A mathematical model for

rspa.royalsocietypublishing.org fitting and predicting

relaxation modulus and
Research
simulating viscoelastic
Cite this article: Xu Q, Engquist B. 2018 A
responses
mathematical model for fitting and predicting Qinwu Xu1 and Björn Engquist1,2
relaxation modulus and simulating viscoelastic
1 Institute for Computational Engineering and Sciences, and
responses. Proc. R. Soc. A 474: 20170540.
2 Department of Mathematics, The University of Texas at Austin,
http://dx.doi.org/10.1098/rspa.2017.0540
Austin, TX 78712, USA
Received: 11 August 2017 QX, 0000-0002-1366-934X
Accepted: 6 March 2018
We propose a mathematical model for relaxation
modulus and its numerical solution. The model
Subject Areas: formula is extended from sigmoidal function
civil engineering, materials science, mechanics considering nonlinear strain hardening. Its physical
meaning can be interpreted by a macroscale elastic
Keywords: network-viscous medium model with only five model
mathematical model, relaxation modulus, parameters in a simpler format than the molecular-
viscoelasticity, finite element, displacement, chain-based polymer models to represent general
solid materials. We also developed a finite-element
stress
(FE) framework and robust numerical algorithm to
implement this model for simulating responses under
Authors for correspondence: both static and dynamic loadings. We validated the
Qinwu Xu model through both experimental data and numerical
e-mail: [email protected] simulations on a variety of materials including asphalt
concrete, polymer, spider silk, hydrogel, agar and
Björn Engquist
bone. By satisfying the second law of thermodynamics
e-mail: [email protected] in the form of Calusius–Duhem inequality, the model
is able to simulate creep and sinusoidal deformation
as well as energy dissipation. Compared to the
Prony series, the widely used model with a large
number of model parameters, the proposed model
has improved accuracy in fitting experimental data
and prediction stability outside of the experimental
range with competitive numerical stability and
computation speed. We also present simulation
results of nonlinear stress–strain relationships of
spider silk and hydrogels, and dynamic responses of
a multilayer structure.

Electronic supplementary material is available

online at https://doi.org/10.6084/m9.figshare. 1. Introduction
c.4088969.
Almost all materials exhibit viscoelastic behaviour more
or less. Viscoelastic material exhibits both elastic and

2018 The Author(s) Published by the Royal Society. All rights reserved.
 (a) (b) (c) (d)
s(t) 2

E1 E2 ... En

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
E•
h2 hn
h1
s(t)

Figure 1. Material viscoelastic models: (a) friction beads–spring model; (b) tube model; (c) arm-retraction model; and
(d) generalized Maxwell model. (Online version in colour.)

viscous characteristics, and its deformation is temperature dependent. Relaxation modulus E(t) is
a characteristic of material viscoelasticity as used to describe the stress relaxation of materials with
time (t). It is important to accurately simulate the stress relaxation and viscoelastic deformation
of subjects in order to evaluate and design materials. Thermal transitions of viscoelastic materials
can be described either in terms of free volume changes or relaxation time [1]. The Crankshaft
model [2] simulates polymer molecules as a series of jointed segments involving in a few stages
of thermal transition with elevated time or temperature. For example, with the elevation of
temperature, the first or γ transition may start with the local motion of molecules, and then
the β transition may appear when E(t) drops slightly due to the bend and stretch of molecules
with elevated temperatures. Consequently, the glass or α transition occurs when E(t) significantly
reduces until reaching the rubbery stage, and lastly, the terminal transition may exit when the
polymer melts into liquids due to the slippage of molecular chains. Different materials may
present different thermal transition behaviours, but generally they all include the glass transition
at the solid state, which is also the focus of this study. Different theories and physical models for
describing E(t) were proposed primarily in the field of polymer science. The majority of these
models focused on linear viscoelasticity. Among these, the Rouse model simulated the single
chain diffusion of polymers as Brownian motion of a beads–harmonic-springs system based on
the theory of molecular dynamics (figure 1a) [3], and the Kremer–Grest model used hundreds
of chains and beads in its simulation framework [4] The well-known (entangled) tube model
has defined entangled polymer chains that are confined in tubes with permanent topological
interactions and move along tubes (figure 1b) [5]. This model especially concerned the phenomena
associated with polymers of complex topology or long chain branching [6]. The stress relaxation of
each chain is calculated as the fraction of the tube that has not been vacated, where the relaxation
time is related to the molecular mass of the cube [7]. The arm-retraction model [8] described the
entangled monomers (unit of polymer macromolecular) as retracted by arms (figure 1c), which
was validated for the rheology of entangled polymer liquids.
Maxwell model proposed by Maxwell and Wiechert in late nineteenth century consists of
a linear spring and a dashpot in series. With an additional spring in parallel, it is known as
the standard linear solid or Zener model. For the dashpot itself stress has a linear relationship
with strain rate as σ (t) = ηε̇(t), where η is viscosity. However, the Maxwell model can only
capture the relaxation behaviour of polymers in a very limited time range. Accordingly, a series
of such models have been assembled in parallel, delivering a physical system referred to as
the generalized Maxwell (GM) model (figure 1d) to improve accuracy in fitting experimental
data. The combination of elastic springs and viscous dashpots in this physical system could be
interpreted by variable molecular-chain-segment lengths under different time distributions for
polymer structures [9]. The GM model expresses E(t) as follows:

n 
n
E(t) = E∞ + Ei e−(Ei /ηi ) t = E∞ + Ei e−t/τi , (1.1)
i=1 i=1

where E∞ is the model’s elastic modulus at t = ∞, Ei is the elastic modulus of the spring and ηi
is the viscosity of the dashpot for the ith series, respectively, n is the number of spring–dashpot
series, and τi = ηi /Ei is retardation representing a time range for modulus reduction from Ei to 0
3
for a single spring–dashpot series.
The E(t) value of this physical system is the sum of Ei distributed at variable time spectrums.

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
This formula with accumulated exponential terms in a discrete spectrum with a finite number of
series is also referred to as Prony series (PS). The model parameters are determined by fitting on
experimental data using mathematical optimization skills. Usually a higher n value may achieve
a higher fitting accuracy, but also involves more complexity and variability as will be discussed
later. The GM model or the PS has been widely used to describe the linear viscoelastic behaviour
of a broad range of solid materials. Examples of these materials include polymers [10], waterproof
membrane [11], glasses [12], poroelastic medium [13], asphalt concrete (AC) [14], ligament [15]
and skin [16], etc. The numerical implementation of the PS is computationally convenient due to
its exponential format for a closed-form solution of time integration [17]. It has been implemented
as a standard viscoelastic material input in the popular finite-element (FE) software (e.g. ANSYS
and ABAQUS). Dynamic modulus E*(ω) is also used to describe the viscoelastic behaviour based
on frequency domain [18], which is not covered in this paper.
Many other viscoelastic models have also been proposed to simulate specific materials. Iyo
and Sasaki et al. proposed a Kohlraush–Williams–Watts function based model for biomaterials as
follows [19]:
 ϑ ϑ

E(t) = E0 A1 e−(t/τ1 ) + (1 − A1 ) e−(t/τ2 ) , (1.2)

where τ 1 and τ 2 are characteristic time of relaxation process, and ϑ is a model parameter.
The fractional derivative model was proposed to express the stress–strain relationship of
viscosity as a fractional order of time as follows [20]:

ηdεθ (t)
ση (t) = , (1.3)
dtθ
where θ is a non-integer parameter. The stress–strain relationship of the standard viscoelastic
system using a fractional derivative order can be expressed as follows:

