Machine-Learning-Enabled Comparative Modelling of the Creep Behaviour of Unreinforced PBT and Short-Fibre Reinforced PBT Using Prony and Fractional Derivative Models

Abstract

This study presents an approach based on data-driven methods for determining the parameters needed to model time-dependent material behaviour. The time-dependent behaviour of the thermoplastic polymer polybutylene terephthalate is investigated. The material was examined under two conditions, one with and one without the inclusion of reinforcing short fibres. Two modelling approaches are proposed to represent the time-dependent response. The first approach is the generalised Maxwell model formulated through the classical exponential Prony series, and the second approach is a model based on fractional calculus. In order to quantify the comparative capabilities of both models, experimental data from tensile creep tests on fibre-reinforced polybutylene terephthalate and unreinforced polybutylene terephthalate specimens are analysed. A central contribution of this work is the implementation of a machine-learning-ready parameter identification framework that enables the automated extraction of model parameters directly from time-series data. This framework enables the robust fitting of the Prony-based model, which requires multiple characteristic times and stiffness parameters, as well as the fractional model, which achieves high accuracy with significantly fewer parameters. The fractional model benefits from a novel neural solver for fractional differential equations, which not only reduces computational complexity but also permits the interpretation of the fractional order and stiffness coefficient in terms of physical creep resistance. The methodological framework is validated through a comparative assessment of predictive performance, parameter cheapness, and interpretability of each model, thereby providing a comprehensive understanding of their applicability to long-term material behaviour modelling in polymer-based composite materials.

1. Introduction

Short-fibre-reinforced composites (SFRCs) are widely used in various engineering fields like the aerospace and automotive industries. Their mechanical properties, including high stiffness and strength, together with their suitability for cheap, large-scale automated production recently increased the interest in SFRC [1,2,3,4,5,6]. SFRC materials consist of two or more different materials that are combined during a mould injection process [7,8]. However, due to mould injection, the fibres are spatially distributed within the component [8]. This results in heterogeneity of mechanical properties at the component level [9], which is derived from its heterogeneous microstructure [10,11]. The probabilistic characteristics of this microstructure are characterised by probability density functions like fibre length, fibre diameter, and fibre orientation [11,12].

While thermoplastic polybutylene terephthalate (PBT) serves as a matrix material, the short glass fibres reinforce the structure, leading to increased stiffness and strength. However, polymer-based composites like PBT GF30 (with 30% fibre mass fraction) exhibit time-dependent deformation under sustained loads, known as viscoelasticity or creep. Typically, creep is caused by the matrix material [13,14,15]. To predict the long-term performance of components, it is important to accurately characterise the creep behaviour, especially for the automotive, aerospace, and electronics industries [1,2,3,4,5].

Viscoelastic behaviour of SFRC is typically described using the generalised Maxwell model [16], which combines springs and dampers to represent elastic and viscous elements [16,17]. The Maxwell model is standardised and widely used in commercial Finite Element Method (FEM) software packages [18,19]. Its implementation in FEM relies on the Prony series [16,20,21], where the Prony series reflects the solution of an Ordinary Differential Equation (ODE), which results from the combination of springs and dampers. As these elements are typically considered to be linear, the ODE solution can be modelled as a sum of exponential functions. The effectiveness of this approach requires a large number of parameters to capture the full spectrum of material behaviour on varying time scales. This increases the complexity of parameter identification, as well as computational cost [20,21,22].

While Prony-based models approximate the non-local nature of creep effects by the superposition of several spring and damper elements, fractional derivative models can inherently model this behaviour. A fractional derivate model, also referred as a fractional element, describes dynamics that lie between a spring and a damper [16,23,24]. However, the practical application of fractional models remains challenging: normally, these models are solved indirectly by transforming them into the Fourier domain to fit experimental data. When established, this approach requires more complex data collection setups as the material behaviour has to be measured across different frequencies in the frequency domain. This makes simple tensile tests insufficient for data collection. Furthermore, Fourier analysis is limited to linear models, which complicates the representation of non-linear behaviour [25,26].

Recent advances in machine learning (ML) have enabled us to learn fractional dynamics directly from data. Coelho et al. [27] introduced neural fractional differential equations (NFDEs). NFDEs combine a solver for Fractional Dynamical Equations (FDEs) with neural nets (NNs) to learn the underlying FDE from arbitrary data. Building on this foundation, Zimmering et al. [28] improved the solver in accuracy and resource efficiency, enabling the use of the solver in real-world scenarios.

In this study, the need for special experiments that capture the frequency spectrum of SFRCs is overcome by fitting a parametric fractional element model directly to experimental tensile test data. The optimised NFDE solver of Zimmering et al. is employed to perform this fitting. The solver’s general capability to fit arbitrary dynamical models (including NNs) additionally offers flexibility in capturing non-linear creep behaviour from arbitrary data in future studies.

Following this, the contributions of this paper are as follows.

• First, the accuracy of the fractional element in describing the creep behaviour of PBT GF0 and PBT GF30 is compared with the generalised Maxwell model. This comparison aims to determine the extent to which fractional derivative models can effectively capture the viscoelastic response of such materials over time.
• Second, practical methodologies are studied to efficiently learn the parameters of fractional element models. It is sought to identify computational techniques that enable a stable parameter identification process while maintaining accuracy.
• Third, the advantages that fractional derivative models may offer over the Prony models are discussed for SFRCs. It is evaluated if fractional models provide improved predictive capabilities and enhance the representation and interpretability of material properties over extended time scales.

This paper is structured as follows: Section 2 outlines the theoretical framework for modelling creep behaviour as well as the data fitting methods for the Prony fractional model. Section 3 details the experimental setup for the data acquisition methods. In Section 4, the results are presented. Here, it is focused on the determination of Prony coefficients and coefficients derived from fractional models. Finally, in Section 5, the comparative performance of the models is discussed, and conclusions about their practical applications in material characterisation are drawn.

The developed software is based on Python 3.11 and open-source software libraries and is available along with the experimental data at https://github.com/zimmer-ing/PBTGF-Creep. This allows others to reproduce the results and apply the presented methods in their own work.

2. Materials and Methods

2.1. Theoretical Framework for Creep Behaviour

Creep is a time-dependent deformation process that occurs under constant load and temperature, involving both elastic and viscous contributions. This phenomenon is particularly relevant in polymer-based materials, which exhibit viscoelastic behaviour, which combines time-independent elasticity and time-dependent viscous flow.

2.1.1. Viscoelastic Effect

Creep occurs when a material undergoes time-dependent deformation under constant load and temperature. This effect occurs in polymer-based materials, which have both elastic and viscous behaviours, commonly referred to as viscoelastic behaviours [16,21,29,30]. In general, creep deformation typically occurs in three stages, primary, secondary, and tertiary creep.

