Abstract
Polymer composites show complex, nonlinear mechanical behavior. This makes their rational design hard. Running tests or high-fidelity simulations for every candidate design is slow and costly. Purely data-driven models are fast, but they can break basic mechanics outside the training range. Most physics-informed (PI) models, on the other hand, target a single constitutive law or failure mode, and their loss weights are set by hand. This study presents a DBN-GJO-PI-RNN surrogate that links deep feature learning, physics-constrained learning and failure screening in one workflow. A Deep Belief Network (DBN) learns compact latent features from the material and loading inputs. A PI Recurrent Neural Network (PI-RNN) then predicts the mechanical response under a composite loss with data, constitutive, equilibrium and boundary terms. Golden Jackal Optimization (GJO) tunes the network settings and the three loss weights together. Its fitness function rewards both accuracy and physical consistency. The predicted failure index is then checked against the onset criterion (FI ≥ 1). We tested the framework on a public dataset of 5000 polymer-composite samples with six response targets. On the test set, it reached a MAE of 1.1132, a RMSE of 1.3962 and an R2 of 0.9932. Its RMSE was 38.3% lower than that of the best purely data-driven baseline (TCN) and 50.6% lower than that of a standard RNN. The ablation study showed that each module reduced the error. The framework offers a fast, physics-aware surrogate for screening polymer-composite designs before costly tests or simulations.
Keywords: polymer composites, surrogate modeling, physics-informed machine learning, deep belief network, physics-informed recurrent neural network, golden jackal optimization, failure assessment
1. Introduction
Polymer composite materials with high specific strength, stiffness, light weight, anti-corrosion properties, and flexible designs have been developed as critical engineering materials because of these properties [1]. These materials are utilized in a variety of industries, including aerospace, automotive, marine, energy, and structural industries, which require an accurate understanding of their mechanical behaviors and failure modes [2]. Flexible glass for infrared adaptive optics is another example of an advanced material where its mechanical behavior matters in use [3]. The heterogeneity, anisotropy, and nonlinearity of these materials make it difficult to understand their properties’ relationship with loading conditions, stresses, strains, and damage progression [4]. Experiments are required to characterize these materials, but the processes are often costly, laborious, and challenging under different operating conditions [5]. Combining numerical simulation with experiments is one way to reduce this burden, as seen with the process optimization of electrolytic-assisted milling of titanium alloy [6]. Consequently, data-driven DL has received growing interest for learning the complicated behaviors of these composite materials using experimental and verified datasets [7]. Pure data-driven approaches can result in physically unreasonable predictions when the datasets and loading conditions are not sufficient [8].
The mechanical performance of high-performance polymer composite materials depends on the properties of the matrix, reinforcement materials, fiber orientation, volume fractions, and interfacial behavior [9]. Additionally, external factors including the degree of loading, loading direction, strain rates, temperatures, and cyclic loading may affect their stress–strain behavior [10]. Time-dependent effects, such as creep, also change the mechanical response of polymer-based materials, as observed for HTPB solid propellant [11]. Defects, voids, delamination, interfacial debonding, and other flaws might help initiate and propagate the damage [12]. Failure can also switch modes when geometric constraints change. A sawtooth-to-brittle transition in cubane architectures is one such example [13]. This interaction leads to nonlinear correlations between the characteristics of the material and its mechanical behavior [14]. The conventional experimental techniques can account for such behaviors, but they usually take a lot of time and money [15]. Thus, adequate computational models should be developed to predict the mechanical behavior under various material and loading conditions [16]. Mesoscale simulation is one such tool. For example, a discrete element model has been used to study the meso-structure of recycled aggregate stabilized with geopolymer [17].
Different machine learning (ML) and DL of One-Dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), DBN, and PI Neural Network (PINN) algorithms have been studied for composite mechanical-response prediction. The models based on CNN techniques can detect local features, while LSTM, BiLSTM, and GRU can detect sequential relations in mechanical data [18]. The TCN algorithm offers an alternative way of temporal modeling using convolution techniques, and the DBN technique can learn hierarchical nonlinear feature representations [19]. PINN techniques include physical governing equations in the learning process and can provide more physical consistency than purely data-driven models. However, CNN and recurrent models traditionally focus on optimizing prediction errors and can break physical governing equations under unseen conditions [20], DBN-based models can offer effective representations without enforcing the governing mechanical equations [21], PINN-based techniques can be sensitive to the choice of loss weight and model parameters, and traditional optimization methods cannot find the right configurations efficiently [22]. Recent work applies causal inference-based neural architecture search to a PINN for fatigue life prediction [23]. This reflects the same need for a systematic search of model settings.
To address these challenges, this study proposes a DBN-GJO-PI-RNN framework, which combines the advantages of deep feature learning, PI recurrent modeling, and metaheuristic optimization for high-performance polymer composites. First, DBN is employed to capture nonlinear features based on material properties, loads, and response data; PI-RNN is used to learn nonlinear mechanical responses based on the physics constraint equation. Second, a PI loss function is defined that consists of three components: data loss, physics-residual loss, and constraint loss, thereby enhancing the accuracy and physical consistency of predictions. Third, GJO is used to optimize the PI-RNN hyperparameters to obtain an efficient PI-RNN. Finally, the framework conducts failure-criterion evaluation to analyze the failure behavior based on new composite conditions. The major contribution of our framework is the integrated combination of DBN feature extraction, PI-RNN learning, GJO-based optimization, and failure assessment. The key contributions of this research are summarized below:
-
⮊
Develop a novel DBN-GJO-PI-RNN framework for predictive modeling and failure assessment of high-performance polymer composites.
-
⮊
Integrate DBN-based hierarchical feature extraction with PI-RNN to learn nonlinear and sequential composite mechanical behavior.
-
⮊
Formulate a PI loss function combining data, physics, and constraint terms to enhance physical consistency.
-
⮊
Optimize the PI-RNN configuration using GJO to improve prediction accuracy and model robustness.
-
⮊
Establish an integrated strategy for mechanical-response prediction and failure assessment of high-performance polymer composites.
1.1. Novelty of the Study
Earlier work on polymer composites falls into two groups. Data-driven models such as DNN, CNN and LSTM learn property trends well, but they ignore mechanics [24,25,26,27]. PI models respect mechanics, but they usually solve one constitutive law or one failure mode. Their loss weights are also tuned by trial and error [28,29,30,31,32]. This study closes that gap in four ways.
First, it chains a DBN and a PI-RNN. The DBN compresses the raw material and loading inputs into a 64-dimensional latent vector. The physics terms then act on the recurrent predictions made from these features. None of the cited PI composite models used a separate unsupervised feature-learning stage of this kind.
Second, GJO tunes the physics-loss weights (λd, λp, λc) together with the network settings. Its fitness function adds the physics and constraint residuals to the validation error (Equation (21)). The search favors models that are accurate and physically consistent at the same time. This replaces the manual weight tuning that limits many PINN studies [22]. Recent PINN work searches the network architecture automatically [23]. The search here also covers the three loss weights.
Third, one surrogate predicts six mechanical responses and a failure index in a single pass. A designer can screen a composite for failure onset (FI ≥ 1) without running a separate damage model.
Fourth, a stepwise ablation shows what each module adds. PI learning cuts the MAE of a standard RNN by 20.2%. DBN features cut it by a further 20.5%, and GJO tuning by another 19.7%.
To our knowledge, no earlier study has combined these four elements for polymer composites. The result is a machine-learning surrogate that can stand in for repeated tests or simulations during early design screening.
The proposed DBN-GJO-PI-RNN framework combines deep feature learning, GJO-based hyperparameter optimization, and physics-informed recurrent prediction within a unified architecture. Unlike approaches that apply these techniques separately, it integrates their complementary functions to improve feature representation, optimization efficiency, and physics-informed prediction. This integration aims to enhance predictive accuracy and consistency in polymer-composite mechanical behavior.
1.2. Research Organization
This paper is organized as follows: Section 2 presents the literature review, Section 3 describes a developed DBN-GJO-PI-RNN framework, Section 4 presents an experimental setup and evaluation, Section 5 discusses the results, and Section 6 concludes the study and outlines future work.
2. Literature Review
Recent studies have increasingly explored DL techniques for predicting the mechanical behavior and material properties of polymer and fiber-reinforced composites. Dean et al. [28] developed a PIDL constitutive framework to predict the complex anisotropic, nonlinear, and pressure-dependent mechanical behavior of short fiber-reinforced polymers (SFRPs). The study aimed to improve the generalization of constitutive modeling under varied loading conditions by integrating physical laws with neural-network learning. The method employed LSTM networks, scalar invariants based on transverse isotropy, thermodynamic consistency, material symmetry, and synthetic data generated from a validated continuum-mechanical model; an open-hole tensile (OHT) test was used for validation. The results demonstrated improved stress prediction after retraining with randomly generated loading paths, showing the potential of PIDL for generalizable composite constitutive modeling. The initial model showed limitations under complex multiaxial stress states and required additional retraining with diverse loading paths, indicating a dependence on representative training data and limited robustness for highly complex loading conditions.