σ (t) = E0 ε(t) + E1 Dθ [ε(t)], (1.4)

where D is a fractional derivative, and E0 and E1 are two modulus parameters. The fractional
derivative model can reduce the number of model parameters while attaining a desired accuracy.
It is especially appropriate for predicting the dynamic behaviour of polymers as the non-integer
parameter, θ , offers nonlinearity and flexibility to capture a wide range of relaxation time. In
addition, it may simulate the laboratory-measured asymmetric peak of loss modulus by using
different fractional terms for stress and strain, respectively [21]. However, the fractional derivative
model involves more complexity in fitting experimental data and numerical implementation. The
Huet–Sayegh model expresses E*(ω) in fractional order [22], but it is not designed to simulate
stress relaxation with time domain [23]. Nonlinearity of viscoelastic models have been considered
for solids with large deformation or properties that change with deformation [24]. Schapery’s
model considers the spring’s elastic modulus as a nonlinear function of time [25]. Other models
instead present the dashpot’s stress as a nonlinear function of strain rate such as following the
power-law [26] or exponential function [27]. However, computation instability may arise for
nonlinear models [13,28].
Based on literature review, a few research questions may arise for existing models and
approaches. The (multi-molecular) chain-based models and theories discussed above (e.g. Rouse
model, tube model and arm-retraction model) were primarily derived from and validated for
monomers and polymers. These models may also involve difficulty in characterizing model
parameters and numerical solutions at macroscale. More importantly, the widely used PS and
its modified ones can raise a few research questions. Firstly, the PS only describes linear
viscoelasticity within a small strain range. Secondly, it is known that the PS produces unsatisfied
curve fitting on experimental data [23,29] due to its discrete spectrum as a finite set of exponential
functions. This is especially true when a relatively small number of terms is used as a common
practice (e.g. n = 7 with 15 model parameters or even less [30,31]). With a larger n number
4
data fitting may get more accurate. However, it becomes difficult to fit a large number of
model parameters mathematically with multiple solutions because the optimization intrinsically

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
involves more variability [29]. It also becomes difficult to provide a clear physical interpretation of
this extensive spring–dashpot system, although it can improve data fitting accuracy. In addition,
the PS is primarily used for fitting experimental data rather than predicting E(t) outside of the
experimental range, while prediction may be warranted when experimental tests have limitations
to capture E(t) at the low and high relaxation time ranges. Lastly, strain hardening or softening
may appear in some special materials [32]. This hardening or softening feature has not been
acknowledged for its potential effect in the PS and many other existing (nonlinear) viscoelastic
models [25–27].
Accordingly, the objective of this research is to develop an alternative material model in
a relatively simple format with fewer model parameters for describing E(t) and simulating
viscoelastic responses of solid materials to improve accuracy and efficiency. We extended
our study from an early version published on the arXiv website [33] by conducting more
comprehensive theoretical and experimental validations, including model derivation, detailed
experimental validations and simulations of the stress–strain relationships of broad materials. We
tested the model through experimental data on a variety of materials including asphalt concrete
(AC), shape-memory polymer, hydrogels, spider silk, agar and bone. We also developed a FE
scheme to implement this model for simulating responses. The completion of the model also
included a check for numerical accuracy, stability and convergence of the proposed model versus
that of the PS.

2. Model development
(a) Model formula
Before deriving the model, we reviewed a few continuous-time-spectrum based mathematical
models. Nutting [34] proposed a general stress–strain relationship as relating to a power function
of time that ε = σ tA . E(t) derived as the reciprocal of a power function is called the power law
model [35]:
A
E(t) = , (2.1)
1 + tδ
where A and δ are model parameters. The power law model has the advantage of simplicity
by using only two model parameters. However, it is unable to capture E(t) at the high and low
time ranges (thresholds) accurately [23]. The S-shape sigmoidal function may capture these two
thresholds more accurately as:
1
S(t) = . (2.2)
1 + et
By including additional model parameters into the sigmoidal function, the modified model can
more accurately control the model shape and trend. For example, the modified model is proposed
to fit the absolute value of dynamic modulus for asphalt materials [36,37]:
α
log(|E∗ ( f )|) = δ + , (2.3)
1 + eβ+γ log( f )
where δ, α, β, and γ are fitted model parameters, and f = ω/2π is frequency. The modified
sigmoidal function model has been successfully implemented for fitting experimental data of
materials including bitumen and asphalt. However, as a mathematical function it does not
describe material viscoelasticity based on a continuum-mechanics framework. For example, even
the unit of the left side of equation (2.3) does not match the right side physically. As a result, it
has not been used to simulate viscoelastic responses of structures other than mathematical fitting
on experimental data.
 By extending the continuous-time-spectrum-based models (i.e. the power-law and sigmoidal
5
function), we introduced a new mathematical model with physical mechanism considered to
describe the material viscoelasticity at both experimental and numerical scales, for improving

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
accuracy of existing models. With validation of the model formula on experimental data of a
variety of materials, we proposed the model as follows, where the log–log scale is applied for
improving accuracy:
1
log(E(t)) = log(E∞ ) + (log(E0 ) − log(E∞ )) , (2.4)
1 + μeα log(t/τ0 )
where E∞ is the minimum modulus at the rubbery stage, E0 is the instantaneous modulus at the
glassy stage, μ and α are unitless model parameters, log = log10 for our implementation and τ 0 is
a characteristic time of relaxation process (in seconds) as related to the time after loading required
for the strain to reach its equilibrium, and the value. t/τ 0 becomes unitless such that the left and
right part of the equation equals physically. One shall note that in this model the exponential
term eα log (t/τ0 ) can be converted to a power-law term as proportional to (t/τ 0 )α . However,
the logarithmic scale expression (the same as equation (2.3)) is adopted here for the purposes
of: (1) using logarithmical expression for both E(t) and t to be continuous; (2) the logarithmical
scale of the model improves accuracy in fitting experimental data; and (3) attaining computation
convenience for numerical implementation with the exponential format.
In addition, some viscoelastic materials may exhibit strain hardening behaviour [38] that
represents the strengthening of materials with strain [39]. Accordingly, the proposed model can
be extended to consider the nonlinear strain hardening or softening, in which E(t) decreases with
strain for strain softening as opposite to that of strain hardening, as follows:
1
log E(t) = log E∞ + (log E0 − log E∞ )(1 + ε)β . (2.5)
1 + μeα log(t/τ0 )
where β is a strain hardening or softening coefficient. By setting τ 0 as a constant value such as 1 (s)
for the simplicity purpose, the model parameters are reduced to five. To consider the temperature
effect, a shift factor α T is included to convert time to a reduced time tr at the reference temperature
according to the temperature–time superposition principle as:
(log E0 − log E∞ ) (log E0 − log E∞ )
log E(t) = log E∞ + (1 + ε)β = log E∞ + (1 + ε)β . (2.6)
1 + μeα log(αT t/τ0 ) 1 + μeα log(tr /τ0 )
Here we adopted Arrhenius law to describe the temperature–time superposition [40] for α T
as:  
Ea 1 1
log(αT ) = − − , (2.7)
R T T0
where T is temperature (K), T0 is the reference temperature (K), Ea is a fitted model parameter
that may be referred as an ‘energy’ (J Mol−1 ) required for thermal transition, R is a constant
(like the gas constant in J K−1 Mol−1 ), and Ea /R is a model parameter index (K). Other formulae
for α T have also been proposed including the WLF (Williams–Landel–Ferry) model [41] and the
polynomial function [37]. The WLF equation might be a better description for temperatures above
the glass transition, while the Arrhenius equation could be more suitable at temperatures below
glass transition [42]. The formula of α T shall be properly chosen as dependent on the specific
material and temperature range. In this study, we have not evaluated the effect of α T formula.

(b) Physical interpretation

Network model-based physical mechanisms have been often used for polymers, rubber and
elastomers, although the dynamics of polymer networks is complex and not fully understood
yet in polymer physics. Examples include the coarse-grain polymer network model, in which
the mesh-like networks are stretched in the viscous medium [43], the symmetrically growing
treelike micro-network model built on sub-chains and the non-affine microspheres [44]. However,
 (a) (b) (c) 6

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
(d) (e) (f)

spongious bone
a
E0 Æ E•, b
m
cortical bone

marrow
(g) (h)
100 000
glassy stage glassy stage 1/m increases m = 0.003

10 000 m = 0.049
m = 0.279
a = 0.2
1000 m = 0.768
a = 0.4
log E (t)

a = 0.7
100 E• = 1.5
a = 0.9 a µ ∂E/∂t E0 = 24 723
E• = 1.5 a = 0.4
E0 = 24 723 DE
10 b=0
m = 0.049 Dt
b=0 a increases
1 rubbery stage rubbery stage
1 × 10
1 × 10–14 0.0000001 1 10 000 000 1 × 1014 1 × 10–14 0.0000001 1 10 000 000 1 × 1014
log t log t