The total strain ${\epsilon }_{\mathrm{total}}\left(t\right)$ over time of a sample can be expressed additively as

$$\epsilon \left(t\right)={\epsilon }_{\mathrm{total}}\left(t\right)={\epsilon }_{\mathrm{linear} \mathrm{elastic}}+{\epsilon }_{\mathrm{creep}}\left(t\right).$$

(1)

Here, ${\epsilon }_{\mathrm{linear} \mathrm{elastic}}$ represents the immediate strain response upon the application of stress, following Hooke’s law for linear elasticity. This component remains constant over time and does not contribute to the time-dependent deformation behaviour of the material. However, the creep deformation ${\epsilon }_{\mathrm{creep}}\left(t\right)$ accounts for the viscoelastic deformation. In this work, the focus is on the primary and secondary stages because capturing the tertiary stage is time-consuming and experimentally challenging.

The mathematical models are derived using rheological elements that describe the stress–strain relations for the linear viscoelastic behaviour [16,31]. These elements are shown in Figure 1, where the spring element represents Hooke’s law with stiffness $K_{eq}$. A dashpot represents a purely viscous material according to Newton’s law, characterised by dynamic viscosity $\eta$. The fractional element is described by the parameters C and $\alpha$.

Generally, there are two equivalent approaches to characterise the mathematical relationships between stress and deformation in linear viscoelastic materials [16,21,29,30,32]. One approach employs integral formulations to establish these relationships, while the other uses first-order ODE to connect stresses and strains. The integral formulation is where Boltzmann first articulated the superposition principle, commonly referred to as the Boltzmann superposition principle [16,33]. This integral formulation for linear viscoelastic materials describes the response to strain as a function of past stresses and the stiffness of the material. It is written as [16]

$$\epsilon \left(t\right)={\int }_0^{t}K(t-\tau )d\sigma \left(\tau \right),$$

(2)

where $K\left(t\right)$ represents the creep stiffness. If the roles of stress and strain are reversed, Equation (2) can be expressed as

$$\sigma \left(t\right)={\int }_0^{t}J(t-\tau )d\epsilon \left(\tau \right),$$

(3)

where $J\left(t\right)$ denotes the creep compliance. The material function $K\left(t\right)$ depends on the material and can be obtained experimentally. In the context of experimentally measured creep data, the creep stiffness $K\left(t\right)$ can be determined as follows:

$$K\left(t\right)={\displaystyle \frac{\sigma }{\epsilon \left(t\right)}},$$

(4)

where $\sigma$ is the sustained load and $\epsilon \left(t\right)$ gives the measured time-dependent creep elongation.

2.1.2. Generalised Maxwell Model and Prony-Series

The basic Maxwell model, shown in Figure 2a, consists of a spring and a dashpot connected in series. This approach can represent simple viscoelastic behaviour, but it cannot fully capture the complex time-dependent response of materials. To address this limitation, multiple Maxwell models are arranged in parallel with an additional spring $K_{eq}$, as shown in Figure 2b. This extended configuration is known as the generalised Maxwell model [21,34,35]. A first-order ODE governs each Maxwell element [21,35].

The total stress $\sigma$ is modelled as the sum of the contributions of the Maxwell elements N, each representing a viscoelastic branch of the material. The constitutive behaviour of the i-th Maxwell element is governed by a first-order differential equation that relates stress and strain through elastic and viscous parameters. The governing equation for the i-th branch is given by [34,36]

$${\displaystyle \frac{d{\sigma }_i\left(t\right)}{dt}}+{\displaystyle \frac{1}{{\tau }_i}}{\sigma }_i\left(t\right)=K_i{\displaystyle \frac{d\epsilon \left(t\right)}{dt}},$$

(5)

where ${\sigma }_i\left(t\right)\in \mathbb{R}$ represents the stress contribution of the i-th Maxwell element, $K_i\in {\mathbb{R}}^{+}$ is the elastic modulus of the i-th spring, and ${\tau }_i\in {\mathbb{R}}^{+}$ is the relaxation time of the i-th Maxwell element, defined as ${\tau }_i={\displaystyle \frac{{\eta }_i}{K_i}}$, with ${\eta }_i\in {\mathbb{R}}^{+}$ being the viscosity of the dashpot.

For the generalised Maxwell model, as detailed in [16,21,29,35,36,37], the stress–strain relationship for viscoelastic materials in the form of a differential equation is formulated as follows:

$${\displaystyle \frac{d\sigma \left(t\right)}{dt}}+{\displaystyle \frac{K_i}{{\eta }_i}}\sigma \left(t\right)=(K_{eq}+K_i){\displaystyle \frac{d\epsilon \left(t\right)}{dt}}+{\displaystyle \frac{K_{eq}{K}_i}{{\eta }_i}}\epsilon \left(t\right).$$

(6)

This differential Equation (6) effectively models the viscoelastic response of complex materials, encompassing both stress relaxation and creep behaviour [36].

A widely used approach to describe the time-dependent viscoelastic response is the Prony series, which can be derived by applying the Laplace transform to the differential formulation of the generalised Maxwell model, such as Equation (6) [16]. The Prony series effectively describes the time-dependent deformation behaviour of materials, which is written as [16]

$$K\left(t\right)=K_{eq}+\sum _{i=1}^{N_i}{K}_i{e}^{-t/{\tau }_i}.$$

(7)

The relaxation function $K\left(t\right)$ characterises the time-dependent decrease in material stiffness due to viscoelastic effects. Each exponential term corresponds to a Maxwell element with a distinct relaxation time ${\tau }_i$, capturing different rates of stiffness relaxation [31,34]. In the following sections, Equation (7) at $t=0$ is referred as $K_{total}$.

In practical applications, the Prony series parameters $K_{eq}$, $K_i$ ${\tau }_i$ are determined by fitting the Prony series to experimental data. This fitting process typically involves optimisation techniques to minimise the deviation between model predictions and observations, yielding a viscoelastic model tailored to the material under investigation.

2.1.3. From the Generalised Maxwell Model to the Fractional Viscoelastic Element

The generalised Maxwell model is limited in its ability to model continuous memory effects. To overcome this limitation, fractional derivatives are introduced, leading to the definition of the fractional viscoelastic element.

The governing equation for the fractional viscoelastic element is obtained by replacing the integer-order derivative in the Maxwell model with a fractional derivative:

$${\phantom{D_t}}_0^C{D}_t^{\alpha }\epsilon \left(t\right)={\displaystyle \frac{\sigma \left(t\right)}{C}},t\in {\mathbb{R}}^{+},\sigma \left(0\right)={\sigma }_0,\epsilon \left(0\right)=0,$$

(8)

where ${\phantom{D_t}}_0^C{D}_t^{\alpha }$ denotes the Caputo fractional derivative of order $\alpha \in (0,1)$, $C\in {\mathbb{R}}^{+}$ is the elastic modulus, $\sigma \left(t\right)\in \mathbb{R}$ is the stress, and $\epsilon \left(t\right)\in \mathbb{R}$ is the strain in the material at time t. The initial condition $\epsilon \left(0\right)=0$ represents the initial strain.