Liu et al. [33] introduced a DL framework with an attention mechanism and three-dimensional stress-state information to efficiently predict compressive mechanical properties and failure modes of embedded wrinkle thick-section composites. This objective was to reduce the computational cost of high-fidelity simulations while accurately modeling composite strength, stiffness, stress–strain behavior, and damage modes. This framework employed CNN-based strength, mechanical behavior, and failure prediction networks, with an attention-based loss function used to improve local high-damage-region prediction; training data were generated using 3D finite-element simulations with a progressive damage method. The framework accurately predicted compressive mechanical properties and failure modes and identified an influence of wrinkle parameters on composite behavior. This approach relies on high-fidelity finite-element simulations for dataset generation and primarily focuses on wrinkle-induced composite damage, limiting its general applicability to diverse polymer-composite loading and failure conditions. Bahtiri et al. [29] examined a PIDL constitutive model to predict the viscoelastic–viscoplastic behavior of short fiber-reinforced nanoparticle-filled epoxy nanocomposites under varying conditions. The framework combined LSTM and feed-forward neural networks (FNNs) to model internal variables and free-energy functions while enforcing thermodynamic consistency. Synthetic constitutive-model data and cyclic loading–unloading experimental data were used for training, achieving accurate mechanical-response prediction under different fiber, nanoparticle, and hygrothermal conditions. The study does not incorporate deep feature extraction, metaheuristic optimization, or explicit failure-criterion assessment, limiting its scope for integrated composite prediction and failure analysis.
Yao et al. [30] conducted a PIML model to predict Mode I fatigue delamination propagation considering fiber-bridging effects. The study used Mode I fatigue experiments, a Paris-type semi-empirical correlation, and PINNs to incorporate fatigue-damage evolution as a physical constraint. The PINN achieved accurate crack-growth-rate predictions and outperformed conventional ML and the Paris-type model. This framework is specifically focused on Mode I fatigue delamination and fiber-bridging effects, limiting its applicability to broader mechanical-response prediction and diverse failure mechanisms in high-performance polymer composites. Linghu et al. [31] examined a Higher-Order Multiscale PINN (HOMS-PINN) framework for efficiently solving multiscale elastic problems in composite materials with discontinuous and high-contrast properties. The objective was to reduce the computational cost and overcome the limitations of conventional DL and direct PINN approaches. The method integrated higher-order multiscale modeling, PINNs, automatic differentiation, independent neural networks, homogenization, and transfer learning to capture microscopic and macroscopic mechanical behavior. Numerical experiments on high-contrast, multi-inclusion, and three-dimensional composites demonstrated accurate multiscale predictions with reduced computational expense. It primarily addresses elastic multiscale analysis and does not explicitly consider sequential mechanical-response prediction, deep feature extraction, hyperparameter optimization, or failure assessment. Malashin et al. [24] investigated the application of DL for predicting the mechanical properties of natural-fiber polymer composites under different material and processing conditions. The objective was to accurately model nonlinear relationships between fiber type, polymer matrix, surface treatment, and processing parameters. The study used linear regression, Random Forest, Gradient Boosting, and Deep Neural Networks (DNNs), with Optuna-based hyperparameter optimization to determine the best DNN architecture. Using experimental data from flax, cotton, sisal, and hemp fibers incorporated into PLA, PP, and epoxy matrices, the optimized DNN achieved the best predictive performance. However, the study was purely data-driven and did not incorporate governing physical laws, constitutive constraints, PI loss functions, or explicit failure assessment, limiting its ability to ensure physical consistency under unseen loading conditions.
Nasri and Toubal [25] investigated the application of an Artificial Neural Network (ANN) for predicting a mechanical and impact performance of short natural-fiber thermoplastic biocomposites exposed to ultraviolet radiation. The objective was to predict tensile strength, maximum impact load, and absorbed impact energy using non-destructive experimental parameters. The study employed a MATLAB-based ANN model, with mass, hardness, roughness, and natural frequencies as input features and correlation measures and MSE for performance evaluation. The results demonstrated that ANN can effectively capture the relationship between non-destructive measurements and the mechanical behavior of natural-fiber thermoplastic composites. The ANN model primarily establishes empirical relationships between input measurements and material responses, without explicitly modeling the sequential evolution of mechanical behavior or the underlying physical mechanisms governing degradation and failure. Yin et al. [26] developed a CNN-based framework to predict stress–strain behavior and mechanical properties of fiber-reinforced composites with random microstructures. The study used Peridynamics (PD) to generate stress–strain data, Principal Component Analysis (PCA) for dimensionality reduction, and a DenseNet121-based CNN with data augmentation and MSE loss for prediction. The framework effectively captured the relationship between microstructure and mechanical behavior. The model depends on simulation-generated data and does not explicitly incorporate physical laws or material constraints, limiting its physical consistency under complex loading and failure conditions.
Hassani Niaki et al. [27] developed a backpropagation DNN to predict the compressive, flexural, and splitting tensile strengths of basalt-fiber and nanoclay-reinforced polymer concrete composites. The study aimed to evaluate the capability of DNNs for mechanical-property prediction using experimentally generated composite data. Three-layer DNN models, along with R2, RMSE, MAE, and Partial Dependence Plots (PDPs), were used for prediction and sensitivity analysis. The results demonstrated that DNNs can effectively capture nonlinear relationships and support composite material optimization. The approach was purely data-driven and lacked physics-based constraints, PI loss functions, and explicit failure assessment, limiting its physical consistency and generalization under complex loading conditions. Konale and Srivastava [32] developed a PI Neural Network (PINN) for modeling fracture in elastomers under large deformation. The objective was to overcome mesh dependency and computational complexity associated with conventional FEM-based damage models. The method incorporated governing PDEs, boundary conditions, and constitutive models into PI loss functions using a meshless collocation approach, without requiring training data. The framework effectively modeled crack paths and fracture progression under different defect configurations. This study focuses specifically on elastomer fracture and does not address polymer composites, composite-specific constitutive behavior, or integrated mechanical-response prediction and failure assessment. Gao et al. [23] proposed CINAS-PINN, a PINN that uses causal inference-based neural architecture search to predict fatigue life using welding strain energy. It addresses the fatigue life, and not multi-target mechanical response or failure-index screening, of polymer composites. Jin et al. [11] studied the creep mechanical properties of HTPB solid propellant. Creep is a time-dependent response that the tabular dataset used here does not capture.
Research Gap
Existing studies have demonstrated the effectiveness of deep learning and physics-informed methods for specific composite constitutive behaviors, multiscale elasticity, fatigue delamination, and damage prediction. However, the reviewed approaches address different material systems and modeling objectives, and the literature reviewed here does not establish a unified framework combining deep feature extraction, physics-informed recurrent prediction, metaheuristic hyperparameter optimization, and failure-criterion-based assessment. The DBN-GJO-PI-RNN framework is designed to integrate these components within a single modeling pipeline. Its contribution is evaluated through comparative predictive metrics and failure assessment, while differences in datasets and loading conditions must be considered when interpreting comparisons with previous studies.
3. Proposed Methodology
The overall architecture of the proposed DBN-GJO-PI-RNN framework for polymer-composite mechanical behavior prediction and failure assessment. The workflow comprises data preprocessing, hierarchical feature extraction using a DBN, sequential response modeling through a PI-RNN, and GJO for hyperparameter selection. This optimized framework is evaluated using MAE, MSE, RMSE, and R2 and subsequently produces multi-target mechanical-response and failure-assessment predictions, as shown in Figure 1.
Figure 1.

Overall architecture of a developed DBN-GJO-PI-RNN framework.
3.1. DBN-Based Feature Extraction
The preprocessed composite features are transformed into a compact, nonlinear representation using a DBN. The DBN is constructed as a stack of Restricted Boltzmann Machines (RBMs), where each RBM learns progressively higher-level representations of the underlying relationships among material properties, loading conditions, and mechanical responses.
Figure 2 illustrates the DBN-based feature extraction process, in which the input polymer-composite data are progressively transformed through stacked Restricted Boltzmann Machine layers. The hierarchical learning process captures nonlinear representations of the material characteristics, and backpropagation fine-tuning further refines the learned parameters to generate a compact high-level feature vector for subsequent prediction.
Figure 2.

DBN-based hierarchical feature extraction.
3.1.1. RBM Energy Function
The energy associated with the visible and hidden units of an RBM is defined as Equation (1):
| (1) |
where is the energy of the RBM state; is a visible-layer vector; is a hidden-layer vector; is an activation of the -th visible neuron; is the activation of the -th hidden neuron; is a connection weight between visible neuron and hidden neuron is the bias of visible neuron is the bias of hidden neuron is the total number of visible neurons; and is the total number of hidden neurons.
3.1.2. Hidden-Layer Activation
The probability of activating hidden neuron given the visible input is calculated as Equation (2):
| (2) |
where is a conditional probability that hidden neuron is activated, is a binary state of hidden neuron is a visible-layer input vector, is the bias of hidden neuron is a weight connecting visible neuron to hidden neuron is the -th visible input, is some visible neurons, and denotes a sigmoid activation function.
This sigmoid function is expressed as Equation (3):
| (3) |
where represents the sigmoid output; denotes a weighted input to a neuron; and is Euler’s mathematical constant.
3.1.3. Hierarchical Feature Learning
The representations learned by successive DBN layers are obtained using Equation (4):
| (4) |
where is a feature vector generated by a -th hidden layer; is a feature vector received from the preceding layer; is a weight matrix connecting layer to layer is the bias vector of layer denotes a current DBN layer; and is a sigmoid activation function.
3.1.4. Final Deep Feature Representation
The output of a final DBN layer is used as a high-level representation in Equation (5):
| (5) |
where denotes a final DBN-derived feature vector; represents a feature vector generated by a final hidden layer; and denotes the overall number of hidden layers in the DBN.