Figure 2. Proposed physical mechanism for different viscoelastic materials: (a) agar cake made from algae; (b) a spider silk web;
(c) natural asphalt and its chemical structure; (d) polymer network with cross-linked branched chains; (e) proposed network-
viscous medium schematic; (f ) bone; (g) effects of model parameter α; and (h) effect of model parameter μ on E(t). (Online
version in colour.)

molecular and micromechanical models are more appropriate for specific type of materials,
e.g. the multi-molecular-chain-network models specifically for polymers as discussed above.
We studied a few different materials as plotted in figure 2: (1) agar material (e.g. agar cake,
figure 2a), spider silk (figure 2b) and polymer which belong to the multi-chain cross-linked
network (figure 2d). Spider silk is a protein fibre consisting of short polypeptide stretches of
about 10 to 50 amino acids. These patterns can be repeated more than 100 times within one
individual protein, and each polypeptide duplication results in extremely durable spider silk
threads [45]; and (2) other materials that cannot be represented by the multi-chain network
including organic asphalt (figure 2c) and bone (figure 2f ). Bone, a rigid and lightweight composite,
primarily consists of hydroxyapatite-like mineral particles (e.g. calcium) embedded in a matrix of
collagen fibres. The cortical or compact bone, which is the dense outside layer has noticeable
viscoelasticity as contributed by the collagen fibres and non-fibrous proteins in the bone matrix
[46–48]. Asphalt is heavy organic presented in petroleum and performs as a viscous fluid at high
temperature and viscoelastic solid at normal or low temperature. Asphalt is amorphous with
complex molecular structure and low molecular weight. Organic asphalt cannot be separated
into individual components or narrow fractions [49]. Therefore, its viscoelastic behaviour may
not be properly interpreted by the molecular-chain-based models, which is also true for many
other materials with different morphologies than polymers and rubber-like materials.
 We interpret the physical meaning of the proposed model to represent general materials with
7
various morphologies based on a relatively simple and network-based model. As illustrated
in figure 2e, the model consists of an elastic network with viscous medium filled between the

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
network meshes. The elastic network carries E(t) ∈ [E∞ , E0 ], which transits between the glassy
stage and the rubber stage ( lim E(t) = E∞ and lim E(t) = E0 ). According to the temperature–time
t→∞ t→0
(T − t) superposition principle, the temperature effect is equivalent to the reduced relaxation time
tr (i.e. a higher T corresponds to a higher tr and vice versa). At higher T or tr the model domain
becomes ‘softer’ with lower E(T) value until reaching E∞ at the infinite time (rubbery stage). At
the zero time or infinitely low T, the network carries the highest modulus of E0 (glassy stage). The
β coefficient characterizes the strain hardening or softening behaviour. The α coefficient controls
the slope or rate of E(t) change such that ∂E(t)/∂t ∝ α. A higher α value indicates a higher thermal
sensitivity, resulting in more E(t) reduction (figure 2g). The μ coefficient is related to friction,
and a higher 1/μ value indicates a higher friction force and viscosity leading to a higher E(t)
value (figure 2h). The mathematical model formula (on the non-logarithmical scale) may also be
represented by a three-element of nonlinear spring–dashpot system with the constitutive stress–
strain relation satisfying the Calusius–Duhem inequality [50] for thermodynamic consistency
as proofed in the electronic supplementary material a). By extending the model to a three-
dimensional (3D) domain, the shear and bulk relaxation moduli G(t) and K(t) can be attained
using the same model formula as that of E(t) [51]:

1
log G(t) = log10 G∞ + log10 (G0 − G∞ )(1 + ε)βG (2.8)
A0G + μG eαG log(t/τ0 )

and
1
log (K(t)) = log10 K∞ + log10 (K0 − K∞ )(1 + ε)βK , (2.9)
A0k + μK eαK log(t/τ0 )

where G∞ is shear modulus at t = ∞, and G0 is the instantaneous shear modulus, K∞ is bulk

modulus at t = ∞, K0 is the instantaneous bulk modulus and α G , α K , β G , β K , μG and μK
are model parameters. When compared to the modified sigmoidal-function model in equation
(2.3), the advantages of the proposed model include: (1) it includes a physical mechanism to
interpret material viscoelastic behaviours, (2) we developed a FE scheme to implement the model
for simulating responses based on a continuum-mechanics perspective; and (3) it includes the
nonlinear strain hardening to model broader materials including spider silk. Other referred
models [43,44] were designed primarily for polymers at microscale. In comparison, our model
is designed to model a broad range of general materials for experimental data fit and prediction,
and numerical simulation at macroscale. However, we have not evaluated the model for multiple
relaxation processes.

3. Experimental validation
For some materials like AC the E*(ω) is more often measured than E(t) for practical purpose,
and E(t) can be derived from E*(ω) via the interconversion. However, for other materials like
polymer and biomaterials E(t) are more often measured and then fitted by the PS. In addition,
it is important to predict E(t) values outside of the experimental range to evaluate material
property and simulate responses for a wider range of relaxation time. Therefore, it is useful to
compare the proposed model and the PS in both fitting and predicting the experimental data of
E(t). When part of the experimental data within a middle time range is used as training data to
fit model parameters, the rest of the experimental data at the lower and higher time ranges is
predicted using the fitted model parameters. We used the nonlinear reduced gradient method for
both the proposed model and the PS to determine model parameters as illustrated in electronic
 local oscillation measurement
35 000 n = 1, c2red = 11.14% 8000 8
n = 3, c2red = 1.95%

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

n = 5, c2red = 0.30% 6000

...................................................
30 000 n = 17a
n = 7, c2red = 0.22%
n = 14, 30
n = 14, c2red = 0.18% 4000
25 000 n = 30, c2red = 0.18%
sharp transition
n = 17a, c2red = 0.18% 2000
E (t) (MPa)

20 000 proposed model

n increase
0
15 000 0.1 100 105
loading
aggregate
aluminium cup void asphalt
10 000
transducer 15 cm
specimen
5000 10 cm
steel plate
note: the third transducer
is on the back of the specimen
0
1 × 10–9 0.000001 0.001 1 1000 1 000 000
log time (s)

Figure 3. Laboratory test results andmodel fit of asphalt concrete using the PS with different term numbers versus proposed
model (a using larger seed values of Ei for the PS with n = 17). (Online version in colour.)

supplementary material b). In addition to the physical-ground based evaluation, we used the
reduced χ 2 statistics to quantify the goodness of fit [52] as follows:

2 1 f (x1 , x2 , . . . , xm )
χfit = , (3.1)
L Var
where Var is the variance of measurements Êi (t), and L is degree of freedom as L = m − N − 1
(m is the total number of data points, N is the number of model parameters). A lower χfit 2 value

indicates a higher fitting accuracy. It is known that optimization is generally dependent on seed
values and it may turn out multiple results of model parameters all satisfying the first-order
optimal condition [14]. This is especially true for the model with a relatively large number of
unknown parameters such as the PS.
In figure 3 we evaluate the fitting accuracy of the PS versus the proposed model for
experimental data of E(t) (converted from the measured E*(ω)). E(t) data are fitted by the PS
with variable term n ranging from 1 to 30 (for a total of 3 to 61 model parameters). Results
have shown that with a higher n the PS intends to capture E(t) at a wider range of time more
accurately. However, as n is relatively large (i.e. n ≥ 14) the χRed
2 value has almost no change
without improving fitting accuracy further (e.g. n=14 and 30 attain almost identical E(t) values).
Different seed values may result in different E(t) shapes although with similar χfit 2 value and

fitting accuracy. For example, the fitted E(t) values at the low time range using relatively large

seed values of Ei for n = 17 are significantly higher than that fitted using smaller seed values
for n = 14, but both have almost the same χfit 2 values. Local oscillations of the fitted curves are

also observed due to the finite terms of the PS in discrete spectrums. PS may also produce sharp
transitions to E0 and E∞ especially when n is relatively small (figure 3). With a larger n the E(t)
curve may become smoother when using proper seed values. However, it becomes more difficult
to properly estimate plenty of model parameters without unique solution. In comparison, the
proposed model can achieve a smoother, more accurate and unique curve fitting using much less
model parameters.
In the following we present important tests to validate the model on another three materials:
polymer, agar and bone. For the PS, a term number of n = 14 is used for all materials in the
following sections with satisfied accuracy (the model formula and data fitting/prediction are
 700 20 9
b)
600 15

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
500 prediction 10

E (t) (MPa)
5 gap
400
gap 0
300 5 × 102 5 × 103 5 × 104 5 × 105
measurements
200 Prony series fit
Prony series predict
100 prediction
proposed model fit
proposed model predict fitting
0
10–8 10–6 10–4 10–2 1 102 104 106 108
log time t (s)

Figure 4. Polymer E(t): model fitting and predictions where the proposed model yields higher accuracy and stability than the
PS. (Online version in colour.)