The Caputo fractional derivative is defined as

$${\phantom{D_t}}_0^C{D}_t^{\alpha }f\left(t\right)={\displaystyle \frac{1}{\Gamma (1-\alpha )}}{\int }_0^t{(t-\tau )}^{-\alpha }{\displaystyle \frac{df\left(\tau \right)}{d\tau }}d\tau ,0<\alpha <1,$$

(9)

where $\Gamma \left(·\right)$ is the gamma function, defined as $\Gamma \left(n\right)=(n-1)!$ for $n\in \mathbb{N}$. For $\alpha =1$, the Caputo derivative reduces to the classical first-order derivative, recovering the classical Maxwell model. For $\alpha =0$, the fractional derivative acts as the identity operator, describing purely elastic behaviour (i.e., a spring).

Applying a unit stress step to Equation (8) yields the creep–compliance [16] (Equations (6.2)–(7), p. 340)

$$J\left(t\right)={\displaystyle \frac{t^{\alpha }}{C\Gamma (1+\alpha )}}.$$

(10)

Here, C is considered as fractional (pseudo-) modulus in the unit $\mathrm{Pa}{\mathrm{s}}^{\alpha }$. Hence C controls the amplitude of the creep curve, where a larger C implies a smaller strain under a given load. In the limits $\alpha \to 0$ and $\alpha \to 1$, C reduces to the instantaneous elastic modulus and to the Newtonian viscosity, respectively [16].

The fractional differential equation can be expressed as

$${\phantom{D_t}}_0^C{D}_t^{\alpha }\epsilon \left(t\right)=f_{\theta }\left(\epsilon \left(t\right)\right),\alpha \in (0,1),$$

(11)

where $f_{\theta }$ represents the model of the system dynamics, parametrised by $\theta$ (e.g., C). $f_{\theta }$ is not restricted to linear functions as in the case of the fractional element Equation (8).

To solve the fractional differential equation, one formulates the corresponding Initial Value Problem (IVP)

$$\epsilon \left(t\right)=\epsilon \left(0\right)+{\displaystyle \frac{1}{\Gamma \left(\alpha \right)}}{\int }_0^t{(t-\tau )}^{\alpha -1}{f}_{\theta }\left(\epsilon \left(\tau \right)\right)d\tau .$$

(12)

This IVP can typically only be solved numerically by fractional differential equation solvers [38]. Implementations in modern computational frameworks [28] are employed to solve this equation, allowing the estimation of model parameters like $\alpha$, C. This approach enables the fractional viscoelastic model to capture the complex, continuous memory nature of creep processes in viscoelastic materials.

2.2. Parameter Fitting Approach

As described in the previous sections, the experimental tensile data is used in tests to learn parameters for two models capturing the long-term deformation of SFRCs:

• Model 1: A Maxwell model where the model’s solution is represented by a Prony series. Here, a classical approach with numerically approximated gradients to fit the Prony series to the data is used.
• Model 2: A fractional derivative-based model where the NFDE solver of Zimmering et al. [28] is used. This solver supports exact gradient computation and can thus be used within an optimisation procedure to fit the model’s parameters to the data.

First, the general procedure for parameter fitting and explaining the underlying optimisation concepts used in this study is reviewed. This is followed by the specific approaches for the Prony series and fractional model. Lastly, the differences between the approaches used for the Prony series and fractional derivative models are highlighted.

2.2.1. General Procedure and Concept of Optimisation

Parameter fitting in viscoelastic modelling is an optimisation problem where the objective is to identify a set of model parameters that minimises the differences between experimental measurements and predictions generated by the model. Let

$${\mathbf{y}}_{\exp }=\{y_{\exp }\left(t_1\right),y_{\exp }\left(t_2\right),\dots ,y_{\exp }\left(t_n\right)\}$$

(13)

represent the experimentally measured strain values at discrete time points $\{t_1,t_2,\dots ,t_n\}$. Furthermore, let $\mathcal{M}$ be a model that is parametrised by a parameter vector $\mathit{\theta }$ and outputs predictions

$${\mathbf{y}}_{\mathrm{model}}\left(\mathit{\theta }\right)=\{y_{\mathrm{model}}(t_1;\mathit{\theta }),y_{\mathrm{model}}(t_2;\mathit{\theta }),\dots ,y_{\mathrm{model}}(t_n;\mathit{\theta })\}.$$

(14)

During the fitting process of $\mathcal{M}$, an objective (i.e., the loss) function L is minimised. Often, L is defined as the sum of squared errors

$$L\left(\mathit{\theta }\right)=\sum _{i=1}^n{\left(y_{\exp }\left(t_i\right)-y_{\mathrm{model}}(t_i;\mathit{\theta })\right)}^2.$$

(15)

Gradients of the parameters ${\nabla }_{\mathit{\theta }}$, which provide information about the direction and magnitude of the steepest decrease in $L\left(\mathit{\theta }\right)$, guide the optimisation. The gradient vector is defined as

$${\nabla }_{\mathit{\theta }}L\left(\mathit{\theta }\right)=\left[{\displaystyle \frac{\partial L}{\partial {\theta }_1}},{\displaystyle \frac{\partial L}{\partial {\theta }_2}},\dots ,{\displaystyle \frac{\partial L}{\partial {\theta }_k}}\right],$$

(16)

where $\mathit{\theta }=\{{\theta }_1,{\theta }_2,\dots ,{\theta }_k\}$ represents the model parameters.

In the simplest case, the parameters of the models are updated according to

$${\mathit{\theta }}^{(k+1)}={\mathit{\theta }}^{\left(k\right)}-\eta {\nabla }_{\mathit{\theta }}L\left(\mathit{\theta }\right),$$

(17)

where $\eta >0$ is the learning rate, controlling the step size. The choice of optimisation strategy depends on the model complexity, as well as the number of parameters to be optimised.

Gradient Computation Approaches. In this work, two complementary strategies for obtaining gradient information are employed:

• Numerical Gradient Approximation: Gradients are estimated, using finite-difference methods, as described by Nocedal and Wright [39] (Chapter 8.1): a partial derivative with respect to the i-th parameter can be approximated by perturbing only that parameter by a small step size $ϵ$, i.e.,

  $${\displaystyle \frac{\partial f}{\partial x_i}}\left(x\right)\approx {\displaystyle \frac{f\left(x_1,\dots ,x_i+ϵ,\dots ,x_n\right)-f(x_1,\dots ,x_i,\dots ,x_n)}{ϵ}}.$$
  This procedure is used in many standard libraries (e.g., scipy.optimize in Python). It is efficient when the number of parameters is moderate or when closed-form expressions for the equations are available.
• Exact Gradient Computation: To compute gradients, automatic differentiation (AD) is used. AD is implemented in ML frameworks such as PyTorch [40]. It generalises backpropagation by systematically applying the chain rule [41]. Compared to other methods, it enables highly accurate gradients for complex or high-dimensional models (e.g., fractional differential solvers and/or neural networks). A good introduction into AD, also discussing its distinction from alternative approaches, can be found in Baydin et al. [42].