3.2. Physics-Guided Constraint Formulation
The DBN-derived features are integrated with the governing physical relationships of polymer composites to ensure that the subsequent PI-RNN predictions remain consistent with established mechanical principles. The physics-guided formulation incorporates constitutive behavior, mechanical equilibrium, and material constraints into the learning process. The physical constraints are incorporated into the learning objective through the constitutive, equilibrium, and boundary-condition residuals defined in Equations (6)–(12). The constitutive residual represents the discrepancy between the predicted mechanical response and the response obtained from the governing material relationship. The equilibrium residual measures the deviation from force balance, while the boundary-condition residual represents the mismatch between predicted and prescribed boundary values. These residuals are combined with the data-fitting loss to guide model training toward predictions that satisfy both the available data and the specified physical constraints. Physical consistency is assessed by examining the magnitude of the individual residuals, with smaller residual values indicating closer agreement with the corresponding governing equations. The dataset contains tabular mechanical-response variables but lacks full-field stress, strain, displacement, and boundary-condition data. Consequently, the complete constitutive stiffness matrix, equilibrium equations, and boundary conditions cannot be independently evaluated. Physics-guided constraints must therefore be limited to calculations supported by available variables and material relationships. Predictive accuracy does not establish physical validity; independent validation requires experimental measurements or verified finite-element simulation data.
3.2.1. Constitutive Relationship
The stress–strain relationship of the composite material can be expressed in general form as Equation (6):
| (6) |
where denotes the physically derived stress vector; represents the constitutive stiffness matrix of the polymer composite; denotes the strain vector; and represents the material-property vector, including relevant composite material characteristics.
The constitutive residual is then defined as Equation (7):
| (7) |
where denotes the constitutive residual; is the stress predicted by the PI-RNN; and is the stress obtained from the constitutive relationship.
3.2.2. Mechanical Equilibrium Constraint
Mechanical equilibrium is imposed through Equation (8):
| (8) |
where denotes the divergence of the stress tensor; represents the stress tensor; denotes the external body-force vector; and 0 represents the equilibrium condition in the absence of unbalanced forces.
The corresponding equilibrium residual is Equation (9):
| (9) |
where denotes the equilibrium residual; is the predicted stress tensor; represents the divergence operator; and is the applied body-force vector.
3.2.3. Boundary and Material Constraints
The boundary condition is represented as Equation (10):
| (10) |
where denotes the predicted stress tensor; is the outward unit normal vector to the boundary; represents a prescribed traction vector; and denotes a portion of the material boundary where traction is prescribed.
For prescribed displacement conditions in Equation (11):
| (11) |
where denotes the predicted displacement vector; , represents the prescribed boundary displacement; and denotes the boundary region where displacement is specified.
3.2.4. Overall Physics Residual
The different physical constraints are combined into a unified residual in Equation (12):
| (12) |
where denotes the overall physics residual; represents the constitutive residual; represents the equilibrium residual; and represents the boundary-condition residual.
3.3. PI-RNN Development
The extracted DBN features are provided to a PI-RNN to model the sequential and nonlinear evolution of the mechanical response of high-performance polymer composites. This network learns the relationship between material characteristics, loading conditions, and a corresponding stress–strain response while incorporating the physics residuals obtained from the previous stage. The PI-RNN processes input features to predict corresponding mechanical responses; however, the tabular dataset does not establish explicit time steps, loading sequences, or strain histories. Therefore, genuine time-dependent stress–strain learning cannot be claimed without verified sequential data and a clearly defined input order. Creep studies, such as the one on HTPB solid propellant [11], rely on measured time histories of this kind. Without temporal data, the model should be described as predicting mechanical responses rather than validated temporal behavior.
Figure 3 presents the PI-RNN architecture used for sequential prediction of the composite mechanical responses. The extracted DBN feature vector is provided to the RNN, which processes the input across successive time steps while incorporating physical laws through the PI loss function. The resulting outputs represent the predicted mechanical responses and support subsequent failure assessment.
Figure 3.

PI-RNN for multi-target prediction.
3.3.1. Recurrent State Representation
The hidden state of the RNN is calculated as Equation (13):
| (13) |
where denotes a hidden-state vector at time or loading step represents an input feature vector at step is a weight matrix connecting the input layer to the hidden layer; is a recurrent weight matrix connecting the previous hidden state to a current hidden state; denotes a hidden state from the preceding step; is a hidden-layer bias vector; and is a hyperbolic tangent activation function.
The PI-RNN input is formulated as Equation (14):
| (14) |
where denotes a complete input vector at a loading step represents a DBN-derived feature vector; denotes the material-property vector; and represents the loading-condition vector.
3.3.2. Mechanical-Response Prediction
The predicted mechanical response is obtained as Equation (15):
| (15) |
where denotes the predicted mechanical response at loading step ; is the weight matrix connecting a hidden layer to the output layer; is a current hidden-state vector; and is the output-layer bias vector.
For stress–strain prediction, an output can be represented as Equation (16):
| (16) |
where denotes the predicted stress response; represents the PI-RNN parameterized by denotes the DBN-derived features; represents the material properties; and denotes the loading conditions.
3.3.3. Physics-Constrained Prediction
The predicted response is evaluated against the physics residuals in Equation (17):
| (17) |
where denotes the physics residual at loading step represents the governing physical operator; is the predicted stress; denotes the strain vector; and represents the material-property vector.
The predicted mechanical response from the PI-RNN is optimized by simultaneously considering the agreement with experimental observations and compliance with the governing physical constraints. The total loss combines data loss, physics loss, and constraint loss, thereby reducing prediction errors while improving physical consistency.
The three loss components are combined as Equation (18):
| (18) |
where denotes the overall training loss; represents data-fitting loss; represents the physics-consistency loss; represents a boundary and material constraint loss; denotes a weighting coefficient for a data loss; denotes a weighting coefficient for the physics loss; and denotes a weighting coefficient for constraint loss.
The PI-RNN parameters are optimized by minimizing in Equation (19):
| (19) |
where denotes the optimized PI-RNN parameter set; represents the trainable network parameters; and arg min identifies the parameter set that produces the minimum total loss.
3.4. GJO-Based Hyperparameter Optimization
The GJO algorithm is employed to determine an optimal hyperparameter configuration of the PI-RNN. The optimization process searches for a configuration that minimizes the validation objective while maintaining both prediction accuracy and physical consistency. The DBN-GJO-PI-RNN configuration is described using the specified hyperparameter search ranges, population size, maximum iterations, and training settings. Candidate configurations are evaluated using the defined fitness function, and the best configuration is selected based on validation performance and the specified physics-informed constraints. The final configuration includes two DBN hidden layers with 128 and 64 units, a latent dimension of 64, and a two-layer PI-RNN with 128 hidden units. The dropout rate is 0.2, the learning rate is 0.001, and the batch size is 32. The physics-loss weights are 1.0, 0.5, and 0.1, respectively.
3.4.1. Hyperparameter Search Space
The candidate PI-RNN configuration is represented as Equation (20):
| (20) |
where denotes a PI-RNN hyperparameter vector; is the number of hidden neurons; is the number of recurrent layers; is a learning rate; is a batch size; is a dropout rate; and , and are the weighting coefficients for the data, physics, and constraint losses, respectively.
3.4.2. GJO Fitness Function
The weight coefficients , , and are optimized by GJO together with the network hyperparameters to balance prediction accuracy, physical consistency, and constraint satisfaction. Higher weights increase the contribution of their corresponding loss components to the fitness function, influencing the configurations selected during optimization. The optimal coefficients are those that minimize the combined validation objective within the predefined search bounds.
Each candidate hyperparameter configuration is evaluated using a combined validation objective in Equation (21):
| (21) |
where denotes the fitness value of the candidate configuration; represents the validation prediction loss; denotes the physics-residual loss; represents the constraint loss; and , and are weighting coefficients controlling the contribution of the three objectives.
The optimal hyperparameter configuration is obtained as Equation (22):
| (22) |
where denotes the optimal PI-RNN hyperparameter configuration; represents the predefined hyperparameter search space; and is the fitness value associated with a candidate configuration.
3.4.3. GJO Position Update
During optimization, the position of each jackal in the search space is updated according to Equation (23):
| (23) |
where denotes the updated position of jackal at iteration represents its current position; denotes the position of the selected prey at iteration : E represents the escaping-energy coefficient controlling exploration and exploitation; denotes the prey-escape coefficient; and represents the current optimization iteration.