600 220
prediction fitting
200
500 prediction
180
Prony series fit
400 Prony series predict 160
g
gap
E (t) (kPa)

proposed model fit 140

300 proposed model predict
120
measurements 2
Agar cake
200

100

0
0.0001 0.001 0.01 0.1 1 10 100 1000
log time (s)

Figure 5. Data fitting and prediction of E(t) for agar material showing that the proposed model slightly improves accuracy at
the high time range than the PS (measurement data were reproduced with permission from [54]). (Online version in colour.)

illustrated in electronic supplementary material c) Excel spreadsheet). Figure 4 presents the

model fitting and prediction of a polyurethane (PU) polymer material. E(t) was measured at six
temperatures: −30, −22.5, −20, −17.5, −15 and −12.5°C (data from [53]) and ‘shifted’ to that at the
reference temperature of −17.5°C through the temperature–time superposition. In comparison,
for the PS, the fitted E(t) curve has a ‘sudden’ jump at the low time range of 10−5 ≤ t ≤ 10−4 s
outside of the experimental data range (figure 4). It intends to converge to the glassy and rubbery
stages sharply once it is out of the experimental data range. However, it still achieves a high
fitting accuracy (for the experimental data range) with a χfit2 value of only 0.65%. The proposed

model yields higher prediction accuracy though it under-predicts E(t) at the high time range. This
result may imply that the PS has lower stability and higher dependency on seed values than the
proposed model especially for the prediction outside of the experimental data range. Figure 5
presents both the fitted and predicted E(t) of agar materials. The PS and proposed model have
relatively close values, which may be due to the relatively high linear relationship of E(t) versus
(a) 20 (b) 16
local oscillations 10
18 14 prediction
16

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

12

...................................................
prediction
14 prediction
E (t) (GPa)

measurement
prediction 10 sharp
12 Prony series fit transition
measurement
10 Prony series fit 8 Prony series predict gap
8 Prony series predict proposed model fit
proposed model fit 6 proposed model predict
6 proposed model predict
4 sharp 4
transition
2 2
8

01

10

00

00

10 9
10 –

10 1
00

00

00

00
0.

00

10
00

0.
10

00

1×
00
1×

1×
0.

10
log time (s) 10 log time (s)

Figure
 6. Data fitting and prediction of bovine femoral cortical bones: (a) using higher Ei seed values, and (b) using lower
Ei seed values, both showing that the proposed model achieves higher accuracy in fitting E(t) and more stability in prediction
outside of experimental range than the PS (measurement data reproduced with permission from [46]). (Online version in colour.)

Table 1. Model fitting parameters of proposed model.

material AC PU agar bone

E ∞ (MPa) 0.6499 1.26×10−3 1.02×10−3 1000
..........................................................................................................................................................................................................

E 0 (MPa) 6122 701 0.5298 16 514

..........................................................................................................................................................................................................

α 0.3800 0.3999 1.4517 0.5619

..........................................................................................................................................................................................................

μ 0.8191 0.1242 0.5651 0.7729

..........................................................................................................................................................................................................

t for this specific experimental dataset. The PS has slightly overpredicted E(t) at the high time
range. Figure 6 presents fitted and prediction of E(t) of the cortical bone material. We used two

groups of seed values for fitting the PS parameters, and the one with higher seed values ( Ei )
produced less smooth curves in fitting experimental data although with similar fitting accuracy
to that using lower seed values (figure 6b versus 6a) as indicated by their χfit2 value of 0.34 and

0.21%, respectively. The PS significantly overpredicted E(t) at the high time range as it converged
to E∞ sharply outside of the fitting data range. In comparison, the proposed model produced a
smoother and more accurate prediction, and it predicted E0 as 16.5 GPa which falls within the
range of 11–21 GPa as measured by other researchers [55].
Table 1 lists the fitted model parameters for different materials. Bone is the stiffest material
with the highest E0 and E∞ values and poses the longest relaxation time spectrum, while agar
is the softest one with the lowest modulus among these materials. Agar poses a higher α value
than the others, indicating its relatively higher thermal sensitivity for viscosity, followed by the
PU and then the AC, and lastly the bone material which has the elasticity dominant.
The norm-based method was often used to evaluate the accuracy of model prediction, and a
lower norm value indicates higher prediction accuracy [56,57]. We use the L2 norm to evaluate the
model prediction as γ = (Ei − Êi )/EiL2 , where Ei is the predicted modulus and Êi is the measured
one. The L2 values of the PS/proposed models are 0.42/0.24, 4.20/0.89 and 0.69/0.15 for agar, PU
and bone, respectively, indicating higher prediction accuracy of the proposed model with lower
L2 value. The improved prediction accuracy and stability of the proposed model may be explained
as follows. The proposed model captures E(t) in a smooth S curve constrained within the [E∞ , E0 ]
range, with the variation slope and trend controlled by α—the thermal sensitivity parameter and
u—the friction coefficient. As a result, the model predicts E(t) growth outside of the experimental
range naturally to capture the glass transition. In comparison, the PS intends to attain mostly
accurate fitting only within the time range of the available data and it may sharply converge
11
to the more likely false E0 and E∞ values once it is outside the data range (figures 4–6). Given
E0 and E∞ values as constraints (then it is not a true prediction), the PS may achieves higher

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
prediction accuracy, but a sharp converge may still be seen due to its discrete finite set of time
spectrums. Nevertheless, this comparison is fair because both models use the same mathematical
optimization scheme and constraints for determining model parameters. One may argue that
given a very large term number for the PS (e.g. n > 30) the fitting and prediction accuracy may
be improved, and the pre-smooth method may also be used to improve the accuracy of the PS
fitting [29]. However, the proposed model could be superior to the PS for experimental data fitting
and prediction regarding the following aspects: (1) a model with less parameters and satisfied
accuracy is always preferred for simplicity; (2) a large number of model parameters of the PS
produces more instability for data fitting and prediction; (3) the large spring–dashpot systems
of the PS may represent multi-chains of polymer and other macromolecular materials physically,
but not for many others with different morphology.

4. Numerical solution method

(a) Finite-element solution
After the theoretical demonstration and experimental validation, we implemented the proposed
model for numerical simulation under both static and dynamic loadings. The PS has the closed-
form solution for the time integration, while for the proposed model numerical discretization
is necessary for both the time and space domains. Different numerical methods with different
advantage(s) may exist for implementing the same model including the one based on frequency-
domain [58]. The objective of our numerical solution is to achieve competitive computation speed
and numerical stability when compared to the PS. We proposed a FE solution based on time-
domain to integrate E(t) without domain transformation, which is fast and more straightforward
than the frequency-domain method. We proposed a combined time discretization scheme with
the Houbolt, forward and backward finite difference methods for shortening time-step length,
as extended from a previous work for the PS model [59]. We also conducted error analysis to
compare the simulation accuracy of the proposed model versus the PS.
For solids in a continuum state, the strong form of the governing state equation for the dynamic
problem can be formed as follows:

∇ · σ + b = ρ ü(t) on Ω ∈ R3 × [t0 , td ], (4.1)

where σ is stress tensor, u is displacement, b is body force, ρ is material density, Ω ∈ R3 is a 3D

space domain and [t0 , td ] is a time domain. The stress–displacement relationship of the viscoelastic
solid can be expressed as follows [59]:
 
1
t ∂ 1/2[∇u(τ ) + ∇u(τ )T ] − ∇ · u(τ )
3
σ (t) = 2 G(t − τ , ε) dτ
0 ∂τ
t
∂(1/3∇ · u(τ ))
+ 3I K(t − τ , ε) dτ , (4.2)
0 ∂τ
where the first part is the deviatoric stress tensor, and the second term is the hydrostatic stress.
Multiply a test function p(t) on both sides of the strong form (equation (4.1)), and then integrate
with the space and time domains to form a weak form as follows:
 td   td 
(∇ · σ ) · p(t)dΩdt + b · p(t)dΩdt
0 Ω 0 Ω
 td 
− ρ ü(t) · p(t)dΩdt = 0. (4.3)
0 Ω
 According to divergence theory, the divergence of stress can be decomposed into two parts:
12