While both methods are able to be used in gradient-based optimisation, numerical approximations are better suited for simpler models, where the model equations are evaluated in closed form (e.g., Model 1: Prony Series). For Model 2, the fractional derivative model, AD, is better suited as a numerical solver and is therefore integrated into the gradient computation. For such a scenario, AD enables more robust and accurate gradient estimation.

2.2.2. Parameter Fitting for Model 1 (Prony Series)

As introduced in Section 2.1.1, the Prony series expresses creep stiffness as a sum of exponential terms. Here, each term is characterised by a stiffness contribution $K_i$, a relaxation time ${\tau }_i$, and an equilibrium stiffness $K_{\mathrm{eq}}$. To obtain the experimentally measured stiffness $K_{\exp }\left(t_j\right)$,

$$K_{\exp }\left(t_j\right)={\displaystyle \frac{\sigma }{\epsilon \left(t_j\right)}},j=1,\dots ,M$$

is applied. Next, an initial guess for the parameter vector $\left(K_{\mathrm{eq}},\left\{K_i\right\},\left\{{\tau }_i\right\}\right)$ is defined. Let

$$\Delta K_{\exp }=\max \left(K_{\exp }\right)-\min \left(K_{\exp }\right),$$

and set

$$K_{\mathrm{eq}}=\min \left(K_{\exp }\right).$$

$\delta K_{\exp }$ is distributed among n exponential terms using descending weights, ensuring that larger $K_i$ values correspond to earlier terms

$$K_i=\Delta K_{\exp }{w}_i,\mathrm{with}{w}_i\propto (n-i+1)\mathrm{and}\sum _{i=1}^n{w}_i=1.$$

Each relaxation time ${\tau }_i$ is initialised by a simple heuristic such as

$${\tau }_i=10n\left(1/(i+1)\right),$$

which spreads relaxation timescales over several orders of magnitude and helps avoid unphysical parameter values.

The model is then fitted by minimising the weighted least-squares loss function

$$L=\sum _{j=1}^{M}w\left(t_j\right){\left(K_{\exp }\left(t_j\right)-K_{\mathrm{model}}(t_j;K_{\mathrm{eq}},\{K_i,{\tau }_i\})\right)}^2,\mathrm{where}w\left(t\right)=\alpha e^{-t/{\tau }_w}+\beta .$$

(18)

Here, $\alpha =10.0$ is used to assign a high initial weight, ${\tau }_w=15$ to control the decay rate, and $\beta =1.0$ as the baseline weight for later times. This weighting shifts attention to the early time data, as the initial creep behaviour is of high importance.

To ensure physically meaningful parameter values, simple box constraints on all parameters

$$0\le K_{\mathrm{eq}},K_i\le 10·\max \left(K_{\exp }\right),0\le {\tau }_i\le 1000\mathrm{s}$$

are employed. The bounded variant of the Limited-memory Broyden–Fletcher–Goldfarb–Shanno method (L-BFGS-B) [39,43] is used to minimise the loss function shown above. For problems of moderate dimensions, it takes advantage of its low memory requirements and efficient convergence properties. By combining a well-chosen initial guess with a direct, closed-form model and physically reasonable parameter constraints, this approach yields a plausible and fast approximation of the underlying parameters.

2.2.3. Parameter Fitting for Model 2 (Fractional Derivative Model)

As introduced in Section 2.1.3, fractional derivatives generalise integer-order derivatives and thus capture long-term memory effects. The IVP is solved corresponding to the FDE given in Equation (8) numerically using the open-source library FDEint [28]. This PyTorch-based [40] solver supports AD and can thus be used to optimise $\alpha$ and C.

Here, we employ the same weighting function as used for the first model

$$w\left(t\right)=\alpha e^{-t/{\tau }_w}+\beta ,$$

(19)

with $\alpha =10.0$, ${\tau }_w=15$, and $\beta =1.0$. In distinction to the first model, a two-stage optimisation process is used.

First, a global optimisation over all samples is performed simultaneously. Next, the weighted least-squares loss

$$L_{\mathrm{global}}=\sum _{i=1}^{M}w\left(t_i\right){\left({\epsilon }_{\exp }\left(t_i\right)-{\epsilon }_{\mathrm{model}}(t_i;\alpha ,C)\right)}^2,$$

(20)

is minimised. Here, M is the total number of data points, ${\epsilon }_{\exp }\left(t_i\right)$ is the experimentally measured strain, and ${\epsilon }_{\mathrm{model}}(t_i;\alpha ,C)$ is the model prediction. L-BFGS-B is used to find well-scaled initial estimates of $\alpha$ and C.

In a second step, the model is adopted for the individual sample characteristics. $\alpha$ and C are fine-tuned with sample-specific optimisation by minimising

$$L_{\mathrm{sample}}=\sum _{k=1}^{N}w\left(t_k\right){\left({\epsilon }_{\exp }^{\left(j\right)}\left(t_k\right)-{\epsilon }_{\mathrm{model}}^{\left(j\right)}(t_k;\alpha ,C)\right)}^2$$

(21)

for each sample j, where N is the number of data points in that sample. In this stage, the Adam optimiser [44] is used. It needs fewer model evaluations as L-BFGS-B and is more robust to local parameter variations. Each sample-specific fit is initialised with the globally optimised parameters, which ensures stable convergence. This combination of global and sample-specific optimisation balances general behaviour across all datasets with the local adaptability needed for each individual sample.

To enforce physical plausibility during optimisation, C and $\alpha$ are reparametrised using transformations that constrain their values to be positive. Instead of optimising over C, it is optimised over $\log \left(C\right)$ as well as over the inverse sigmoid of $\alpha$ instead of $\alpha$ directly. It is defined as

$${\alpha }_{\mathrm{raw}}=\log \left({\displaystyle \frac{\alpha }{1-\alpha }}\right),$$

(22)

such that during training,

$$C=\exp \left(\log \left(C\right)\right),\alpha =\sigma \left({\alpha }_{\mathrm{raw}}\right),$$

(23)

where $\sigma \left(x\right)={\displaystyle \frac{1}{1+e^{-x}}}$ denotes the sigmoid function. This ensures $C>0$ and $0<\alpha <1$ at all times.

2.2.4. Comparing the Two Approaches

The Prony series approach uses a closed-form expression for creep stiffness. It allows gradients to be computed via numerical approximations and optimised by the L-BFGS-B optimiser. While this work is implemented in Python, this approach is also available in Matlab. The optimiser is computationally efficient for a parametric model and a standard implementation in many tools. These properties make it well suited for engineering applications.

