A Hybrid Experimental–Machine Learning Framework for Designing Fire-Resistant Natural Fiber Composites

Abstract

This work presents an integrated experimental and machine learning study on the fire performance of sisal fiber-reinforced polyester composites treated with magnesium hydroxide as a flame retardant. A total of 43 small-scale fire resistance tests were conducted in a custom-built gas-fired furnace following ISO 834 and NCh935/2 standards. Key parameters—including fiber content, flame retardant proportion, catalyst, and accelerator—were correlated with burn time and mass loss. Linear regression revealed negligible to weak correlations, while nonlinear models (Random Forest, Support Vector Regression, and Deep Neural Network) showed improved predictive capacity. The Deep Neural Network achieved the best performance (MSE = 0.061, R² = 0.334). Experimental results confirm that magnesium hydroxide substantially increases burn time, whereas sisal fiber content alone has a minimal effect on fire resistance. This study highlights an affordable strategy for enhancing the fire safety of bio-based composites and demonstrates the potential of machine learning to optimize material formulations. Future research should expand the dataset and validate the models through standardized large-scale fire tests. However, the findings are limited to small-scale fire resistance tests under controlled laboratory conditions and should not be generalized to full-scale structural applications without further validation.

1. Introduction

Fire safety remains a critical design criterion for building materials, particularly given the increasing risk of urban fires driven by climate change and rapid urbanization [1,2]. Fiber-reinforced polymer (FRP) composites are attractive alternatives to conventional construction materials due to their low weight, high strength-to-weight ratio, and ease of fabrication [3,4,5]. However, their inherent flammability poses significant challenges for structural and non-structural applications, demanding enhanced fire resistance [6].

Recent advances have explored the fire behavior of natural fiber composites beyond empirical testing, emphasizing hybrid modeling strategies and multi-scale simulations. He et al. (2021) conducted a comprehensive review on the predictive modeling of flame-retardant FRPs using machine learning, identifying data sparsity and physical interpretability as major bottlenecks [7]. More recently, Tran et al. (2024) demonstrated the successful application of polymer informatics for predicting the flammability as well as the thermal, mechanical, and electrical properties of polymer composites, achieving high accuracy through comprehensive feature engineering and advanced machine learning architectures [8]. Furthermore, Jafari et al. (2024) presented machine learning frameworks for expediting next-generation fire-retardant polymer composites, highlighting the growing relevance of data-driven modeling in complementing small-scale fire experiments and optimizing flame-retardant formulations for sustainable composite development [9].

Various flame-retardant (FR) technologies have been explored to mitigate the fire hazard associated with FRPs. Halogenated compounds, once popular, have been largely replaced by eco-friendly inorganic fillers such as aluminum trihydrate (ATH), ammonium polyphosphate (APP), and magnesium hydroxide (Mg(OH)₂) due to environmental and regulatory concerns [1,10,11]. Recent studies highlight the synergistic effects of combining Mg(OH)₂ with APP or nanoclays to enhance charring and intumescence, further improving fire performance [12,13]. Additionally, novel intumescent systems and nano-engineered fillers (e.g., nanoclays, graphene oxide) have demonstrated promising flame retardancy with minimal impact on mechanical properties [14,15].

Natural fibers such as sisal, jute, and hemp have gained traction as reinforcements for sustainable composite panels, offering biodegradability and a reduced environmental footprint [16]. Nevertheless, these fibers contribute little to intrinsic flame resistance and often act as additional fuel if not properly treated [17,18]. Therefore, incorporating efficient flame retardants into bio-based FRP systems remains an active area of research [19].

Parallel to material innovations, computational approaches have emerged to complement experimental testing. Machine learning techniques have been increasingly adopted for predicting the mechanical and thermal behaviors of polymer composites, enabling rapid property estimation and optimization with limited experimental data [20,21]. Yan et al. (2023) presented self-enforcing machine learning approaches specifically designed for small datasets in flame retardancy prediction [22]. Some works have combined machine learning with computational fluid dynamics (CFD) simulations and fire dynamics solvers such as FireFOAM and FDS to model flame spread and smoke production in composite structures [23,24]. Despite these advances, few studies have integrated the experimental fire resistance testing of natural fiber composites with machine learning-based predictive frameworks.