(∇ · σ ) · p(t)dΩ = (σ · nds) · p(t) − σ : ∇p(t), (4.4)

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
where (σ · nds) is equivalent to the surface traction f (t). Substitute equation (4.4) into equation
(4.3), and the weak form can be re-expressed as follows:
 td   td 
σ : ∇p(t)dΩdt + ρ ü(t) · p(t)dΩdt
0 Ω 0 td
  td 
= f (t) · p(t)dsdt + b · p(t)dΩdt. (4.5)
∂Ω 0 Ω

Substitute equation (4.2) into equation (4.5) and then re-arrange to reach the following weak
form:
 td   t  td 
R(t − τ , ε)u̇(τ )dτ · p(t)dΩdt + ρ ü(t) · p(t)dΩdtt
0 Ω 0 0 Ω
  td 
= f (t) · p(t)dsdt + b · p(t)dΩdt, (4.6)
∂Ω 0 Ω

where

R(t − τ , ε) = R0 (t − τ ) (1 + ε)β , (4.7)

R0 (t − τ ) is a relaxation modulus matrix for linear viscoelasticity in the 3D domain, which is

defined as follows [59]:
t  
1 1
R0 (t − τ ) = 2 [∇ u̇(τ ) + ∇ u̇(τ )T ] − ∇ · u(τ ) dτ
G(t − τ )
0 2 3
t  
1
+ 3I K(t − τ ) ∇ · u̇(τ ) dτ . (4.8)
0 3

u(t) and p(t) are then discretized to that at the FE nodes and presented in a vector format through
the shape function. The discretized p(t) value is arbitrary and can be dismissed on both sides of
the equation. Thus, the following weak form shall satisfy ∀ t ∈ [0, td ]:
 t
R(t − τ , ε)u̇(τ )dτ dΩ + Mü(t) = , (4.9)
Ω 0

where M is mass matrix and is the loading vector such that:

 
= b · p(t)dΩdt + f (t) · p(t)dsdt. (4.10)
Ω ∂Ω

The time domain t ∈ [t0 , td ] is discretized to N time steps for k = 1, 2 . . . N, and for each of them
the sub-time domain τ ∈ [0, t] includes k time steps for j = 1, 2 . . . k. Thus, equation (4.10) can be
rewritten as follows after discretizing τ into k sub-time steps:

k  tj 

R0 (t − τ )(1 + ε)β u̇(τ )dτ + Mü( k) = , (4.11)
tj−1
j=1

where tj and tj−1 are the jth and (j − 1)th time.

The gradient term u̇(τ ) at the sub-time of τ ∈ [tj−1 , tj ] can be discretized u̇(τ ) =
[u(j) − u(j − 1)]/t according to the explicit Euler’s rule. This discretized gradient at the sub-time
can be taken out of the time integration of τ ∈ [tj−1 , tj ], and thus equation (4.11) is re-expressed as
13
follows:
k

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

J(j)(u(j) − u(j − 1))(1 + ε)β + Mü(k) = ,

...................................................
(4.12)
j=1

where u(j) is displacement at the jth sub-time step for j = 1, 2, 3 . . . k, and J(j) is a viscoelastic
stiffness matrix defined as follows:
 tj
J(j) = R0 (t − τ )dτ . (4.13)
tj−1

Poisson’s ratio can be time dependent for different test modalities [60]. For viscoelastic
materials the temperature effect can be converted to a relaxation time through the temperature–
time superposition rule. Given a constant temperature and strain rate within a short loading time
period, it can be reasonable to consider Poisson’s ratio as constant for the simplicity purpose
in numerical implementation. Therefore, G(t) and K(t) values can be directly derived from E(t)
following the relationship among the elastic Young’s, shear and bulk moduli [59]. Therefore, J(t)
can be rederived as follows:
 tj
J(j) = Ce E(t − τ ) dτ , (4.14)
tj−1

where Ce is a discretized elastic matrix at FE nodes of the fourth-order elasticity tensor. For the
proposed model, substitute equation (2.4) into equation (4.14), the viscoelastic stiffness matrix
JP ( j) can be calculated as follows:
  tj 
1
JP (j) = Ce E∞ + (E0 − E∞ ) αlog(t−τ )
dτ dΩ. (4.15)
Ω tj−1 1 + μe

Following the trapezoidal rule, JP (j) can be discretized as:


(E0 − E∞ ) 1 1
JP (j) = Ce E∞ + α log(t−t )
+ α
t, (4.16)
2 1 + μe j 1 + μe log(t−tj−1 )
where t = tj − tj−1 is the time-step length. While for the PS, substitute equation (1.1) into
equation (4.14), J(j) can be attained as follows:


N
JPS (j) = Ce E∞ t + ηi [e−(Ei /ηi )(t−tj ) − e−(Ei /ηi )(t−tj−1 ) ] . (4.17)
i

The Houbolt method is adopted for time discretization of acceleration due to its low time-step
dependency and high stability [61]:
[2u(k) − 5u(k − 1) + 4u(k − 2) − u(k − 3)]
ük = , (4.18)
t2
where u(k) is displacement at the kth time step for k = 1, 2, 3 . . . n. A relatively great time-step
length (i.e. 0.001 second) was used to improve computation speed satisfying numerical accuracy.
For solving the nonlinear viscoelastic displacement, substitute equation (4.18) into equation (4.12)
to attain the discretized:

k−1
J(k)[u(k) − u(k − 1)](1 + ε (k))β+1 + J(j)[u(j) − u(j − 1)](1 + ε (j))β
j=1

M
+ [2u(k) − 5u(k − 1) + 4u(k − 2)−u(k − 3)] − = 0. (4.19)
t2
We employed the Cholesky factorization with the iterative refinement method to solve the
global linear system. At each iteration step Newton’s method can be applied to solve u(k) by
calculating the tangent of the function ∂f /∂u(k) [62]. At t = 0 u(0) = 0 and ü(0) = 0. The acceleration
14
at time zero is discretized as follows:
∂ 2u [u(1) − 2u(0) + u(−1)]

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
= = 0. (4.20)
∂t2 |t=0 t2
At t = 0 the governing state equation satisfies:

J(0)[u(0) − u(−1)](1 + ε(0))β /t + Mü(0) = (0). (4.21)

Given ε (0) = 0, u( − 1) and u(1) can be solved from equation (4.20) and (4.21), then u(2) is solved
following the Houbolt time discretization, and so on for u(3), u(4) . . . u(N). With displacements
calculated at all FE nodes, strain can be solved as (∇u + ∇uT )/2 through the strain–displacement
matrix. Consequently, stress is solved as follows:
t
∂ε
σ (t) = R0 (t − τ )(1 + ε)β dε. (4.22)
0 ∂t
We developed the VBA/FORTRAN coding to implement the numerical solution (see computer
code in electronic supplementary material d) for one-dimensional stress solution and the
subroutine of stiffness matrix for two-dimensional solution).