As the fractional derivative model requires solving a fractional differential equation, AD has to be used for gradient computation. To ensure this, the FDEint solver is used in a PyTorch-based (version number 2.5.1) optimisation setup. AD computes exact gradients through the solver, enabling stable gradients and efficient optimisation but at a higher computational cost. Compared to Model 1, which is already the solution of an ODE, this approach is not limited in model complexity. Hence, the fractional model provides a powerful mechanism for capturing long-term memory effects but demands more extensive computational resources and a more involved implementation strategy.

2.3. Trend Analysis Using the Master Curve and Its Standard Deviation

To analyse the trend of the creep behaviour of the samples, a master curve is calculated from the creep curves. For each material, all $M=20$ samples are averaged along the time axis. The master curve allows us to check general trends and analyse the shape of the creep curves as noise is reduced. As all time values $t_{ij}$ of the measured curves are consistent within the data set, the average at each discrete time step $t_{ij}=t_i$ is computed. Formally, the master curve is written as

$${\overline{y}}_i={\displaystyle \frac{1}{M}}\sum _{j=1}^M{y}_{ij},$$

(24)

where $y_{ij}^{\exp }=y_{ij}$ gives the experimental values at each discrete time $t_i$ for the j-th measurement. Averaging these values across all measurements $j=\{1,2,\dots ,M\}$ for each $t_i$, the mean value ${\overline{y}}_i$ is obtained.

Additionally, the standard deviation ${\sigma }_i$ for each $t_i$ is analysed. The standard deviation is defined as

$${\sigma }_i=\sqrt{{\displaystyle \frac{1}{M}}\sum _{j=1}^M{(y_{ij}-{\overline{y}}_i)}^2}.$$

(25)

Here ${\sigma }_i$ measures the deviation of individual data points $y_{ij}$ from the mean value ${\overline{y}}_i$. This metric quantifies the experimental variability.

These two additional metrics help to characterise the general material behaviour as well as the experimental quality.

3. Experimental Setup

This section details the experimental investigation conducted to characterise the creep behaviour of unreinforced composites and SFRCs under controlled sustained loading. The specimen preparation and the experimental setup are also presented here.

3.1. Materials and Specimen Preparation

In this work, standardised tensile test specimens according to DIN EN ISO 527-2 type 1A [45] were used. Figure 3 illustrates the schematic configuration of these specimens, highlighting the measurement of creep behaviour within a gauge length $l_0$ of 75 $\mathrm{m}$$\mathrm{m}$. The matrix material is a thermoplastic PBT, representing the primary component, marketed under the trade name ULTRADUR B 4520, supplied by BASF. The reinforcing component of the composite is E-Glass, designated as FGCS 3540 and provided by STW.

All specimens were produced using an injection moulding process at Kunststoff–Zentrum–Leipzig gGmbH on an Arburg Allrounder 370E machine (Arburg GmbH & Co. KG, Lossburg, Germany). Key parameters from the injection moulding process for both PBT and PBT GF30 are summarised in

Table 1. Injection moulding parameters for the PBT and PBT GF30 specimens.

Parameter	PBT (Unreinforced)	PBT GF30
Material Conditioning and Machine Setup
Drying temperature [°C]	120	120
Drying time [h]	5	5
Mould temperature [°C]	60	60
Melt temperature [°C]	265	265
Screw diameter [mm]	30	30
Number of cavities	2	2
Clamping force [kN]	550	550
Injection Phase
Dosing volume [cm3]	35.0	35.0
Back pressure [bar]	80	80
Dosing speed [mm/s]	300	300
Injection speed [cm3/s]	25	25
Injection pressure [bar]	556 (spec.)	913 (spec.)
Injection time [s]	1.22	1.23
Dosing time [s]	3.93	3.83
Dosing delay [s]	1.00	1.00
Packing and Holding Phase
Switchover point [cm3]	7.0	7.0
Holding pressure [bar]	550	600
Holding time [s]	20.0	20.0
Residual cushion [cm3]	3.50	3.64
Cooling and Cycle Summary
Cooling circuits—nozzle side	2	2
Cooling circuits—ejector side	3	3
Cooling time [s]	24.0	24.0

. These values are taken directly from the production protocols and ensure consistent processing history across specimens. Two types of material were tested, each with 20 specimens, as detailed below.

• Specimen without fibres (PBT);
• Specimen with fibre mass fraction of 30% (PBT GF30);

In the case of the PBT GF30 samples, the fibres were introduced during a compounding process performed before the injection moulding process [46,47]. During compounding, matrix material and fibres were combined and processed into uniform pellets containing a 30% fibre mass fraction. Subsequently, these compounded pellets were used in the injection moulding process to produce the final tensile test specimens with fibres.

3.2. Creep Testing and Measurement Setup

Tensile creep experiments were conducted according to the DIN EN ISO 899-1 [48] standard, which ensures consistency and reliability in creep testing. As specified in the standard, the full force must be applied within 1 $\mathrm{s}$ to 5 $\mathrm{s}$ to ensure consistency in initial loading conditions. For PBT samples, a force of 300 $\mathrm{N}$ was applied, while PBT GF30 specimens, with higher stiffness, required a force of 1025 $\mathrm{N}$. This adjustment ensures comparable elongation between materials, facilitating meaningful comparisons. For both materials, the force was fully applied within 3 $\mathrm{s}$. The creep behaviour test was measured over 300 $\mathrm{s}$ at room temperature under controlled conditions (see

Table 2. Summary of experimental conditions for creep behaviour testing of PBT and PBT GF30 specimens.

Type of Material	Force [N]	Time to Apply the Force [s]	Duration of Creep Measurement [s]
PBT	300	3	300
PBT GF30	1025	3	300

).

The creep behaviour of the specimens was measured using the ARAMIS 3D digital image correlation (DIC) system (Carl Zeiss GOM Metrology GmbH, Braunschweig, Germany). The experimental setup is depicted in Figure 4a. This optical measuring system is specifically designed for a non-contact, full-field deformation analysis. The setup used two high-resolution 12 MP cameras positioned with 700 $\mathrm{m}$$\mathrm{m}$ in the specimen. The force was applied and controlled using the ZWICK/ROELL Z050 tensile test machine (ZwickRoell GmbH & Co. KG, Ulm, Germany).

The ARAMIS system uses a non-contact optical technique to analyse the displacement of a stochastic speckle pattern applied to the surface of the sample [49,50]. As illustrated in Figure 4b, the distribution of the strain (${\epsilon }_y$) on the surface of the PBT specimen under a sustained load of $300 \mathrm{N}$ is visualised. The strain distribution is calculated using correlation algorithms that track changes in the position of the speckle features between the reference state (unloaded) and the deformed state (loaded) [50,51].

4. Results

This section presents the results of the experimental investigation of creep behaviour for unreinforced (PBT) and fibre-reinforced (PBT GF30) specimens under sustained loading. The experimental setup and conditions are summarised in Table 2. Additionally, the predictive performance of two modelling approaches, the Prony series and the fractional derivative method, is evaluated to approximate the experimental data.