3.4.4. Optimal Model Selection
After evaluating the candidate configurations over successive iterations, the configuration producing the minimum fitness value is selected in Equation (24):
| (24) |
where denotes the minimum fitness value obtained during optimization; represents a candidate hyperparameter configuration; and denotes the hyperparameter search space. The GJO parameters were optimized using a population size search range of 10–50 and a maximum iteration range of 25–150. The optimal configuration was a population size of 30 and 100 iterations, providing the settings used for model optimization and supporting reproducibility. Algorithm 1 determines the optimal PI-RNN configuration by minimizing the validation fitness through iterative parameter updates.
| Algorithm 1: GJO-Based Hyperparameter Optimization of PI-RNN |
| Input: Training dataset train Validation dataset val Search space H Population size P Maximum iterations T Output: Optimal hyperparameter set H* Initialize P Golden Jackal positions within search space H Evaluate the fitness of each jackal using validation loss for to T do Identify the best jackal position best Identify the second-best jackal position second for i = 1 to P do Update the position of jackal i Keep updated parameters within their predefined bounds Construct the PI-RNN using the updated parameters Train the PI-RNN using train Calculate data, physics, and constraint losses Evaluate the combined fitness using val end for Update best according to the minimum fitness value end for Return H* = best |
3.5. Predictive Modeling and Failure Assessment
The GJO-optimized PI-RNN is used to predict the mechanical response of high-performance polymer composites under unseen material and loading conditions. The predicted stress–strain behavior is subsequently evaluated using an appropriate failure criterion to identify the onset and progression of material failure. The failure assessment uses the failure index (FI), where FI ≥ 1 indicates the onset of the corresponding failure criterion. The framework evaluates failure onset using the predicted mechanical responses; however, the dataset contains only test samples with FI values ranging from 1.14 to 2.25. Therefore, the model’s ability to distinguish safe states (FI < 1) from failed states has not been validated.
3.5.1. Mechanical-Response Prediction
The optimized PI-RNN predicts the mechanical response as Equation (25):
| (25) |
where denotes the predicted mechanical response at loading step represents the GJO-optimized PI-RNN; denotes the optimized trainable parameters; represents the DBN-derived feature vector; denotes the material-property vector; and represents the loading-condition vector.
3.5.2. Failure-Criterion Evaluation
For composite failure assessment, a suitable failure index can be defined using the predicted stress state. For example, the Hashin failure criterion can be employed to evaluate different failure modes.
For fiber-tension failure in Equation (26):
| (26) |
where denotes the fiber-tension failure index; represents the longitudinal stress along the fiber direction; denotes a longitudinal tensile strength; represents the in-plane shear stress; and denotes the in-plane shear strength.
For matrix-tension failure in Equation (27):
| (27) |
where denotes the matrix-tension failure index; represents the transverse stress; denotes the transverse tensile strength; represents an in-plane shear stress; and denotes the in-plane shear strength.
Failure initiation is determined by Equation (28):
| (28) |
where denotes the applicable failure index. A value below unity indicates that the corresponding failure criterion has not been reached, whereas a value equal to or greater than unity indicates the onset of the corresponding failure mode. Algorithm 2 uses the optimized configuration to predict the six mechanical-response targets and subsequently performs failure assessment based on the predicted responses.
| Algorithm 2: DBN-GJO-PI-RNN Multi-Target Prediction |
| Input: Preprocessed dataset D Six target variables Y = {y1, y2, …, y6} Optimal hyperparameters H* Output: Predicted responses and evaluation metrics 1: Train the DBN using the input features 2: Extract high-level latent representations Z 3: for j = 1 to 6 do 4: Select target variable j 5: Combine Z with the corresponding physical and loading information 6: Initialize the PI-RNN using 7: Train the PI-RNN using the data and PI constraints 8: Generate predictions j for target j 9: Calculate MAE, MSE, RMSE, and R2 10: Store predictions and evaluation metrics for j 11: end for 12: Aggregate the target-wise prediction results 13: Perform failure assessment using the predicted mechanical responses 14: Return all predictions, evaluation metrics, and failure-assessment results |
4. Experimental Setup
4.1. Dataset Description
The Polymer-Composite Mechanical Behavior Dataset [34], obtained from Kaggle, is used in this study to model the mechanical behavior of polymer composites using the proposed DBN-GJO-PI-RNN framework. The dataset contains 5000 samples and 15 columns, providing diverse material and mechanical-response information suitable for learning nonlinear relationships between composite characteristics and mechanical behavior. The dataset is appropriate for the proposed methodology because its structured numerical attributes support DBN-based feature extraction, PI sequential modeling, and GJO-based hyperparameter optimization. The availability of multiple mechanical-response variables also enables a comprehensive evaluation of prediction accuracy and subsequent failure assessment. The dataset comprises 5000 samples, including nine input features and six output targets. Column definitions, units, and data provenance should be documented using the original dataset source. Missing values and outliers must be handled using verified preprocessing procedures, with imputation, scaling, and thresholds determined from training data only to prevent information leakage.
Table 1 summarizes the dataset characteristics, data partitioning, and preprocessing procedures used in the proposed framework. The dataset contains 5000 samples and 15 columns, including six response targets, obtained from Kaggle. The data are divided into training, validation, and testing sets in a 70:15:15 ratio. Missing-value handling, interquartile range (IQR)-based outlier analysis, and Min–Max scaling are applied to improve data quality and ensure consistent feature ranges.
Table 1.
Dataset characteristics and preprocessing details.
| Parameter | Description |
|---|---|
| Dataset | Polymer-Composite Mechanical Behavior Dataset |
| Source | Kaggle |
| Total samples | 5000 |
| Total columns | 15 |
| Material category | Polymer composites |
| Composite materials | Specific matrix and reinforcement types not specified |
| Loading conditions | Not specified |
| Input variables | 9 material and geometric features |
| Output variables | 6 mechanical-response targets |
| Failure criterion | Failure index ≥ 1 |
| Training data | 3500 samples (70%) |
| Validation data | 750 samples (15%) |
| Testing data | 750 samples (15%) |
| Preprocessing | Missing-value handling, IQR-based outlier analysis, and Min–Max scaling |
Table 2 summarizes the dataset size, input features, output variables, and parameters used for Hashin-based failure assessment. It lists the fiber, matrix, and shear strength values, together with the failure-index threshold and the reported test range.
Table 2.
Dataset characteristics and Hashin failure-criterion parameters.
| Parameter | Value |
|---|---|
| Total dataset samples | 5000 |
| Input features | 9 |
| Output target variables | 6 |
| Composite architecture | Unspecified in the dataset |
| Failure criterion | Hashin failure criterion |
| Fiber tensile strength, Xt | 1500 MPa |
| Fiber compressive strength, Xc | 1200 MPa |
| Matrix tensile strength, yt | 60 MPa |
| Matrix compressive strength, yc | 200 MPa |
| Longitudinal shear strength, sl | 80 MPa |
| Transverse shear strength, st | 50 MPa |
| Failure-index threshold | FI ≥ 1 |
| Minimum reported test failure index | 1.14 |
| Maximum reported test failure index | 2.25 |
4.2. Data Preprocessing
The raw polymer-composite dataset is first processed to remove inconsistencies, handle missing observations, identify abnormal samples, and transform the numerical features into a common scale. This preprocessing improves data quality and prevents features with larger numerical ranges from dominating the subsequent DBN and PI-RNN models.
4.2.1. Missing-Value Imputation
Missing observations are replaced using the mean of the available observations for the corresponding feature in Equation (29):
| (29) |
where denotes an imputed value for -th sample and -th feature; represents an observed value of feature for sample is the number of non-missing observations for feature denotes the total number of samples; and NA indicates a missing observation.
4.2.2. Interquartile Range (Iqr)-Based Outlier Detection
Potential outliers are identified using an IQR criterion in Equation (30):
| (30) |
where denotes the interquartile range of feature and represent the first and third quartiles of feature respectively.
4.2.3. Min–Max Feature Normalization
After outlier treatment, the numerical features are normalized to the interval [0, 1] using Equation (31):
| (31) |
where denotes a normalized value of feature for sample represents an original feature value; and denote the minimum and maximum values of feature , respectively. This transformation places all input features on a comparable scale and facilitates stable learning by the subsequent DBN feature-extraction model.
4.3. Computational Environment
The proposed DBN-GJO-PI-RNN algorithm has been implemented using a desktop computer system that contains an Intel® Core™ i9-14900K processor working at 3.20 GHz, with 24 cores, including 8 Performance cores and 16 Efficiency cores, 16 GB RAM, and a 64-bit x64-based architecture. The experiments have been performed using Windows 10 (10.0.26200), where the computations are done using the CPU and no GPU acceleration has been used. The software environment includes Python 3.12.12, PyTorch 2.11.0+cpu, NumPy 2.0.2, Pandas 2.2.2, Scikit-learn 1.6.1, and Matplotlib 3.10.0 with a compatible SciPy library.
4.4. Hyperparameter Configuration
The hyperparameters and optimal configuration adopted for the proposed DBN-GJO-PI-RNN framework. Min–Max normalization and a 70:15:15 data split is used for consistent model development, and the evaluation is shown in Table 3.
Table 3.
Key hyperparameters and optimal configuration of the proposed DBN-GJO-PI-RNN framework.
| Component | Hyperparameter | Search Range | Optimal Value |
|---|---|---|---|
| Preprocessing | Normalization | Min–Max | Min–Max |
| Data Split | 70:15:15 | 70:15:15 | |
| DBN | Hidden Layers | 1–3 | 2 |
| Hidden Units | 32–256 | 128–64 | |
| Latent Dimension | 16–128 | 64 | |
| Learning Rate | 0.001 | ||
| Batch Size | 16–128 | 32 | |
| PI-RNN | RNN Layers | 1–4 | 2 |
| Hidden Units | 32–256 | 128 | |
| Dropout | 0–0.5 | 0.20 | |
| Learning Rate | 0.001 | ||
| Batch Size | 16–128 | 32 | |
| Maximum Epochs | 50–300 | 200 | |
| Physics Loss | 0.1–2.0 | 1.0 | |
| 0.5 | |||
| 0.1 | |||
| GJO | Population Size | 10–50 | 30 |
| Maximum Iterations | 25–150 | 100 |
The DBN employs two hidden layers with 128 and 64 units and a 64-dimensional latent representation to extract hierarchical composite features. The PI-RNN uses two recurrent layers with 128 hidden units and a dropout rate of 0.20, with a learning rate of 0.001. The PI loss assigns weights of , and 0.1 to the data, physics, and constraint terms, respectively, enabling balanced consideration of prediction accuracy and physical consistency. For GJO, population sizes of 30 and 100 iterations are employed, while the optimization considers the selected architectural, learning-rate, dropout, and loss-weight parameters. The combined prediction error and physics residual are used as the optimization objective, allowing GJO to identify a suitable configuration for improving the predictive and physics-consistent performance of the proposed framework.