(b) Error analysis

In this section, we analyse the simulation error for the proposed model versus the PS, in which
strain hardening or softening is not considered. The major numerical error of the viscoelastic
solution lies in the computation of the viscoelastic stiffness matrix J(j). Using the trapezoidal rule
[63] the truncation error of the proposed model for the discretization of ∫tt21 J(x)dt can be calculated
as follows:
(t2 − t1 )3
ErrP = J (ξ ) ξ ∈ [t1 , t2 ], (4.23)
12N2
where N is the number of time grids. Substitute equation (4.16) into equation (4.23):

t3 ∂ 2 E(t − ξ )
ErrP = − Ce . (4.24)
12 ∂ξ 2
The first derivative of E(t − ξ ) with respect to ξ can be derived as:

∂E(t − ξ ) γ (t − ξ )γ −1
= ln(10)A(E)E(t − ξ ) , (4.25)
∂ξ (1 + A(t − ξ )γ )2
α
where A = μα , and γ = − . Then the second-order derivative can be derived as:
ln10
∂ 2 E(t − ξ ) ∂E(t − ξ ) γ (t − ξ )γ −1
= ln(10)AE
∂ξ 2 ∂ξ [1 + A(t − ξ )γ ]2
 
−γ (γ − 1)(t − ξ )γ −2 2Aγ 2 (t − ξ )2(γ −1)
+ E(t − ξ ) + . (4.26)
[1 + A(t − ξ )γ ]2 [1 + A(t − ξ )γ ]3
Substitute equation (4.26) into equation (4.24), ∀ ξ ∈ [tj−1 , tj ]:

−ln(10)AEt3 ∂E(t − ξ ) γ (t − ξ )γ −1
ErrP (t) = Ce
12 ∂ξ [1 + A(t − ξ )γ ]2
 
−γ (γ − 1)(t − ξ )γ −2 2Aγ 2 ξ 2(γ −1)
+ E(t − ξ ) + . (4.27)
[1 + A(t − ξ )γ ]2 [1 + A(t − ξ )γ ]3
In comparison the PS’s J(j) has a closed-form solution without needing time discretization.
However, the PS produces computational errors due to its less accurate data fitting, for which
 10
15

J (t) and error (kPa s Ce)

1

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

ΩPS J (t) errorΩ

...................................................
0.1
proposed model J (t) error
0.01
experimental
0.001 range

0.0001 prediction range

0.00001
0 1 2 3 4 5 6 7 8 9 10
time (s)

Figure 7. Errors of viscoelastic stiffness matrix of agar material where the proposed model yields lower errors than the PS (the
fluctuation of the PS’s J(t) error is due to the variation of fitted E(t) values at different time ranges.). (Online version in colour.)

the computation error can be calculated as follows:

 tj
ErrPS (t) = Ce (Eps (t − τ ) − E(t − τ ))dτ , (4.28)
tj−1

where Eps is the PS fitted modulus, E is the true value and Ce is elastic tensor.
Here we present the error analysis results of agar material as a numerical example because the
E(t) values of agar fitted by the PS are closer to that of the proposed model than other materials
(figure 5). Figure 7 plots the J(t) errors of the proposed model versus that of the PS during 10 s.
Results indicate that the J(t) errors of the proposed model are negligible (i.e. average 0.016% and
max 0.46%) and smaller than that of the PS.

5. Numerical analysis and implementation

With the numerical algorithm implemented and laboratory-measured material property as
inputs, we simulate the displacement, stress–strain relationship and hysteresis considering strain
hardening, and energy dissipation of a variety of materials. We evaluate the computation
accuracy, speed and stability for convergence of the proposed model versus the PS.

(a) Creep and cyclic deformations

In order to validate the accuracy of the developed FE algorithm, we first compared its simulation
results with that of ANSYS—one of the most powerful commercial FE software using the PS
model. They attain almost identical results although they used different algorithms (e.g. different
time discretization methods). We simulated a unit agar body as illustration to evaluate the
numerical stability and convergence. Figure 8a presents the simulated deformations under a
constant Heaviside loading (1 kPa) for t ∈ [0, 10 s]. Note that the initial deformation at zero time
is the instantaneous elastic responses. Different time-step lengths dt ∈ [0.01, 1 s] are used, and
both models converge at dt = 0.01 s although the PS is slightly less sensitive to dt. Figure 8b
presents simulated deformations under the sinusoidal loading f (t) = 0.5sin(π t − π /2) + 0.5 with
an amplitude of 1 kPa. Again simulation results show that both models converge at the same
dt = 0.1 s. The constant loading requires a shorter dt for convergence than the cyclic loading due
to the resulting larger creep deformation. Figure 8c compares the simulated deformations of the
two models with dt = 0.01 s. The difference is as high as 4.2 and 8.1% for the constant Heaviside
loading and sinusoidal loading, respectively. This discrepancy is primarily introduced by the PS’s
less accurate E(t) model fitting on experimental data.
(a) proposed model Prony series 16
(i) 0.009 (ii)
0.008

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
deformation (mm)

0.007
0.006 dt = 0.01 s
0.005 dt = 0.10 s
0.004 dt = 0.01 s dt = 0.20 s
dt = 0.50 s
0.003 dt = 0.10 s
dt = 0.20 s dt = 1.00 s
0.002
dt = 0.50 s
0.001 dt = 1.00 s
0
0 1 2 3 4 5 6 7 8 9 10 0 2 4 6 8 10
(b)
(i) 0.006 1.2 (ii) 0.006
time
0.005 retardation
1.0 0.005
deformation (mm)

deformation (mm)
0.004 0.8 0.004
0.003 0.6 0.003
dt = 0.01 s
0.002 dt = 0.01 s 0.4 0.002 dt = 0.10 s
dt = 0.10 s
dt = 0.20 s dt = 0.20 s
0.001 dt = 0.50 s 0.2 0.001 dt = 0.50 s
dt = 1.00 s
loading dt = 1.00 s
0 0 0
0 1 2 3 4 5 6 7 8 9 10 0 2 4 6 8 10
time (s) time (s)

(c) 0.009
proposed model sinusoidal load
0.008 PS sinusoidal loading
proposed model constant load –4.2%
0.007
deformation (mm)

PS constant load
0.006 ANSYS sinusoidal load

0.005
0.004
0.003
0.002
0.001 –8.1%
0
0 1 2 3 4 5 6 7 8 9 10
time (s)

Figure 8. Simulated viscoelastic deformations: (a) under constant loading, (i) results of the proposed model and (ii) results of
the PS both converged at dt = 0.01 s; (b) under sinusoidal loading, (i) results of the proposed model and (ii) results of the PS both
converged at dt = 0.1 s, showing time retardation of displacement to loading; and (c) deformation comparison of the proposed
model to that of the PS with dt = 0.01 s showing notable difference at some time points (note: the additional simulation results
of the ‘ANSYS sinusoidal load’ is to show that the proposed model attains (almost) identical results as ANSYS software). (Online
version in colour.)

(b) Stress–strain relationship of materials

We simulated stress–strain curves of AC, PU polymer, agar and bone as plotted in
figure 9a–d, clearly showing material viscoelastic behaviours. Note that different strain ranges are
used for different materials due to their different viscoelastic ranges [64–67]. We also simulated
the nonlinear stress–strain relationship of different materials. In the laboratory test, a spider silk
thread was glued to a glass substrate with the glycoprotein glue (figure 9e) and was subjected
to a stretching loading (data reproduced from [68] with permission). Results have shown that the
model is able to capture the nonlinear strain hardening behaviour of spider silk (figure 9e), and the
nonlinear strain softening and then hardening of the polyampholyte-formed hydrogel material
(a) (b) (c) (d )
0.40 a) 9
17
1800 180
0.35 8 1600 160
0.30 7 1400 140
AC PU agar bone

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

stress (MPa)

6 1200 120

...................................................
0.25
5 1000 100
0.20
4 800 80
0.15 3 600 60
0.10 2 400 40
0.05 1 200 20
0 0 0 0
0 0.0005 0.0010 0.0015 0.0020 0.0025 0 0.02 0.04 0.06 0.08 0.10 0.12 0 0.2 0.4 0.6 0 0.01 0.02 0.03 0.04 0.05 0.06
strain strain strain strain

(e) 700
( f ) 1.0
600 0.9
strain hardening
0.8 strain hardening
coat simulation
500 simulation
experimental data 0.7 experimental data
stress (MPa)

400 0.6
spider axial strain rate = 10 mm min–1
thread, 20 mm 0.5
300 E• = 89.1 MPa
stretching 0.4 E• = 2.14 Pa
E0 = 2818 MPa E0 E0 = 1.995 MPa
200 0.3
m = 1.2 m = 0.01052
glycoprotein a = 0.5 0.2 a = 1.19
100 b = 1.43 b = 0.092
0.1 softenning
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 2 4 6
strain strain

Figure 9. Simulated stress–strain relationships: (a) asphalt concrete; (b) PU polymer; (c) agar; and (d) bone; (e) spider web silk
laboratory test set up and simulation results versus measurements with strain hardening; and (f ) hydrogel showing nonlinear
strain softening and then hardening as simulated by the new model. (Online version in colour.)

(figure 9f, laboratory data reproduced from [68] with permission). This hydrogel performed
viscoelastic behaviour with stress recovery for strains up to 700% [68]. These materials may
perform plastic deformation and fracture at larger strain levels which are studied by our model.
Finally, we extended the model to simulate the dynamic responses of a multilayered pavement
structure as illustrated in the Supplementary material e).

