
deformation58, yielding a chip with a Young’s modulus of59 approximately water to demonstrate the switching behaviour. The concentrations of
3 MPa. Both the PDMS chip and the glass substrate were cleaned with the dyes were adjusted for similar flow rate under the same pressure.
isopropanol and treated by plasma for 90 s before bringing them into The flow rate measurements in Supplementary Fig. 10 were performed
contact. Once the PDMS chip was sealed against the glass slide, the device using isolated channels constructed from Flexdym and SU-8 photore-
was placed in an oven for 30 min at 74 °C to improve bonding quality. sist, respectively.
The mould used was a silicon wafer containing microchannel pat-
terns created by soft photolithography using a negative photoresist60,61. Fluorescence imaging
A 4-inch silicon wafer (test grade, University Wafer, Boston, MA) was Fluorescent polyethylene microspheres (10–20 μm) were suspended
cleaned with acetone and isopropanol and dried with nitrogen gas. The in Tween 80 solution (Cospheric LLC, Santa Barbara, CA) and pumped
wafer was then coated with SU-8 50 negative photoresist (MicroChem through a single microfluidic channel with obstacles by an Elveflow OB1
Corp., Newton, MA) on a spin coater (Laurell Technologies Corp., North pressure controller. Two different pressures were applied, 3 kPa and
Wales, PA) operating at 600 rpm for 30 s. After a pre-exposure bake at 100 kPa, to demonstrate different flow profiles around the obstacles.
65 °C and subsequently at 95 °C, each for 60 min, the coated wafer was Fluorescence images were captured with an Olympus BX51 microscope
exposed to UV light (Autoflood 1000, Optical Associates, Milpitas, CA) equipped with a NIBA filter through an Infinity 3 CCD camera.
through a negative transparent photomask that contained the desired
channel design. Following a 3.5 min post-exposure bake at 95 °C, the Measured flow rate data and statistics
wafer was developed in SU-8 developer (MicroChem Corp., Newton, Savitsky–Golay filtering was applied to all flow rate data collected
MA) for 60 min to obtain the pattern. through experiments, using a window length of 11 data points and a
second-order polynomial. For each of the fixed pressures presented
Flexdym channel fabrication in Fig. 4b–d, a 60 s time series of flow rate data was collected at each
Flexdym (Blackholelab Inc., Paris) is a thermoplastic elastomer (Young’s of the outlets with a sampling rate of 10 Hz. Over the 60 s interval, the
modulus of 1.18 MPa) with a rapid and easy moulding process for micro- linking channel valve was sequentially opened/closed every 15 s. For
fluidic devices62. After fabrication of the silicon wafer mould containing each time series, the 15 s intervals in which the valve was open (closed)
the channel designs, a sheet of Flexdym (6 cm × 4 cm) was placed directly were averaged to create a single 15 s time series for each outlet. The total
above the mould with another sheet of unpatterned PDMS (about 1 mm flow rate (Q4 + Q5) was calculated when the valve is open and closed,
thick) placed above the Flexdym for protection. The whole set was then respectively, by summing the 15 s time series for the two outlets point-
placed on a heat press between two Teflon sheets. The plate on the heat by-point. The statistics presented in Fig. 4 are the average and standard
press was heated to 175 °C before starting to mould the Flexdym. Once deviation of the resulting series. For Supplementary Fig. 12, the flow rate
the target temperature was reached, the lever on the heat plate was at each of the two outlets was measured experimentally at a sampling
locked down with a timer set for 5 min. After the process was finished, rate of 100 Hz over a 180 s interval, during which the linking channel
the lever was released and the Flexdym sheet was inspected visually to was sequentially opened/closed every 30 s. The total flow rate in Sup-
make sure that no bubbles were trapped around the channel. The chip plementary Fig. 12c was calculated by summing, point-by-point, the
was allowed to cool down for 5 min before unfolding the layers. The data in Supplementary Fig. 12a, b.
Flexdym was permanently sealed with a glass slide by following the same
sealing procedure used for the PDMS channels. The dimensions of the Parameters in simulations and experiments
cross-section of the channels were 201 μm (width) × 166 μm (height), In the simulations, we set ρ = 103 kg m−3, μ = 10−3 Pa s, ν = μ/ρ = 10−6 m2 s−1,
and the diameter of the obstacles was 99 μm. w = 500 μm for the width of all channels, and r = 100 μm for the radius of
all obstacles, unless otherwise noted. In all experiments, DI water was
SU-8 photoresist channel fabrication used as the working fluid. The other undeclared dimensions were as fol-
To make microfluidic channels directly from SU-8 photoresist, an inverse lows. In Fig. 2a, b, the length of the (partially shown) channel was 1.25 cm.
mask was designed and printed on transparency. The desired channel In Fig. 2c–e, the channel length was 4.3 cm, and in Fig. 2f the channel
was printed on the inverse mask in black with transparent dots mark- length was 2.0 cm (see Methods section ‘PDMS channel fabrication’ for
ing the obstacles, and the rest of the mask was left transparent. The the remaining dimensions). In Fig. 3, L1 = 0.17 cm, L2 = 1.0 cm, L3 = 0.1 cm,
same procedure used to make the silicon wafer master as described in L4 = 1.25 cm, and L5 = 1.0 cm. In Fig. 4, L1 = 0.6 cm, L2 = 2.9 cm, L4 = 1.4 cm
Methods section ‘PDMS channel fabrication’ was followed to fabricate and L5 = 1.4 cm. For the linking channel, the switch valve was connected
the channels on glass slides. The chip was then sealed by 3M VHB tape to the two parallel channels through 15 cm of round tubing and 0.7 cm of
to another glass slide with holes for connections. The dimensions of the microchannel on each side. Each inlet was connected to the pressurized
cross-section of the channels were 209 μm (width) × 196 μm (height), vials through 62 cm of tubing, and each outlet was attached to 50 cm of
and the diameter of the obstacles was 90 μm. The Young’s modulus of tubing. The inner diameter of all tubing was 0.79 mm.