This research addresses this gap by presenting a hybrid experimental–machine learning approach to study the fire performance of sisal fiber-reinforced polyester composites treated with magnesium hydroxide. The study demonstrates a practical framework for sustainable, fire-safe composite design by quantifying the effects of fiber, flame-retardant, catalyst, and accelerator contents on burn time, comparing linear and nonlinear machine learning models, and providing insights for rapid formulation screening that reduces experimental workload while guiding more targeted testing. The integrated approach represents an advancement over previous studies by combining cost-effective small-scale fire testing with data-driven predictive modeling specifically tailored for natural fiber composites.

2. Materials and Methods

2.1. Materials

The composite panels were manufactured using a thermosetting unsaturated polyester resin as the matrix. Natural sisal fiber served as the reinforcement, while magnesium hydroxide (Mg(OH)₂) was added as the flame retardant. Cobalt octoate acted as an accelerator, and methyl ethyl ketone peroxide (MEKP) was used as the catalyst to initiate polymer crosslinking at room temperature [6]. All chemicals were of laboratory-grade purity and complied with ISO 834 [25] and NCh935/2 [26] standards for fire testing materials. A precision electronic balance with an accuracy of ±0.01 g was employed for weighing the raw materials.

Magnesium hydroxide was selected as the flame retardant for this study based on several key factors. First, Mg(OH)₂ represents an environmentally friendly, halogen-free alternative to traditional flame retardants, addressing growing regulatory and environmental concerns [1,10,11]. Second, its dual-action flame-retardant mechanism operates through endothermic decomposition (Mg(OH)₂ → MgO + H₂O) that absorbs heat and releases water vapor to dilute combustible gases while simultaneously forming a protective magnesium oxide layer that insulates the polymer matrix from further thermal degradation [12,13]. Third, magnesium hydroxide demonstrates excellent commercial availability and cost-effectiveness compared to other inorganic flame retardants such as aluminum trihydrate (ATH) while offering superior thermal stability with decomposition temperatures above 350 °C [14]. Finally, previous studies have confirmed the compatibility of Mg(OH)₂ with polyester resin systems, showing minimal adverse effects on mechanical properties when used in appropriate concentrations [15,16].

2.2. Composite Fabrication

Prior to fabrication, mold surfaces were coated with a thin layer of mold release wax to facilitate easy demolding. Sisal fibers were cut to the required dimensions and arranged in layers within the mold cavity to ensure uniform fiber orientation and distribution (Figure 1). Magnesium hydroxide powder was sieved to remove moisture and agglomerates, ensuring homogenous dispersion in the resin.

The complete preparation process is illustrated in Figure 2. The resin, accelerator, and catalyst were prepared in separate beakers. The polyester resin was first mixed with the accelerator and then blended with magnesium hydroxide using manual stirring until a uniform mixture was obtained. The catalyst was added last to initiate curing. The fiber-reinforced resin mixture was carefully poured into the mold and subjected to manual pressing to eliminate air voids. The mold was kept under pressure at ambient temperature for 24 h to allow complete curing. After curing, the composite panels were demolded using hex wrenches (Figure 3).

2.3. Small-Scale Fire Resistance Setup

Fire resistance tests were performed in the Thermal Fluids Laboratory of the Pontificia Universidad Católica de Valparaíso (PUCV), employing a custom-built small-scale gas-fired furnace designed to comply with ISO 834 [25] and NCh935/2 [26] standards. The furnace chamber measured 1 m × 1 m × 2 m (Figure 4).

A forced-draft propane burner (Riello 521 T1, GAS 3/2, Riello, Legnago, Italy) served as the heat source, which is capable of delivering thermal power in two stages: 80–175 kW (low flame) and 130–350 kW (high flame) (Figure 5).

Test specimens were mounted in a robust steel specimen holder with adjustable clamps, which was positioned at a minimum distance of 0.09 m from the flame jet as specified by NCh935/2 [26]. Five type-K thermocouples were fixed on the exposed surface of each specimen to record real-time temperatures at 1 s intervals. An infrared thermographic camera (Fluke TiX580) was deployed to monitor the unexposed side temperatures and capture thermal images during combustion (Figure 6 and Figure 7). Figure 8 shows a representative thermal image captured during testing, displaying the temperature distribution on the unexposed side of the specimen.

2.4. Test Procedure

Before each test, the ambient temperature and specimen initial mass were measured using a 300 g capacity balance with 0.01 g precision. The furnace was controlled to follow the ISO 834 [25] heating curve, maintaining steady temperatures between 500 °C and 550 °C, as verified by thermocouples.