6. Conclusion
We proposed a mathematical model for the data fitting and prediction of E(t) and numerical
solution of viscoelastic responses for a broad range of materials. We interpret its physical
mechanism as an elastic network–viscous medium system with five model parameters to
represent general materials considering nonlinear strain hardening. We then developed a
Galerkin FE method and robust numerical algorithm to implement the model for simulating
dynamic responses. We have validated the model on both experimental data and numerical
simulation for a variety of materials including asphalt concrete, polymer, spider silk, agar and
bone. The model has shown some advantages when compared to the PS—the widely used model
including: (a) it improves the fitting and prediction accuracy for experimental data while using
much less model parameters; (b) its numerical solution has competitive computation speed and
numerical stability for convergence; and (c) it has considered the nonlinear strain hardening
behaviour of some special materials. The model is able to simulate the creep and sinusoidal
responses reasonably under both static and dynamic loadings. Therefore, the model may serve as
an alternative to simulate the viscoelastic behaviours of solids at both experimental and numerical
scales with improving accuracy while reducing complexity.
Data accessibility. Part of the data along with the model formula is presented in the electronic supplementary
materials c and d, and the other data can be attained by replacing the model parameters.
Authors’ contributions. Q.X. developed the model, performed analysis and wrote the paper. B.E. advised the model
development and reviewed the paper with revision advice.
Competing interests. We have no competing interests.
Funding. Part of the data collection was supported by the Federal Highway Administration project DTFH61-
07-C-0032.
Acknowledgements. We appreciate the courtesy of those researchers or publishers [46,53,55,63,68] for offering
part of experimental data.
References 18
1. Flory P. 1953 Principles of polymer chemistry. Ithaca, NY: Cornell University Press.

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

2. McCrum N, Williams B, Read G. 1991 An elastic and dielectric effects in polymer solids. New York,

...................................................
NY: Dover.
3. Likhtman AE, Sukumaran SK, Ramirez J. 2007 Linear viscoelasticity from molecular dynamics
simulation of entangled polymers. Macromolecules 40, 6748–6757. (doi:10.1021/ma070843b)
4. Kremer K, Grest GSJ. 1990 Dynamics of entangled linear polymer melts: a molecular-
dynamics simulation. J. Chem. Phys. 92, 5057–5086. (doi:10.1063/1.458541)
5. de Gennes PG. 1971 Reptation of a polymer chain in presence of fixed obstacles. J. Chem. Phys.
55, 572. (doi:10.1063/1.1675789)
6. McLeish TCB. 2010 Tube theory of entangled polymer dynamics. Adv. Phys. 51, 1379–1527.
7. Pokrovskii VN. 2008 Reptation and diffusive modes of motion of linear macromolecules.
J. Exp. Theor. Phys. 106, 604–607. (doi:10.1134/S1063776108030205)
8. de Gennes PG. 1979 Brownian motions of flexible polymer chains. Nature 282, 367–370.
(doi:10.1038/282367a0)
9. Roylance D. 2001 Engineering viscoelasticity. Cambridge, MA: Massachusetts Institute of
Technology.
10. Chen X, Ashcroft IA, Wildman RD, Tuck CJ. 2015 An inverse method for determining the
spatially resolved properties of viscoelastic–viscoplastic three-dimensional printed materials.
Proc. R. Soc. A 471, 20150477. (doi:10.1098/rspa.2015.0477)
11. Xu Q, Zhou Q, Medina C, Chang GK, Rozycki DK. 2009 Experimental and numerical analysis
of a waterproofing adhesive layer used on concrete-bridge decks. Int. J. Adhes. Adhes. 29, 525–
534. (doi:10.1016/j.ijadhadh.2008.12.001)
12. Koontz E, Blouin V, Wachtel P, Musgraves JD, Richardson K. 2012 Prony series spectra of
structural relaxation in N-BK7 for finite element modeling. J. Phys. Chem. A 116, 12 198–12 205.
(doi:10.1021/jp307717q)
13. Jänicke R, Larsson F, Runesson K, Steeb H. 2016 Numerical identification of a viscoelastic
substitute model for heterogeneous poroelastic media by a reduced order homogenization
approach. Comput. Methods Appl. Mechan. Eng. 298, 108–120. (doi:10.1016/j.cma.2015.09.024)
14. Xu Q. 2014 Development of a computational method for inverting dynamic moduli of
multilayer systems with applications to flexible pavements. Dissertation, University of Texas
at Austin, Austin, TX.
15. Provenzano PP, Lakes RS, Corr DT, Vanderby Jr R. 2002 Application of nonlinear
viscoelastic models to describe ligament behavior. Biomech. Model. Mechanobiol. 1, 45–57.
(doi:10.1007/s10237-002-0004-1)
16. Xu F, Lu TJ, Seffen KA. 2008 Biothermomechanics of skin tissues. J. Mech. Phys. Solids 56,
1852–1884. (doi:10.1016/j.jmps.2007.11.011)
17. Slanik ML, Nemes JA, Potvin MJ, Piedboeuf JC. 2000 Time domain finite element simulations
of damped multilayered beams using a PS representation. Mechan. Time-Depend. Mater. 4,
211–230. (doi:10.1023/A:1009826923983)
18. Huang X, Zhou S, Sun G, Li G, Xie Y. 2015 Topology optimization for microstructures
of viscoelastic composite materials. Comput, Methods Appl. Mechan. Engin. 283, 503–516.
(doi:10.1016/j.cma.2014.10.007)
19. Iyo T, Maki Y, Sasaki N, Nakata M. 2004 Anisotropic viscoelastic properties of cortical bone.
J. Biomech. 37, 1433–1437. (doi:10.1016/j.jbiomech.2003.12.023)
20. Bagley RL, Torvik PJ. 1983 A theoretical basis for the application of fractional calculus to
viscoelasticity. J. Rheol. 27, 3, 201–210. (doi:10.1122/1.549724)
21. Pritz T. 2003 Five-parameter fractional derivative model for polymeric damping materials.
J. Sound Vib. 265, 935–952. (doi:10.1016/S0022-460X(02)01530-4)
22. Pronk AC. 2003 Revival of the Huet-Sayegh response model-notes on the Huet & Sayegh
rheological model. DWW-2003-29.
23. Xu Q, Solaimanian M. 2009 Modeling linear viscoelastic properties of asphalt concrete by the
Huet–Sayegh model. Int. J. Pavement Eng. 10, 401–422. (doi:10.1080/10298430802524784)
24. Kim S. 2011 Viscoelastic behaviors in polymeric nanodroplet collisions. Phys. Rev. E 83, 041302.
(doi:10.1103/PhysRevE.83.041302)
25. Schapery RA. 1969 On the characterization of non-linear viscoelastic materials. Polym. Eng.
Sci. 9, 295–310. (doi:10.1002/pen.760090410)
26. Monsia MD. 2011 A simplified nonlinear generalized Maxwell model for predicting
19
the time dependent behavior of viscoelastic materials. World J. Mech. 1, 158–167.
(doi:10.4236/wjm.2011.13021)