SU-8 photoresist is 2 GPa (from the table of properties for SU-8 perma-
nent photoresists, MicroChem Corp., Newton, MA, available at http://
microchem.com/pdf/SU-8-table-of-properties.pdf). Data availability
The datasets generated and/or analysed during the current study are
Flow rate measurement available from the corresponding author on reasonable request.
Experimental measurements in Figs. 2 and 4 were made with the sys-
tem shown in Fig. 4a. When measuring the relation between pressure
and flow rate, the linking channel valve was closed to allow separate Code availability
measurement of the channel with and the channel without obstacles. Custom Python code is available from the corresponding author on
Deionized (DI) water was pumped through each channel and a pressure request.
scan from 0 to 100 kPa was performed using an Elveflow OB1 pressure
controller. The flow rate was measured by an Elveflow MFS5 flow sen- 48. OpenFOAM v4.1 (OpenFOAM Foundation, 2016).
49. Geuzaine, C. & Remacle, J.-F. Gmsh: a three-dimensional finite element mesh generator
sor (0.2–5 ml min−1). To verify Braess’s paradox, the same instruments with built-in pre- and post-processing facilities. Int. J. Numer. Methods Eng. 79, 1309–1331
were used and the pressure was set constant while recording the flow (2009).
50. Oh, K. W., Lee, K., Ahn, B. & Furlani, E. P. Design of pressure-driven microfluidic networks 61. Duffy, D. C., McDonald, J. C., Schueller, O. J. A. & Whitesides, G. M. Rapid prototyping
using electric circuit analogy. Lab Chip 12, 515–545 (2012). of microfluidic systems in poly(dimethylsiloxane). Anal. Chem. 70, 4974–4984
51. Zeitoun, R. I., Langelier, S. M. & Gill, R. T. Implications of variable fluid resistance caused (1998).
by start-up flow in microfluidic networks. Microfluid. Nanofluidics 16, 473–482 (2014). 62. Lachaux, J. et al. Thermoplastic elastomer with advanced hydrophilization and bonding
52. Zovatto, L. & Pedrizzetti, G. Flow about a circular cylinder between parallel walls. J. Fluid performances for rapid (30 s) and easy molding of microfluidic devices. Lab Chip 17,
Mech. 440, 1–25 (2001). 2581–2594 (2017).
53. Gervais, T., El-ali, J., Gunther, A. & Jensen, K. F. Flow-induced deformation of shallow
microfluidic channels. Lab Chip 6, 500–507 (2006).
54. Christov, I. C., Cognet, V., Shidhore, T. C. & Stone, H. A. Flow rate–pressure drop relation Acknowledgements This research was supported by the US National Science Foundation
for deformable shallow microfluidic channels. J. Fluid Mech. 841, 267–286 (2018). (grants PHY-1001198 and CHE-1900011), the Simons Foundation (award number 342906) and a
55. Amstad, E., Datta, S. S. & Weitz, D. A. The microfluidic post-array device: high throughput Northwestern University Presidential Fellowship.
production of single emulsion drops. Lab Chip 14, 705–709 (2014).
56. Haudin, F., Callewaert, M., De Malsche, W. & De Wit, A. Influence of nonideal mixing Author contributions D.J.C., J.-R.A. and A.E.M. designed the overall study and formulated
properties on viscous fingering in micropillar array columns. Phys. Rev. Fluids 1, 074001 the theory. Y.L. and I.Z.K. designed and performed the experiments. D.J.C. implemented the
(2016). numerical simulations and analyses. All authors contributed to the writing of the
57. Zhao, H., Liu, Z., Zhang, C., Guan, N. & Zhao, H. Pressure drop and friction factor of a manuscript, which was led by D.J.C. and A.E.M. All authors reviewed and approved the final
rectangular channel with staggered mini pin fins of different shapes. Exp. Therm. Fluid manuscript.
Sci. 71, 57–69 (2016).
58. Kim, M., Huang, Y., Choi, K. & Hidrovo, C. H. The improved resistance of PDMS to pressure- Competing interests The authors declare no competing interests.
induced deformation and chemical solvent swelling for microfluidic devices.
Microelectron. Eng. 124, 66–75 (2014). Additional information
59. Johnston, I. D., McCluskey, D. K., Tan, C. K. L. & Tracey, M. C. Mechanical characterization Supplementary information is available for this paper at https://doi.org/10.1038/s41586-019-
of bulk sylgard 184 for microfluidics and microengineering. J. Micromech. Microeng. 24, 1701-6.
035017 (2014). Correspondence and requests for materials should be addressed to A.E.M.
60. Martin, R. S., Gawron, A. J., Lunte, S. M. & Henry, C. S. Dual-electrode electrochemical Peer review information Nature thanks Sujit Datta and Dino Di Carlo for their contribution to
detection for poly(dimethylsiloxane)-fabricated capillary electrophoresis microchips. the peer review of this work.
Anal. Chem. 72, 3196–3202 (2000). Reprints and permissions information is available at http://www.nature.com/reprints.