4.1. Experimental Results of Creep Behaviour

The creep behaviour of PBT and PBT GF30 under sustained load conditions was analysed using data from 20 samples per material type. The results shown in Figure 5 provide a comparative analysis of the time-dependent deformation characteristics of these two materials. In all samples tested, the initial deformation at the beginning of the creep lies in a comparable range of around $0.26$%. This indicates a consistent elastic response prior to viscoelastic deformation. The creep responses of primary and secondary creep stages, as discussed in Section 2.1.1, are observed in both materials and exhibit characteristics. However, the second stage cannot be clearly separated from the tertiary stage of creep. Clear indicators such as an accelerating strain rate or failure are missing. However, a detailed classification of creep stages or associated mechanisms lies beyond the scope of this study and is therefore not pursued further.

The master curves in Figure 6 of the total strain ${\epsilon }_{\mathrm{total}}$ and Figure 7, where the creep strain ${\epsilon }_{\mathrm{creep}}$ is shown, indicate that PBT exhibits a slightly higher strain rate compared to PBT GF30. These master curves, which are introduced in detail in Section 2.3, provide more detail on the general trend of deformation compared to the results in Figure 5. These findings align with previous studies [52,53,54], which reported similar trends in the viscoelastic behaviour of unreinforced and fibre-reinforced materials. The viscoelastic behaviour is governed primarily by the matrix material rather than the reinforcing fibres [53,55,56].

In particular, the master curves in Figure 7 show a slightly steeper creep rate for PBT compared to PBT GF30. This suggests a reduced resistance to time-dependent deformation in the absence of fibre reinforcement, specific to the PBT system studied here. In contrast, the inclusion of short glass fibres in PBT GF30 results in lower creep rates. These findings emphasise the role of fibre reinforcement in altering the viscoelastic response of PBT under constant load conditions.

4.2. Prony Series Coefficients for PBT and PBT GF30

The viscoelastic properties of the PBT and PBT GF30 materials are characterised by fitting their experimental creep data to the Prony series model using a different number of branches $N_i=\{1,2,3\}$ with respect to the Equation (7). The coefficients $K_{eq}$, $K_i$, and ${\tau }_i$ were determined using the fitting procedure detailed in Section 2.2. The resulting coefficients for each specimen are summarised in

Table A4. Parameters for Prony 2—PBTGF0.

Sample	${\mathit{K}}_{\mathit{eq}}$	${\mathit{K}}_1$	${\mathit{\tau }}_1$	${\mathit{K}}_2$	${\mathit{\tau }}_2$	${\mathit{K}}_{\mathbf{total}}$
0	2627	89	120	34	8	2749
1	2609	64	116	44	7	2716
2	2622	60	127	37	8	2719
3	2516	82	118	36	5	2634
4	2572	77	279	53	10	2703
5	2609	75	99	36	5	2720
6	2589	71	137	43	6	2703
7	2638	71	110	33	4	2742
8	2579	76	174	30	10	2685
9	2568	84	216	46	10	2698
10	2522	67	180	33	13	2622
11	2591	69	88	30	7	2690
12	2564	62	156	48	14	2675
13	2627	76	119	31	6	2734
14	2593	63	252	43	10	2700
15	2630	66	50	42	2	2738
16	2571	67	64	37	4	2675
17	2569	79	138	47	6	2695
18	2574	72	113	30	9	2675
19	2587	79	223	50	17	2716

,

Table A5. Parameters for Prony 3—PBTGF0.

Sample	${\mathit{K}}_{\mathit{eq}}$	${\mathit{K}}_1$	${\mathit{\tau }}_1$	${\mathit{K}}_2$	${\mathit{\tau }}_2$	${\mathit{K}}_3$	${\mathit{\tau }}_3$	${\mathit{K}}_{\mathbf{total}}$
0	2606	86	245	37	29	21	4	2750
1	2571	81	416	41	23	25	4	2718
2	2618	59	164	38	12	8	1	2722
3	2478	99	337	38	20	22	2	2637
4	2572	77	279	0	10	53	10	2703
5	2596	72	180	39	15	17	1	2725
6	2589	71	137	0	10	43	6	2703
7	2575	106	623	35	39	28	3	2743
8	2500	122	1000	36	85	27	9	2685
9	2568	84	216	0	24	46	10	2698
10	2522	67	180	0	67	33	13	2622
11	2571	65	242	46	20	15	1	2696
12	2564	62	156	7	14	42	14	2675
13	2612	75	215	30	22	18	3	2735
14	2507	141	1000	44	17	15	1	2708
15	2615	53	169	39	14	34	1	2741
16	2560	53	155	48	15	20	1	2681
17	2562	80	177	13	1	42	9	2698
18	2574	72	113	0	69	30	9	2675
19	2480	172	964	58	26	8	3	2718

, and

Table A6. Mean L1 loss comparison for PBTGF30.

Smp.	Fractional Damper	Prony 1	%Diff	Prony 2	%Diff	Prony 3	%Diff
0	3.65 × 10−4	5.44 × 10−4	48.8%	1.40 × 10−4	−61.7%	1.39 × 10−4	−61.8%
1	2.79 × 10−4	3.67 × 10−4	31.4%	1.19 × 10−4	−57.4%	1.18 × 10−4	−57.6%
2	2.18 × 10−4	5.52 × 10−4	153.9%	1.51 × 10−4	−30.8%	1.51 × 10−4	−30.8%
3	2.06 × 10−4	3.87 × 10−4	87.8%	1.26 × 10−4	−38.8%	1.23 × 10−4	−40.6%
4	2.96 × 10−4	3.82 × 10−4	29.3%	1.61 × 10−4	−45.5%	1.61 × 10−4	−45.5%
5	2.06 × 10−4	3.14 × 10−4	52.1%	2.17 × 10−4	5.3%	2.03 × 10−4	−1.7%
6	3.23 × 10−4	3.85 × 10−4	19.2%	1.54 × 10−4	−52.5%	1.54 × 10−4	−52.3%
7	2.36 × 10−4	4.47 × 10−4	89.1%	1.51 × 10−4	−36.3%	1.51 × 10−4	−36.3%
8	3.54 × 10−4	3.83 × 10−4	8.1%	2.41 × 10−4	−31.8%	2.15 × 10−4	−39.4%
9	2.10 × 10−4	3.74 × 10−4	77.7%	1.49 × 10−4	−29.0%	1.49 × 10−4	−29.0%
10	2.06 × 10−4	5.77 × 10−4	179.6%	1.52 × 10−4	−26.2%	1.46 × 10−4	−29.4%
11	3.62 × 10−4	5.83 × 10−4	61.2%	1.60 × 10−4	−55.8%	1.60 × 10−4	−55.8%
12	2.41 × 10−4	3.73 × 10−4	54.6%	1.31 × 10−4	−45.6%	1.31 × 10−4	−45.7%
13	2.32 × 10−4	4.75 × 10−4	104.3%	1.43 × 10−4	−38.5%	1.29 × 10−4	−44.5%
14	3.10 × 10−4	5.17 × 10−4	66.5%	1.70 × 10−4	−45.1%	1.68 × 10−4	−45.8%
15	2.62 × 10−4	3.17 × 10−4	21.3%	2.05 × 10−4	−21.8%	1.83 × 10−4	−30.3%
16	2.51 × 10−4	4.60 × 10−4	83.6%	1.59 × 10−4	−36.4%	1.59 × 10−4	−36.4%
17	2.59 × 10−4	5.80 × 10−4	124.1%	2.15 × 10−4	−16.8%	1.94 × 10−4	−25.0%
18	2.50 × 10−4	3.73 × 10−4	49.6%	2.28 × 10−4	−8.5%	2.28 × 10−4	−8.5%
19	2.67 × 10−4	5.01 × 10−4	87.5%	1.91 × 10−4	−28.6%	1.61 × 10−4	−39.9%
Average	$\mathbf{2.67}\times {\mathbf{10}}^{-\mathbf{4}}$	$\mathbf{4.44}\times {\mathbf{10}}^{-\mathbf{4}}$	71.5%	$\mathbf{1.68}\times {\mathbf{10}}^{-\mathbf{4}}$	−35.1%	$\mathbf{1.61}\times {\mathbf{10}}^{-\mathbf{4}}$	−37.8%