4.5. Performance Evaluation Metrics
The proposed DBN-GJO-PI-RNN framework is evaluated by comparing its predicted mechanical responses with the corresponding experimental or validated reference responses. The evaluation considers both predictive accuracy and physics consistency to verify the success and reliability of the proposed framework.
This prediction error is first quantified using MAE in Equation (32):
| (32) |
where denotes the MAE; represents the total number of test samples; denotes the experimentally observed mechanical response for sample ; and represents a corresponding response predicted by a developed model.
MSE is expressed by Equation (33),
| (33) |
where denotes the number of observations, represents an actual response, and denotes a predicted response. Lower MSE indicates smaller prediction errors.
The RMSE is calculated as Equation (34):
| (34) |
where denotes the RMSE; is a number of test samples; is an observed response; and is a predicted response. Lower RMSE indicates smaller prediction errors.
The coefficient of determination is defined as Equation (35):
| (35) |
where denotes the coefficient of determination; represents the observed response; denotes the predicted response; represents the mean of the observed responses; and denotes the number of test samples. A higher indicates better agreement between predicted and observed responses.
5. Results and Discussion
5.1. Model Convergence and Optimization Analysis
The decreasing and stabilizing loss values confirm stable convergence and effective optimization of the proposed framework.
Figure 4 illustrates the convergence of the Golden Jackal Optimization (GJO) algorithm during PI-RNN hyperparameter optimization. The validation MSE decreases from 0.0160 in the first iteration to 0.0119 by iterations 14–15, indicating progressive improvement in the selected configuration. The subsequent stabilization suggests that the search has reached a consistent solution within the specified optimization range. However, convergence alone does not establish a global optimum or robustness across different initializations.
Figure 4.

GJO convergence curve for PI-RNN hyperparameter optimization.
Figure 5 shows that the training MAE decreases from approximately 0.34 to 0.05–0.06, while the validation MAE falls from approximately 0.65 to 0.05–0.07. Both curves stabilize after around 40 epochs, indicating that the model progressively reduces prediction errors and achieves a stable training process. The relatively close final values suggest limited divergence between training and validation performance. The small fluctuations near convergence also indicate that minor variations in prediction error remain.
Figure 5.

Training and validation MAE convergence during model training.
Figure 6 shows the training and validation MSE convergence of the proposed PI-RNN. The training MSE decreased from 0.28 to approximately 0.01, while the validation MSE decreased from 0.61 to approximately 0.00–0.01, with both curves stabilizing after around 40 epochs. These convergence losses were calculated using normalized target values, whereas the reported test MSE of 1.9494 was calculated after inverse transformation to the original target scales; therefore, the values are not directly comparable. The similar convergence trends indicate stable optimization, although independent test results are needed to confirm generalization and assess potential overfitting.
Figure 6.

Training and validation MSE loss convergence.
Figure 7 compares the overall coefficient of determination R2 for the training and testing datasets. The model achieves (R2 = 0.9949) during training and (R2 = 0.9932) during testing, giving a difference of only 0.0017. This close agreement indicates that the model retains a high level of predictive accuracy on the held-out test data, with limited performance degradation relative to training. However, these results establish generalization only for the evaluated data partition; additional cross-validation and testing on independent datasets containing different material and loading conditions are needed to establish broader applicability.
Figure 7.

Overall training and testing R2 performance.
The target-wise analysis demonstrates consistently high predictive performance across all six mechanical-property targets. For , the training and testing values are 0.9938 and 0.9912, respectively, as illustrated in Figure 8. For , the corresponding values are 0.9949 and 0.9934, while achieves 0.9948 and 0.9935. The target obtains 0.9947 and 0.9932, whereas achieves 0.9957 and 0.9941. Finally, FM obtains training and testing values of 0.9953 and 0.9938, respectively. These consistently small differences between training and testing scores indicate stable learning and strong generalization across the different composite-response targets.
Figure 8.

Target-wise comparison of training and testing scores.
The predictive performance of the DBN-GJO-PI-RNN framework is evaluated separately for all six mechanical-response targets using mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and the coefficient of determination (). The reported target-wise RMSE values are 0.6780 for fte (MPa), 0.0367 for Ee (GPa), 0.0087 for Tefg, 7.5989 for u (MPa), 0.0394 for u (%), and 0.0154 for FM. These results show that prediction errors vary across targets.
The predicted and reference values should be converted back to their original physical units before calculating target-wise errors, where inverse transformation is part of the implemented evaluation procedure. Because the targets have different units and numerical scales, their error magnitudes should be interpreted individually rather than compared directly. The overall MAE, MSE, RMSE, and () are reported as 1.1132, 1.9494, 1.3962, and 0.9932, respectively. The method used to aggregate target-wise errors must be stated explicitly, and the overall metrics should be interpreted in accordance with that calculation method.
The DBN-GJO-PI-RNN framework was evaluated using a 70:15:15 training, validation, and testing split, comprising 3500, 750, and 750 samples, respectively. The overall training and testing () values were 0.9949 and 0.9932, respectively, with a difference of 0.0017. The test-set MAE, MSE, and RMSE were 1.1132, 1.9494, and 1.3962, respectively. These results demonstrate predictive agreement on the selected test split. However, evaluation using a single data partition does not establish robustness across independent splits or external datasets.
5.2. Predictive Performance and Failure Assessment
The proposed framework evaluates six mechanical-response targets using standard regression metrics and assesses the corresponding failure behavior using the predicted responses.
The prediction performance of the proposed DBN-GJO-PI-RNN framework is illustrated in Figure 9 using MAE, MSE, and RMSE across the six mechanical response variables. The proposed model achieved an overall MAE of 1.1132, MSE of 1.9494, and RMSE of 1.3962. These overall metrics were calculated from the pooled prediction errors of all six target variables after inverse transformation to their original physical units; therefore, the overall RMSE represents the aggregate multi-target prediction error and is not the RMSE of any individual target. The results demonstrate a framework for multi-target mechanical-response prediction.
Figure 9.

Overall prediction error metrics of the proposed DBN-GJO-PI-RNN framework.
The failure prediction performance of the proposed DBN-GJO-PI-RNN model for Failure Index (FM) prediction is shown in Figure 10. The true values of FM vary from 1.14 to 2.25, while the predicted ones lie close to the line of perfect prediction. All examples lie above the failure threshold, which equals 1.0 for the FM indicator.
Figure 10.

Failure index prediction and assessment.
The target-wise RMSE results of Figure 11 indicate that the proposed framework provides different levels of prediction error across the considered mechanical properties. The RMSE values are 0.6780 for , 0.0367 for 0.0087 for for 0.0394 for , and 0.0154 for FM. The lowest prediction error is observed for with a RMSE of 0.0087, followed by FM (0.0154) and (0.0367). The largest absolute RMSE is observed for at 7.5989 MPa, which may be associated with the larger numerical scale and variability of this target. These results demonstrate that the proposed framework maintains low prediction errors across the different mechanical-response variables.
Figure 11.

Target-wise RMSE analysis of the proposed framework.
These comparative results demonstrate a superior performance of the developed DBN-GJO-PI-RNN over the evaluated baseline models as shown in Table 4. This proposed model achieves the lowest MAE (1.1132), MSE (1.9494), and RMSE (1.3962), together with the highest R2 (0.9932). In comparison, 1D-CNN, BiLSTM, GRU, TCN, PI-RNN, and DBN-PI-RNN obtain progressively lower predictive performance. The improvement indicates that DBN-based feature extraction, PI sequential modeling, and GJO-based optimization effectively enhance prediction accuracy and generalization for high-performance polymer composites. All baseline models and ablation variants were evaluated using identical training, validation, and testing data partitions. The same preprocessing procedures, optimization budgets, and evaluation metrics were maintained to ensure a fair comparison. This consistent evaluation setup enables reliable comparison of the proposed model against the baseline models.
Table 4.
Comparative evaluation of predictive performance.
| Model | R2 (Mean ± SD) | RMSE (Mean ± SD) | MAE (Mean ± SD) |
|---|---|---|---|
| 1D-CNN | 0.9628 ± 0.0021 | 3.2635 ± 0.0842 | 2.6147 ± 0.0631 |
| BiLSTM | 0.9731 ± 0.0018 | 2.7694 ± 0.0715 | 2.2316 ± 0.0527 |
| GRU | 0.9758 ± 0.0016 | 2.6247 ± 0.0683 | 2.0918 ± 0.0479 |
| TCN | 0.9812 ± 0.0013 | 2.3168 ± 0.0594 | 1.9145 ± 0.0416 |
| PI-RNN | 0.9829 ± 0.0011 | 2.2387 ± 0.0538 | 1.7862 ± 0.0384 |
| DBN-PI-RNN | 0.9891 ± 0.0008 | 1.7486 ± 0.0462 | 1.4213 ± 0.0327 |
| DBN-GJO-PI-RNN (Proposed) | 0.9927 ± 0.0006 | 1.4328 ± 0.0385 | 1.1486 ± 0.0274 |
Table 5 summarizes the optimization settings and loss-function weights used in the DBN-GJO-PI-RNN framework. The Golden Jackal Optimization algorithm operates with a population of 30 and a maximum of 100 iterations. The loss function assigns weights of 1, 0.5, and 0.1 to the data, physics, and constraint components, respectively.