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
27. Chung CW, Buist ML. 2012 A novel nonlinear viscoelastic solid model. Nonlinear Analysis:
Real World Applications. 13, 1480–1488. (doi:10.1016/j.nonrwa.2011.11.011)
28. Haj-Ali RM, Muliana AH. 2004 Numerical finite element formulation of the Schapery non-
linear viscoelastic material model. Int. J. Numer. Methods Eng. 59, 25–45. (doi:10.1002/nme.861)
29. Park SW, Kim YR. 2001 Fitting Prony-series viscoelastic models with power-law
presmoothing. J. Mater. Civ. Eng. 13, 26–32. (doi:10.1061/(ASCE)0899-1561(2001)13:1(26))
30. Al-Qadi I, Yoo P, Elseifi M, Nelson S. 2009 Creep behavior of hot-mix asphalt due
to heavy vehicular tire loading. J. Eng. Mech. 135, 1265–1273. (doi:10.1061/(ASCE)0733-
9399(2009)135:11(1265))
31. Xu Q, Prozzi JA. 2014 Static versus viscoelastic wave propagation approach for
simulating loading effects on pavement structures. Constr. Build. Mater. 53, 584–595.
(doi:10.1016/j.conbuildmat.2013.12.017)
32. Koval’chenko MS. 2000 Dynamics of uniaxial tension of a viscoelastic strain-hardening body
in a system with one degree of freedom. Part 3. Hydraulic loading of the system. Strength
Mater. 32, 27–40. (doi:10.1007/BF02511505)
33. Xu Q, Engquist B. 2016 A mathematical and physical model improves accuracy in simulating
solid material relaxation modulus and viscoelastic responses. (https://arxiv.org/abs/1412.
5225).
34. Nutting PG. 1943 A general stress-strain-time formula. J. Franklin Inst. 235, 513–524.
(doi:10.1016/S0016-0032(43)91483-8)
35. Metzler R, Klafter J. 2002 From stretched exponential to inverse power-law: fractional
dynamics, Cole–Cole relaxation processes, and beyond. J. Non-Cryst. Solids 305, 81–87.
(doi:10.1016/S0022-3093(02)01124-9)
36. National Cooperative Highway Research Program (NCHRP). 2002 Design Guide Draft – 2.4
Modulus of Elasticity for Major Material Groups.
37. Pellinen TK, Witczak MW. 2002 Stress dependent master curve construction for dynamic
(complex) modulus. J. Assoc. Asphalt Paving Technol. 71, 281–309.
38. Sahni V, Blackledge TA, Dhinojwala A. 2010 Viscoelastic solids explain spider web stickiness.
Nat. Commun. 1, 19. (doi:10.1038/ncomms1019)
39. Abboud A, Nassar SA. 2013 Viscoelastic strain hardening model for gasket creep relaxation.
J. Pressure Vessel Technol. 135, 1–9. (doi:10.1115/1.4023501)
40. Lane JMD. 2015 Cooling rate and stress relaxation in silica melts and glasses via microsecond
molecular dynamics. Phys. Rev. E 92, 012320. (doi:10.1103/PhysRevE.92.012320)
41. Williams ML, Landel RF, Ferry JD. 1955 Temperature dependency of relaxation mechanisms
in amorphous polymers and other glass forming liquids. J. Am. Chem. Soc. 77, 3701–3707.
(doi:10.1021/ja01619a008)
42. Li R. 2000 Time-temperature super-position method for glass transition temperature of plastic
materials. Mater. Sci. Eng. A 278, 36–45. (doi:10.1016/S0921-5093(99)00602-4)
43. Gotlib YY, Gurtovenko AA. 1997 The model theory of viscoelastic relaxation properties of
bulk cross-linked polymers. Interchain friction effects. Macromol. Theory Simul. 6, 523–551.
(doi:10.1002/mats.1997.040060214)
44. Miehe C, Goktepe S. 2005 A micro–macro approach to rubber-like materials. Part II: The
micro-sphere model of finite rubber viscoelasticity. J. Mech. Phys. Solids 53, 2231–2258.
(doi:10.1016/j.jmps.2005.04.006)
45. Römer L, Scheibe T. 2008 The elaborate structure of spider silk. Prion 2, 154–161.
(doi:10.4161/pri.2.4.7490)
46. Sasaki N. 2012 Viscoelastic properties of biological materials. In Viscoelasticity - From theory to
biological applications (ed. JD Vincente), pp. 99–122. Intech.
47. Sasaki N, Yoshikawa M. 1993 Stress relaxation in native and EDTA-treated bone as a function
of mineral content. J. Biomech. 26, 77–83. (doi:10.1016/0021-9290(93)90615-L)
48. Iyo T, Sasaki N, Maki Y, Nakata M. 2006 Mathematical description of stress relaxation of
bovine femoral cortical bone. Biorheology 43, 117–132.
49. Mansoori GA. 2010 Heavy Organics Depositions from Petroleum Fluids, University of Illinois
at Chicago. See http://tigger.uic.edu/~mansoori/HOD_html.
50. Haupt P. 1993 Non-equilibrium thermodynamics with applications to solids. In CISM-Course
336 (ed. W. Muschik), pp. 138–140. New York, NY: Springer.
51. Dassault Systemes. 2004 ABAQUS Analysis user’s manual – Volume III: materials, version
20
6.5.
52. Temme K, Verstraete F. 2015 Quantum chi-squared and goodness of fit testing. J. Math. Phys.

rspa.royalsocietypublishing.org Proc. R. Soc. A 474: 20170540

...................................................
56, 012202. (doi:10.1063/1.4905843)
53. Knauss WG, Emri I, Lu H. 2008 Mechanics of polymers: viscoelasticity. In Springer handbook of
experimental solid mechanics (ed. WN Sharpe), pp. 49–96. Springer.
54. Shen J, Cheng P, Gu W, Hao Z. 2013 Stress relaxation measurement of viscoelastic
materials using a polymer-based microfluidic device. Sens. Actuators, A 203, 119–130.
(doi:10.1016/j.sna.2013.08.038)
55. University of Cambridge. 2013 Mechanical properties of bone. Dissemination of IT for
the promotion of materials science. See http://www.doitpoms.ac.uk/tlplib/bones/bone_
mechanical.php.
56. Xu Q, Zhu H, Prozzi JA. 2016 An inverse model and mathematical solution for inferring
viscoelastic properties and dynamic deformation of heterogeneous structures. Comput.
Methods Appl. Mech. Eng. 300, 798–833. (doi:10.1016/j.cma.2015.11.012)
57. Petra N, Zhu H, Stadler G, Hughes TJR, Ghattas O. 2012 An inexact Gauss–Newton method
for inversion of basal sliding and rheology parameters in a nonlinear Stokes ice sheet model.
J. Glaciol. 58, 889–903. (doi:10.3189/2012JoG11J182)
58. Sun L, Pan Y, Gu W. 2013 High-order thin layer method for viscoelastic wave propagation in
stratified media. Comput. Methods Appl. Mech. Eng. 257, 65–76. (doi:10.1016/j.cma.2013.01.004)
59. Xu Q, Prozzi J. 2015 A time domain finite element method for dynamic viscoelastic solution
of layered half space. Comput. Struct. 260, 20–39. (doi:10.1016/j.compstruc.2015.07.005)
60. Lakes RS, Wineman A. 2006 On Poisson’s ratio in linearly viscoelastic solids. J. Elasticity 85,
45–63. (doi:10.1007/s10659-006-9070-4)
61. Chung J, Hulbert GM. 1994 A family of single-step Houbolt time integration algorithms
for structural dynamics. Comput. Methods Appl. Mech. Eng. 118, 1–11. (doi:10.1016/
0045-7825(94)90103-1)
62. Oden JT. 2006 Finite elements of nonlinear continua. Mineola, NY: Dover Publications.
63. Atkinson KE. 1989 An introduction to numerical analysis. New York, NY: John Wiley & Sons.
64. You T, Ai-Rub RKA, Masad E, Little D. 2014 Three-dimensional microstructural
modeling framework for dense-graded asphalt concrete using a coupled
viscoelastic, viscoplastic, and viscodamage model. J. Mater. Civ. Eng. 26, 607–621.
(doi:10.1061/(ASCE)MT.1943-5533.0000860)
65. Dutta J, Naskar K. 2014 Investigation of morphology, mechanical, dynamic mechanical
and thermal behaviour of blends based on ethylene vinyl acetate (EVA) and thermoplastic
polyurethane (TPU). RSC Adv. 4, 60 831–60 841. (doi:10.1039/C4RA07823C)
66. Bhat S, Tripathi A, Kumar A. 2011 Supermacroprous chitosan–agarose–gelatin cryogels: in
vitro characterization and in vivo assessment for cartilage tissue engineering. J. R Soc. Interface
8, 540–554. (doi:10.1098/rsif.2010.0455)
67. Thurner PJ et al. 2007 High-speed photography of compressed human trabecular
bone correlates whitening to microscopic damage. Eng. Fract. Mech. 74, 1928–1941.
(doi:10.1016/j.engfracmech.2006.05.024)
68. Sun TL, Kurokawa T, Kuroda S, Ihsan AB, Akasaki T, Sato K, Haque MA, Nakajima T, Gong
JP. 2013 Physical hydrogels composed of polyampholytes demonstrate high toughness and
viscoelasticity. Nat. Mater. 12, 932–937. (doi:10.1038/nmat3713)