, respectively.

The mean values and standard deviations of the Prony series coefficients for both materials are presented in

Table 3. Mean values and standard deviations (Mean ± StdDev) of Prony-series coefficients for Prony models with 1 to 3 terms.

Prony Model	Parameter	PBTGF0	PBTGF30
Prony 1	$K_{eq}$	$2607.86\pm 30.74$	$9288.97\pm 62.72$
	$K_1$	$83.29\pm 6.54$	$211.06\pm 20.41$
	${\tau }_1$	$35.12\pm 7.97$	$44.36\pm 9.83$
	$K_{\mathrm{total}}$	$2691.15\pm 31.03$	$9500.03\pm 63.12$
Prony 2	$K_{eq}$	$2587.84\pm 32.41$	$9201.04\pm 83.65$
	$K_1$	$72.42\pm 7.65$	$219.80\pm 59.49$
	${\tau }_1$	$143.91\pm 58.90$	$243.57\pm 212.33$
	$K_2$	$39.20\pm 7.21$	$96.43\pm 26.91$
	${\tau }_2$	$8.04\pm 3.63$	$10.57\pm 5.83$
	$K_{\mathrm{total}}$	$2699.47\pm 32.14$	$9517.27\pm 59.73$
Prony 3	$K_{eq}$	$2561.99\pm 41.87$	$9140.27\pm 123.57$
	$K_1$	$84.94\pm 29.50$	$240.49\pm 102.30$
	${\tau }_1$	$348.40\pm 291.28$	$447.15\pm 343.35$
	$K_2$	$27.43\pm 19.02$	$67.33\pm 40.98$
	${\tau }_2$	$26.59\pm 21.48$	$89.64\pm 93.09$
	$K_3$	$27.29\pm 12.52$	$72.53\pm 41.13$
	${\tau }_3$	$5.22\pm 4.16$	$7.95\pm 7.11$
	$K_{\mathrm{total}}$	$2701.65\pm 32.32$	$9520.62\pm 59.09$

. The total stiffness $K_{\mathrm{total}}$ derived from the Prony series representations in these tables corresponds to the effective stiffness experimentally determined in accordance with the DIN EN ISO 527-1 standard [57]. This total stiffness characterises the initial elastic response of the materials and serves as a reference for calibrating Prony series models with varying numbers of branches, specifically $N=1$, $N=2$, and $N=3$. The corresponding Young’s moduli, determined in accordance with the DIN EN ISO 527-1 standard, were obtained from tensile tests on 10 samples per material, yielding mean values of $2674.50$ $\mathrm{M}$$\mathrm{Pa}$ for PBT and $9523.12$ $\mathrm{M}$$\mathrm{Pa}$ for PBT GF30.

Significant differences in the mean Prony series coefficients between PBT and PBT GF30 underscore the role of fibre reinforcement in the modification of viscoelastic properties. Although the matrix material primarily governs time-dependent deformation, the presence of fibres considerably enhances the overall stiffness and resistance to creep. As noted above, the mean relaxation time ${\tau }_1$ and the stiffness distribution $K_1$ (see Table 3) for PBT GF30 are higher compared to PBT, further supporting the non-parallel nature of the master curves shown in Figure 7. These observations are consistent with findings reported in the literature [52,53,54], which similarly demonstrated the influence of fibre reinforcement on the viscoelastic behaviour of the polymer.

A comparative evaluation of experimental creep data with two predictive modelling approaches, the classical Prony series model and the fractional viscoelastic model, is presented in Figure 8 for PBT and in Figure 9 for PBT GF30. These diagrams highlight the ability of each model to reproduce the experimentally observed viscoelastic behaviour over the measured time domain.

For both materials, it is evident that the Prony model with a single branch ($N=1$) does not capture experimental trends with sufficient accuracy. This is particularly apparent in the early and intermediate time regimes, where deviations from the measured data are most pronounced. In contrast, the fractional viscoelastic model provides a significantly improved fit across the full time range, especially for low-order Prony approximations. This can be attributed to the inherent capacity of fractional models to represent broad relaxation spectra with fewer parameters.

However, as the number of branches in the Prony series increases, the model’s ability to replicate the experimental data improves notably. From $N=2$ onward, Prony-based predictions show a better agreement with the measured curves compared to the accuracy of the fractional model. This trend is consistently observed for both material systems, demonstrating the flexibility of the Prony approach when extended with sufficient model complexity. Nevertheless, the trade-off lies in the increased number of fitting parameters and the potential risk of over-fitting or loss of physical interpretability.

In summary, while the fractional model yields an excellent initial approximation of the experimental data with a compact formulation, the Prony series model achieves superior accuracy when configured with multiple branches. This highlights the importance of selecting a modelling strategy based on the required fidelity, parameter efficiency, and application context.

4.3. Fractional Model Coefficients for PBT and PBT GF30

The results for the fractional element in Equation (8) and fitted according to the methodology in Section 2.2.3 can be found in Table 4.

Table 4. Mean values and standard deviations (Mean ± StdDev) of Fractional Element parameters.

Parameter	PBTGF0	PBTGF30
C	5.34 × 105 ± 1.63 × 105	3.55 × 106 ± 9.60 × 105
$\alpha$	0.35 ± 0.05	0.40 ± 0.04

Table 4. Mean values and standard deviations (Mean ± StdDev) of Fractional Element parameters.

Parameter	PBTGF0	PBTGF30
C	5.34 × 105 ± 1.63 × 105	3.55 × 106 ± 9.60 × 105
$\alpha$	0.35 ± 0.05	0.40 ± 0.04

Here, C describes the immediate response of the corresponding material, similar to $K_{total}$ in the previous section. While $K_{total}$ rises by $\approx 3.5$ times for PBT GF30, C rises by 6.65 times, which is significantly higher.