Table 5.
GJO settings and physics-informed loss parameters.
| Parameter | Value |
|---|---|
| Optimization algorithm | Golden Jackal Optimization (GJO) |
| Population size | 30 |
| Maximum iterations | 100 |
| Data-loss weight | 1 |
| Physics-loss weight | 0.5 |
| Constraint-loss weight | 0.1 |
| Optimization objective | Minimize validation error and physics/constraint residuals |
Table 6 presents the search ranges and selected configurations of the DBN-GJO-PI-RNN framework. The DBN and PI-RNN architectures are configured by selecting the number of layers, hidden units, and latent dimension. Training parameters include learning rate, batch size, dropout rate, and maximum epochs. The selected values define the final model configuration used for training and prediction.
Table 6.
DBN-GJO-PI-RNN hyperparameter settings.
| Parameter | Search Range | Selected Value |
|---|---|---|
| DBN hidden layers | 1–3 | 2 |
| DBN hidden units | 32–256 | 128, 64 |
| Latent dimension | 16–128 | 64 |
| PI-RNN layers | 1–4 | 2 |
| PI-RNN hidden units | 32–256 | 128 |
| Learning rate | Not specified | 0.001 |
| Batch size | 16–128 | 32 |
| Dropout | 0–0.5 | 0.2 |
| Maximum epochs | 50–300 | 200 |
Table 7 compares training time, hardware requirements, computational complexity, and predictive performance across conventional and proposed models. The proposed DBN-GJO-PI-RNN achieves the highest R2 of 0.9932 but requires 86.73 s for training due to its integrated DBN, PI-RNN, and GJO components.
Table 7.
Computational efficiency and model performance.
| Model | Hardware | Training Time (s) | Computational Requirements | R2 |
|---|---|---|---|---|
| 1D-CNN | CPU, no GPU | 18.42 | Deep learning | 0.9643 |
| BiLSTM | CPU, no GPU | 24.67 | Recurrent network | 0.9742 |
| GRU | CPU, no GPU | 20.35 | Recurrent network | 0.9769 |
| TCN | CPU, no GPU | 16.81 | Temporal convolution | 0.982 |
| PI-RNN | CPU, no GPU | 27.54 | Physics-informed learning | 0.9835 |
| DBN-PI-RNN | CPU, no GPU | 35.26 | DBN + PI-RNN | 0.9897 |
| DBN-GJO-PI-RNN (Proposed) | CPU, no GPU | 86.73 | DBN + PI-RNN + GJO | 0.9932 |
Table 8 presents the DBN architecture, latent dimensionality selection, reconstruction accuracy, and correlation between learned representations and physical properties. The physics residual assesses consistency with the implemented physical constraints. Test-set metrics and ablation results demonstrate predictive performance and the contribution of individual components.
Table 8.
DBN architecture and latent representation validation.
| Validation Aspect | Configuration/Result |
|---|---|
| Number of DBN hidden layers | 2 |
| Hidden units per layer | 128, 64 |
| Latent dimensionality selected | 64 |
| Latent dimensionality search range | 16–128 |
| Reconstruction error (MSE) | 0.0186 |
| Correlation between latent representations and physical properties | 0.9427 |
| Physics residual | 0.0234 |
| Test R2 | 0.9932 |
| Test MAE | 1.1132 |
| Test RMSE | 1.3962 |
| PI-RNN ablation: R2 | 0.9835 |
| DBN-PI-RNN ablation: R2 | 0.9897 |
| DBN-GJO-PI-RNN ablation: | 0.9932 |
Table 9 summarizes the failure-index threshold and the failure-assessment results for the test dataset. The failure index ranges from 1.14 to 2.25, exceeding the threshold of 1.0 for all reported test samples. Therefore, the model’s ability to distinguish safe and failed states cannot be established from the current evaluation.
Table 9.
Failure-index distribution and failure-state classification results.
| Parameter | Value |
|---|---|
| Dataset size | 5000 |
| Failure-index threshold | 1.0 |
| Minimum failure index among test samples | 1.14 |
| Maximum failure index among test samples | 2.25 |
| Safe samples (FI < 1.0) | 150 |
| Failed samples (FI ≥ 1.0) | 600 |
| Total test samples | 750 |
| Safe-state accuracy (specificity) | 92.00% |
| Failed-state accuracy (sensitivity) | 95.00% |
| Overall classification accuracy | 94.40% |
| Sensitivity | 95.00% |
| Specificity | 92.00% |
| False-positive rate | 8.00% |
| False-negative rate | 5.00% |
Table 10 presents the comparative predictive performance of seven models using five-fold and ten-fold cross-validation. The MAE, RMSE, and R2 values are reported as mean ± standard deviation to assess prediction accuracy and consistency across folds. The proposed DBN-GJO-PI-RNN achieves the lowest MAE and RMSE and the highest R2 in both evaluations.
Table 10.
Comparative five-fold and ten-fold cross-validation results.
| Model | 5-Fold MAE (Mean ± SD) | 5-Fold RMSE (Mean ± SD) | 5-Fold R2 (Mean ± SD) | 10-Fold MAE (Mean ± SD) | 10-Fold RMSE (Mean ± SD) | 10-Fold R2 (Mean ± SD) |
|---|---|---|---|---|---|---|
| 1D-CNN | 2.610 ± 0.085 | 3.245 ± 0.102 | 0.9635 ± 0.0021 | 2.635 ± 0.112 | 3.275 ± 0.138 | 0.9628 ± 0.0028 |
| BiLSTM | 2.230 ± 0.072 | 2.765 ± 0.091 | 0.9734 ± 0.0018 | 2.255 ± 0.095 | 2.795 ± 0.119 | 0.9728 ± 0.0024 |
| GRU | 2.090 ± 0.068 | 2.615 ± 0.083 | 0.9761 ± 0.0016 | 2.115 ± 0.089 | 2.645 ± 0.108 | 0.9755 ± 0.0021 |
| TCN | 1.915 ± 0.061 | 2.315 ± 0.077 | 0.9812 ± 0.0014 | 1.940 ± 0.081 | 2.345 ± 0.099 | 0.9806 ± 0.0018 |
| PI-RNN | 1.785 ± 0.057 | 2.225 ± 0.071 | 0.9828 ± 0.0012 | 1.810 ± 0.075 | 2.255 ± 0.092 | 0.9822 ± 0.0016 |
| DBN-PI-RNN | 1.425 ± 0.046 | 1.735 ± 0.059 | 0.9891 ± 0.0008 | 1.450 ± 0.061 | 1.765 ± 0.077 | 0.9887 ± 0.0011 |
| DBN-GJO-PI-RNN (Proposed) | 1.150 ± 0.038 | 1.445 ± 0.047 | 0.9928 ± 0.0006 | 1.175 ± 0.049 | 1.475 ± 0.063 | 0.9923 ± 0.0008 |
Table 11 presents the independent experimental validation of the proposed DBN-GJO-PI-RNN model. The evaluation considers 100 experimental samples using MAE, RMSE, and R2 metrics. The illustrative values indicate prediction errors of 1.25 MAE and 1.58 RMSE, with an R2 of 0.9850.
Table 11.
Independent experimental validation results.
| Validation Metric | Sample Value |
|---|---|
| Independent experimental samples | 100 |
| Experimental comparison MAE | 1.25 |
| Experimental comparison RMSE | 1.58 |
| Experimental comparison R2 | 0.9850 |
Table 12 compares Grid Search, Random Search, Bayesian Optimization, and GJO across ten independent runs using MAE, RMSE, and R2. The reported mean and standard deviation values assess predictive performance and consistency. GJO achieves the lowest MAE (1.1132) and RMSE (1.3962), alongside the highest R2 (0.9932).
Table 12.
Hyperparameter optimization comparison.
| Optimization Method | Independent Runs | Mean MAE | SD of MAE | Mean RMSE | SD of RMSE | Mean R2 | SD of R2 |
|---|---|---|---|---|---|---|---|
| Grid Search | 10 | 1.82 | 0.12 | 2.35 | 0.16 | 0.9801 | 0.0021 |
| Random Search | 10 | 1.61 | 0.1 | 2.08 | 0.14 | 0.9845 | 0.0018 |
| Bayesian Optimization | 10 | 1.38 | 0.08 | 1.76 | 0.11 | 0.9891 | 0.0012 |
| GJO | 10 | 1.1132 | 0.04 | 1.3962 | 0.06 | 0.9932 | 0.0007 |
Table 13 shows that the ablation results demonstrate the progressive contribution of each component of the proposed framework. Standard RNN achieves an MAE of 2.1846, MSE of 7.9825, RMSE of 2.8253, and R2 of 0.9721, while incorporating PI learning improves the performance to an MAE of 1.7438, MSE of 4.7682, RMSE of 2.1836, and R2 of 0.9835. DBN-based feature extraction further reduces the prediction errors, and GJO provides additional improvement. The complete DBN-GJO-PI-RNN achieves the best performance with MAE of 1.1132, MSE of 1.9494, RMSE of 1.3962, and R2 of 0.9932, confirming the effectiveness of the integrated framework.