The fractional derivate $\alpha$ reflects the mix of elastic ($\alpha =0$) and viscous ($\alpha =1$) behaviour. Its mean values are slightly higher for PBT GF30, but in a similar range. As the corresponding values overlap, due to high standard deviation, it cannot be clearly stated that PBT GF30 shows more viscous behaviour than PBT. The slight increase could be explained by additional micro-slip or friction of the fibres inside the matrix material. Since clear evidence is missing, either more samples or an improved experimental setting is necessary to study the effect in detail.

In summary, the results show that the immediate response C significantly increases for the fibre0enhanced material while the ratio between elastic and viscous behaviour remains similar.

5. Discussion

The measured creep behaviour of fibre-reinforced polymers (PBT GF30) and unreinforced polymers (PBT) is in line with expectations. It confirms that fibre reinforcement plays a role in modulating viscoelastic responses. The PBT-GF30 samples deform less under sustained load than the unreinforced PBT samples, an expected effect, as shown in Figure 7, as the glass fibres stiffen the matrix and inhibit creep strain [52,53,54]. The master curves shown in Figure 6 indicate that while both the fractional derivative model and the Prony series provide reasonable approximations of the creep response, there are notable differences in terms of accuracy and parameter stability.

For single-term Prony models ($N=1$), the approximation of the creep behaviour is less accurate than the fractional model. The fractional model provides a smoother representation of the viscoelastic response. When additional terms are added to the Prony series ($N\ge 2$), the accuracy improves. This increased number of parameters increases potential over-fitting. The higher standard deviations of the Prony coefficients suggest more variability, which also indicates that the Prony coefficients lack robustness. This variability challenges the physical interpretability of the Prony model, as over-fitting becomes a concern with higher N. In contrast, the fractional derivative model demonstrates better approximation accuracy with a lower number of required parameters. The fractional order parameter ($\alpha$) shows stability in different material compositions. This underscores the suitability of fractional calculus-based models to provide a compact and accurate representation of creep behaviour. However, similarly to the Prony series model, greater standard deviations of the parameter C are observed.

Determining the parameters of fractional derivative models presents a challenge due to the complexity of fractional differential equations. This study explores optimisation techniques such as numerical and automatic differentiation-based frameworks to efficiently fit experimental data. While the Prony series benefits from a closed-form solution that simplifies parameter fitting, the fractional model requires solving fractional differential equations numerically, increasing computational demands. To mitigate these challenges, gradient-based optimisation methods are employed, ensuring stable parameter convergence across multiple datasets. Moreover, the application of data-driven techniques, such as NFDE solvers, offers a promising alternative for parameter identification.

Recent developments in the application of artificial intelligence to constitutive modelling have demonstrated significant potential, particularly in areas requiring robust parameter identification from noisy or incomplete data. The novelty of this study lies not in the proposal of new constitutive equations but in the implementation of a machine learning-inspired approach, specifically a neural fractional differential equation (NFDE) solver, to identify the governing parameters of both the Prony and fractional derivative models. This technique combines automatic differentiation with gradient-based optimisation, treating the governing differential equation as an operator constraint. Although the application of AI tools to curve fitting or surrogate modelling in viscoelasticity is a well-established area of research, as evidenced by the work of [58,59], the use of a solver to extract interpretable and physically meaningful parameters from time domain data for fractional models represents a novel contribution. The approach demonstrated stable convergence and interpretability across the two material systems studied, highlighting the emerging role of data-driven methods in enhancing classical mechanical modelling.

One of the key challenges in viscoelastic modelling is to determine the model parameters efficiently. The Prony series requires multiple relaxation times and corresponding stiffness coefficients, increasing the complexity of parameter identification. As the number of branches in the Prony model increases, over-fitting becomes a concern, leading to large standard deviations in the model coefficients. In contrast, the fractional derivative model requires fewer parameters, making it computationally more efficient.

The results suggest that existing numerical optimisation techniques, such as gradient-based or evolutionary algorithms, could be refined to improve the robustness of parameter identification for both models. The need for an improved fitting procedure is evident, as the high variability in Prony coefficients indicates the sensitivity of experimental noise data.

Although the test duration was limited to 300 s, the fractional viscoelastic model used allows reliable extrapolation because of its inherent power-law behaviour, as demonstrated in recent studies [60,61]. Compared to the Prony series, it offers better long-term predictions from short-term data.

The findings of this study confirm that both the fractional derivative model and the generalised Maxwell model based on the Prony series are viable approaches to modelling the creep behaviour of fibre-reinforced and unreinforced materials. However, the results highlight a critical issue in parameter estimation, particularly for the Prony series model, which exhibits a high standard deviation in parameter values due to over-fitting. The fractional derivative model emerges as a promising alternative, providing a more compact reparametrisation while maintaining accuracy in long-term creep predictions.

It is important to emphasise that the conclusions drawn in this study are specific to the thermoplastic polyester PBT and its short-fibre-reinforced variant PBT GF30. While the viscoelastic modelling strategies developed here may generally be applicable, thermoplastic systems vary greatly in molecular structure, crystallinity, and processing sensitivity. Therefore, caution should be exercised when extrapolating these findings beyond the tested materials.

This emphasises the need for a better or more robust fitting procedure that ensures accuracy while preserving physical interpretability, mitigates over-fitting, and enhances the predictive capability of viscoelastic models. Future research should focus on refining optimisation algorithms, improving the consistency of experimental data, and exploring hybrid modelling approaches that integrate the advantages of both methodologies.

6. Conclusions

The present study investigated the tensile creep behaviour of unreinforced PBT and short-fibre-reinforced PBT GF30 (with a fibre mass fraction of 30%) using two viscoelastic modelling approaches, the classical Prony series and a fractional derivative formulation. A comparative analysis was conducted, which revealed that the fractional derivative model outperforms the Prony series in terms of parameter stability and compactness, especially at low model orders. While the Prony model becomes more accurate with additional terms, it also becomes more susceptible to overfitting and parameter variability. In contrast, the fractional model provides accurate results with fewer parameters and demonstrates increased robustness in predicting larger time ranges.

A central contribution of this work lies in the application of a neural fractional differential equation (NFDE) solver for parameter identification. This approach enables the data-driven extraction of physically interpretable parameters from creep data in the time domain and demonstrates stable convergence, offering a viable alternative to conventional curve-fitting techniques.

Despite the limited experimental duration (300 s), the fractional model enables credible extrapolation due to its inherent power-law behaviour. The methods and conclusions presented are specific to PBT and PBT GF30 and should be cautiously extended to other polymer systems due to variations in morphology and processing sensitivity.

Future work should focus on refining optimisation algorithms for parameter identification, improving data robustness, and exploring hybrid modelling strategies that combine the strengths of fractional and classical viscoelastic representations.