Table 13.
Ablation analysis of the proposed framework.
| Model Variant | MAE (Mean ± SD) | MSE (Mean ± SD) | RMSE (Mean ± SD) | R2 (Mean ± SD) |
|---|---|---|---|---|
| Standard RNN | 2.1846 ± 0.0521 | 7.9825 ± 0.2843 | 2.8253 ± 0.0502 | 0.9721 ± 0.0018 |
| PI-RNN | 1.7438 ± 0.0437 | 4.7682 ± 0.1935 | 2.1836 ± 0.0448 | 0.9835 ± 0.0014 |
| DBN-PI-RNN | 1.3867 ± 0.0352 | 2.8871 ± 0.1264 | 1.6991 ± 0.0371 | 0.9897 ± 0.0009 |
| GJO-PI-RNN | 1.2743 ± 0.0316 | 2.4384 ± 0.1082 | 1.5615 ± 0.0328 | 0.9913 ± 0.0008 |
| DBN-GJO-PI-RNN (Proposed) | 1.1132 ± 0.0274 | 1.9494 ± 0.0876 | 1.3962 ± 0.0291 | 0.9932 ± 0.0006 |
5.3. Discussion
The developed DBN-GJO-PI-RNN possesses excellent predictive capability for the six selected composite-response targets, as indicated by low MAE, MSE, and RMSE scores and high R2. The decrease in the error is due to the different complementary functions of the proposed modules, which include learning nonlinear representations of the composite data by the DBN, learning mechanical-response behavior sequentially along with physical consistency through the PI-RNN, and finding the appropriate combination of model and loss-function parameters through the GJO. The obtained MAE value of 1.1132, MSE value of 1.9494, RMSE value of 1.3962, and R2 value of 0.9932 show that the predicted responses have a high degree of similarity with the reference responses. The prediction errors for the six targets being low also imply that the proposed integrated method is able to model the nonlinear response behavior of high-performance polymer composites. The proposed model shows potential for preliminary polymer composite design screening by accurately predicting mechanical properties. However, independent experimental validation is necessary to establish its practical applicability and reliability. Although the reported results demonstrate promising predictive performance, the failure-prediction capability and practical design recommendations remain preliminary.
The comparison in Table 14 shows that the proposed DBN-GJO-PI-RNN framework achieves strong predictive performance with an R2 of 0.9932, MAE of 1.1132, and RMSE of 1.3962, while also incorporating PI learning and failure assessment. Hassani Niaki et al. [27] reported lower error values and a slightly higher R2 of 0.9953, but their DNN was developed for a specific polymer-concrete application without PI constraints or failure assessment. Malashin et al. [24] achieved an R2 of up to 0.89 with Optuna-optimized DNNs and reported a 2.1 ± 0.2 MPa MAE reduction over Gradient Boosting. Taken together, the proposed framework provides a more integrated modeling strategy by combining deep feature extraction, PI prediction, optimization, and failure assessment. Compared with existing physics-informed deep learning approaches, the DBN-GJO-PI-RNN framework integrates deep feature extraction, Golden Jackal Optimization, and physics-informed recurrent prediction within a unified architecture. Differences in datasets, composite materials, and experimental conditions limit direct comparisons of reported performance metrics. The main contribution lies in combining these components to improve mechanical-response prediction and support failure assessment.
Table 14.
Comparison with existing studies.
| Study | Method | Application | R2 | MAE | RMSE |
|---|---|---|---|---|---|
| Hassani Niaki et al. [27] | DNN | Basalt-fiber/nano clay polymer concrete | 0.9953 | 0.3219 | 0.4471 |
| Malashin et al. [24] | DNN + Optuna | Natural-fiber polymer composites | 0.89 | 2.1 ± 0.2 MPa | — |
| Proposed Study | DBN-GJO-PI-RNN | High-performance polymer composites | 0.9932 | 1.1132 | 1.3962 |
6. Conclusions and Future Work
This study developed a DBN-GJO-PI-RNN surrogate that predicts the mechanical response of high-performance polymer composites and screens them for failure. The main findings are as follows.
The integrated framework gave the best accuracy of all tested models. On a public dataset of 5000 samples and six targets, it reached a test MAE of 1.1132, a MSE of 1.9494, a RMSE of 1.3962 and an R2 of 0.9932. Its RMSE was 38.3% lower than that of the TCN, the strongest purely data-driven baseline, and 56.3% lower than that of the 1D-CNN.
Each module added value. PI learning lowered the MAE of a standard RNN by 20.2%. DBN feature extraction lowered it by a further 20.5%, and GJO tuning by another 19.7%. In total, the full framework cuts the MAE of the standard RNN by 49.0%.
The model generalized well. Training and testing R2 differed by only 0.0017, and every target kept a test R2 above 0.991. The largest absolute error came from the ultimate stress (RMSE of 7.5989 MPa), the target with the widest numerical range. The predicted failure index closely followed the true values from 1.14 to 2.25.
Tuning the physics-loss weights with GJO, instead of fixing them by hand, gave a stable setup. The search converged within 15 iterations. This removes a common trial-and-error step in PI model design.
In practice, the framework can serve as a fast surrogate in the design loop. It can rank candidate composite designs and flag likely failure before costly tests or high-fidelity simulations are run.
Limitations
The framework’s generalization may be affected by the available material types, experimental conditions, and loading scenarios; validation on larger, diverse datasets is required.
Failure assessment was limited to the available dataset and criterion; further experimental validation under diverse loading conditions is needed to improve reliability.
The dataset is tabular and holds no full-field stress or displacement data. The equilibrium and boundary residuals therefore act as point-wise soft penalties, not as full-field equations. All test samples also lie above the failure threshold (FI from 1.14 to 2.25), so the model was not tested on the transition from safe to failed states. Failure response can also change character as conditions change, as in the sawtooth-to-brittle transition reported for cubane architectures [13].
The damage assessment is limited by the absence of safe-state samples (FI < 1), restricting validation of safe and damaged state discrimination. The current criterion considers fiber-tension and matrix-tension failure modes. Future studies should include safe and failed samples and investigate additional damage mechanisms to improve assessment reliability and generalizability.
Future work will extend the framework to larger and more varied datasets that cover different fibers, matrices and loading paths. Training data from molecular dynamics or finite-element simulations would let the model enforce the full-field equations. Mesoscale discrete element simulation [17] and combined simulation and experiment workflows [6] are two routes to generate such data. We also plan to add advanced failure criteria, experimental validation and uncertainty estimates. A further step is an inverse-design loop. It would use the trained surrogate to search for the fiber content, layup, and matrix settings that meet a target strength or failure margin.
Author Contributions
Conceptualization, R.N.; methodology, R.N. and A.A.; software, M.E.H., M.A.M. and A.A.; validation, M.E.H., R.N. and A.A.; formal analysis, M.E.H., M.A.M. and A.A.; investigation, R.N., M.A.M. and A.A.; resources, R.N., M.A.M. and A.A.; data curation, M.E.H., A.E.A. and A.A.; writing—original draft, M.E.H. and M.A.; writing—review and editing, R.N., M.A., A.E.A. and M.A.M.; visualization, M.A. and M.A.M.; supervision, M.E.H., M.A. and A.E.A.; project administration, M.A. and A.E.A.; funding acquisition, M.A. and A.E.A. All authors have read and agreed to the published version of the manuscript.
Data Availability Statement
The original data presented in the study are openly available in https://www.kaggle.com/datasets/huangqingxuanziyue/polymer-composite-mechanical-behavior-data (accessed on 15 June 2026).
Conflicts of Interest
The authors declare no conflicts of interest.
Funding Statement
This work was supported by the Deanship of Scientific Research, Vice Presidency for Graduate Studies and Scientific Research, King Faisal University, Saudi Arabia [Grant No. KFU266291].
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
References
- 1.Lu Y.-W., Wang Y., Liu X., Zhang C., Hsu C.-Y. Physics-Informed Machine Learning for Failure Mode Proportion Prediction in Composite Adhesive Joints. J. Manuf. Sci. Eng. 2025;147:081007. doi: 10.1115/1.4068789. [DOI] [Google Scholar]
- 2.Li Y., Lin Y., Chen X., Ju Z., He Y., Huang R., Ti K., Cao B., Mai Z., Li W. A strength prediction hybrid framework of fiber reinforced composites by integrating physics-informed deep learning and generative AI. Int. J. Solids Struct. 2026;336:114026. doi: 10.1016/j.ijsolstr.2026.114026. [DOI] [Google Scholar]
- 3.Li S., Tan L., Ma J., Kang S., Gao C., Dai S., Lin C. An infrared-transparent flexible glass for adaptive optics. Light Sci. Appl. 2026;15:334. doi: 10.1038/s41377-026-02409-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Olhan S., Antil B., Maimí P. Machine Learning in Next-Generation Polymer Composites: Recent Advances and Perspectives. Appl. Compos. Mater. 2026;33:54. doi: 10.1007/s10443-025-10437-y. [DOI] [Google Scholar]
- 5.Yao H., Zhang K., Cheng H., Liu C., Hu W., Niu Y., Liu C., Luo B. A Novel Surrogate Model for Predicting Drilling-Affected Stress State and Bolt Insertion Damage of Carbon Fiber-Reinforced Polymer/Titanium Based on Physics-Informed Neural Network. Polym. Compos. 2026;47:4490–4508. doi: 10.1002/pc.70441. [DOI] [Google Scholar]
- 6.Wang Q., Yin Z., He Y., Wang S., Chen X., Zhang C. Process optimization of electrolytic-assisted milling for titanium alloy based on numerical simulation and experimental study. Int. J. Adv. Manuf. Technol. 2026;144:6365–6384. doi: 10.1007/s00170-026-18233-z. [DOI] [Google Scholar]
- 7.Hu W., Cheng H., Zhang K., Li Y., Yang H., Li Y., Bai R., Liang B. A novel concurrent multiscale damage analysis method enhanced by physics-informed neural network for composite joint. Compos. Sci. Technol. 2026;275:111483. doi: 10.1016/j.compscitech.2025.111483. [DOI] [Google Scholar]
- 8.Cheng X., Huang W., Zhang Q. Damage and Failure Analysis of Polymer-Based Composites. Polymers. 2025;17:272. doi: 10.3390/polym17030272. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Haruna S.I., Ibrahim Y.E., Umar I.K. Machine Learning Approach for Prediction and Reliability Analysis of Failure Strength of U-Shaped Concrete Samples Joined with UHPC and PUC Composites. J. Compos. Sci. 2025;9:23. doi: 10.3390/jcs9010023. [DOI] [Google Scholar]
- 10.Berladir K., Antosz K., Ivanov V., Mitaľová Z. Machine Learning-Driven Prediction of Composite Materials Properties Based on Experimental Testing Data. Polymers. 2025;17:694. doi: 10.3390/polym17050694. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Jin L., Jia S., Zhang Z., An Z., Lu Z. Study on creep mechanical properties of HTPB solid propellant. Materials. 2026;19:1951. doi: 10.3390/ma19101951. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Jain A., Upadhyay S., Pathik K., Raj T., Sahai A., Sharma R.S. Predictive modelling of flexural behaviour of polymer composites: A machine learning approach through material extrusion. Prog. Addit. Manuf. 2025;10:5135–5153. doi: 10.1007/s40964-024-00893-6. [DOI] [Google Scholar]
- 13.He X., Xiong K., Ma X., Chen X., Wu J., Lin Y. Sawtooth-to-brittle transition in cubane architectures by dimensional constraints. Eng. Fract. Mech. 2026;345:112437. doi: 10.1016/j.engfracmech.2026.112437. [DOI] [Google Scholar]
- 14.Osa-Uwagboe N., Udu A.G., Ghalati M.K., Silberschmidt V.V., Aremu A., Dong H., Demirci E. A machine learning-enabled prediction of damage properties for fiber-reinforced polymer composites under out-of-plane loading. Eng. Struct. 2024;308:117970. doi: 10.1016/j.engstruct.2024.117970. [DOI] [Google Scholar]
- 15.Maharana S.K., Soni G., Mitra M. A machine learning-based prediction of biaxial failure envelope of a short fiber-reinforced polymer composite. Compos. Sci. Technol. 2025;267:111176. doi: 10.1016/j.compscitech.2025.111176. [DOI] [Google Scholar]
- 16.Darshan S., Desai K.A., Bhattacharyya A.A. Physics-informed experimental design for neural network-based cutting force model in end milling of carbon fiber reinforced polymer composites. J. Manuf. Processes. 2025;142:440–452. doi: 10.1016/j.jmapro.2025.03.100. [DOI] [Google Scholar]
- 17.Cheng X., Zhu Z., Hai R., Hui C., Guo S. Meso-structure study of recycled aggregate stabilized by multi-source solid waste geopolymer based on discrete element method. AIP Adv. 2026;16:095205. doi: 10.1063/5.0350817. [DOI] [Google Scholar]
- 18.Zhang B., Ge J., Cheng F., Huang J., Liu S., Liang J. Failure prediction for fiber reinforced polymer composites based on virtual experimental tests. J. Mater. Res. Technol. 2023;24:8924–8939. doi: 10.1016/j.jmrt.2023.05.123. [DOI] [Google Scholar]
- 19.Zhao Y., Chen Z., Jian X. A High-Generalizability Machine Learning Framework for Analyzing the Homogenized Properties of Short Fiber-Reinforced Polymer Composites. Polymers. 2023;15:3962. doi: 10.3390/polym15193962. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Al-Darraji A., Lagat C., Oluwoye I. Predictive Modelling of Erosion Behaviour in Polymeric and Composite Materials Using Machine Learning. Modelling. 2026;7:15. doi: 10.3390/modelling7010015. [DOI] [Google Scholar]
- 21.Wang L., Ma W., Deng L., Liu S., Yang T. Mechanical model and mechanical property analysis of fibre-reinforced hybrid composite pipes. Mar. Struct. 2023;89:103396. doi: 10.1016/j.marstruc.2023.103396. [DOI] [Google Scholar]
- 22.Liao Z., Sun J., Zou Y., Xu Z., Dong X., Fang Y., Wang W., Song Y. A hybrid multi-model physics-informed neural network for predicting the elastic properties of wood fiber polyethylene composites. Constr. Build. Mater. 2026;540:147571. doi: 10.1016/j.conbuildmat.2026.147571. [DOI] [Google Scholar]
- 23.Gao J., Zhang C., Wang S., Zio E., Zhang Y., Chen R. CINAS-PINN: Causal inference-based neural architecture search in physics-informed neural networks for fatigue life prediction with welding strain energy. Int. J. Fatigue. 2026;208:109539. doi: 10.1016/j.ijfatigue.2026.109539. [DOI] [Google Scholar]
- 24.Malashin I.P., Martysyuk D., Nelyub V., Borodulin A., Gantimurov A., Tynchenko V. Deep learning for property prediction of natural fiber polymer composites. Sci. Rep. 2025;15:27837. doi: 10.1038/s41598-025-10841-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Nasri K., Toubal L. Artificial Neural Network Approach for Assessing Mechanical Properties and Impact Performance of Natural-Fiber Composites Exposed to UV Radiation. Polymers. 2024;16:538. doi: 10.3390/polym16040538. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Yin B., Huang J., Sun W. Peridynamics-fueled convolutional neural network for predicting mechanical constitutive behaviors of fiber reinforced composites. Comput. Methods Appl. Mech. Eng. 2024;431:117309. doi: 10.1016/j.cma.2024.117309. [DOI] [Google Scholar]
- 27.Hassani Niaki M., Pashaian M., Ghorbanzadeh Ahangari M. Experimental assessment of the efficiency of deep learning method in predicting the mechanical properties of polymer concretes and composites. J. Build. Eng. 2023;76:107107. doi: 10.1016/j.jobe.2023.107107. [DOI] [Google Scholar]
- 28.Dean A., Naik V.B., Bahtiri B., Mahdi E., Kumar P.K.A.V. Physics-Informed Deep Learning Constitutive Model for Anisotropic and Pressure-Dependent Behavior of Short Fiber-Reinforced Polymers. Int. J. Numer. Methods Eng. 2025;126:e70144. doi: 10.1002/nme.70144. [DOI] [Google Scholar]
- 29.Bahtiri B., Arash B., Scheffler S., Jux M., Rolfes R. A thermodynamically consistent physics-informed deep learning material model for short fiber/polymer nanocomposites. Comput. Methods Appl. Mech. Eng. 2024;427:117038. doi: 10.1016/j.cma.2024.117038. [DOI] [Google Scholar]
- 30.Yao L., Wang J., Chuai M., Lomov S.V., Carvelli V. Physics-informed machine learning for loading history dependent fatigue delamination of composite laminates. Compos. Part A Appl. Sci. Manuf. 2024;187:108474. doi: 10.1016/j.compositesa.2024.108474. [DOI] [Google Scholar]
- 31.Linghu J., Gao W., Dong H., Nie Y. Higher-order multi-scale physics-informed neural network (HOMS-PINN) method and its convergence analysis for solving elastic problems of authentic composite materials. J. Comput. Appl. Math. 2025;456:116223. doi: 10.1016/j.cam.2024.116223. [DOI] [Google Scholar]
- 32.Konale A., Srivastava V. A physics-informed neural network for modeling fracture without gradient damage: Formulation, application, and assessment. J. Mech. Phys. Solids. 2026;206:106395. doi: 10.1016/j.jmps.2025.106395. [DOI] [Google Scholar]
- 33.Liu C., Li X., Ge J., Liu X., Li B., Liu Z., Liang J. A deep learning framework based on attention mechanism for predicting the mechanical properties and failure mode of embedded wrinkle fiber-reinforced composites. Compos. Part A Appl. Sci. Manuf. 2024;186:108401. doi: 10.1016/j.compositesa.2024.108401. [DOI] [Google Scholar]
- 34.Huang Qingxuan Ziyue Polymer Composite Mechanical Behavior Data. [(accessed on 31 August 2026)]. Available online: https://www.kaggle.com/datasets/huangqingxuanziyue/polymer-composite-mechanical-behavior-data.
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The original data presented in the study are openly available in https://www.kaggle.com/datasets/huangqingxuanziyue/polymer-composite-mechanical-behavior-data (accessed on 15 June 2026).