During testing, the flame impinged directly on the specimen until structural failure or self-extinguishment. Burn time was recorded from flame contact to complete integrity loss. After cooling, the final mass was measured. The entire experimental setup and test protocol are depicted in Figure 2.

2.5. Data Collection and Variables

The following variables were recorded for each specimen and are summarized in

Table 1. Summary of experimental variables.

Variable	Unit	Label
Fiber content	%	FIBER_CONTENT (%)
Flame-retardant content	%	FLAME_RETARDANT (%)
Accelerator content	%	ACCELERATOR_CONTENT (%)
Catalyst content	%	CATALYST_CONTENT (%)
Initial mass	g	INITIAL_MASS (g)
Final mass	g	FINAL_MASS (g)
Burn time	s	TIME (s)

.

2.6. Machine Learning Models

Given the limited sample size inherent to laboratory-scale composite fire testing, we adopted a supervised machine learning framework suitable for small datasets. This approach is consistent with recent studies that validate the use of regression and tree-based ensemble models in material science applications with fewer than 50 instances [27,28]. While deep learning techniques typically require large datasets, lightweight machine learning algorithms—such as decision trees, random forests, or gradient boosting—have demonstrated high predictive accuracy and robustness when hyperparameter tuning is performed carefully and overfitting is mitigated [29]. In this context, our experimental dataset, although limited in size (n = 43), was sufficient to train regression models with bootstrapped cross-validation. This strategy ensures model generalizability without compromising physical consistency, which is in line with the methodological frameworks proposed in current composite materials research [30].

Three supervised regression algorithms were used: Random Forest, Support Vector Regression (SVR), and Deep Neural Network (DNN) with scikit-learn and TensorFlow. Data preprocessing involved exploratory analysis (Figure 9), outlier detection via boxplots (Figure 10), and log-transformation of burn time (Figure 11) with validation of the transformed distribution (Figure 12). The Pearson correlation heatmap showed variable relationships (Figure 13).

Model performance was assessed using the Mean Squared Error (MSE) and R². The feature importance for each model is summarized in Figure 12, Figure 13 and Figure 14.

3. Results

3.1. Exploratory Data Analysis

Descriptive statistics and boxplots (Figure 9) were used to examine the distribution of input features and the target variable (burn time). The burn time displayed moderate right-skewness and the presence of outliers. A logarithmic transformation (Figure 11) effectively reduced skewness, yielding a more normal-like distribution suitable for regression modeling. The Pearson correlation heatmap (Figure 15) indicated negligible linear correlation between fiber content and burn time (r $\approx -$0.02), while the flame-retardant content showed a weak positive correlation (r ≈ 0.09). Catalyst and accelerator contents exhibited moderate positive correlations (0.3), implying a non-negligible role in the combustion response.

3.2. Linear vs. Nonlinear Regression

Simple linear regressions were applied to quantify the isolated effect of fiber and flame-retardant content on burn time. As shown in the correlation analysis, fiber content produced a negligible negative slope ($\beta$ = −0.0124), confirming no meaningful linear relationship. The flame-retardant content showed a weak positive slope ($\beta$ = 0.086), indicating a slight trend, but this was insufficient for reliable prediction.

Given the low explanatory power of linear models, advanced nonlinear regression methods were explored to better capture complex interactions among variables.

3.3. Machine Learning Model Performance

Three supervised regression models—Random Forest (RF), Support Vector Regression (SVR), and Deep Neural Network (DNN)—were trained using the full experimental dataset. Due to the limited sample size (n = 43), no separate test set was used; instead, model performance metrics are reported for the training set and should be interpreted cautiously.

Table 2. Machine learning model performance metrics.

Model	MSE	R2
Random Forest	0.1014	−1.217
SVR	0.0662	−0.447
DNN	0.0610	0.334

summarizes the model metrics. The RF and SVR models exhibited negative R² values (−1.217 and −0.447, respectively), indicating that their predictions were inferior to the mean baseline. In contrast, the DNN achieved the lowest MSE (0.061) and a modest positive R² (0.334), suggesting it better captured nonlinear relationships between composition and burn time.

The feature importance plots (Figure 12, Figure 13 and Figure 14) consistently identified flame-retardant content as the dominant factor, which was followed by the catalyst and accelerator content. The fiber content contributed negligibly in all models. These trends align with prior findings by Yan et al. (2023), who also reported the crucial role of flame retardant loading in governing fire performance [22].

3.4. Summary of Findings

There are four key findings from this experimental and machine learning analysis:

-

Fiber content alone does not meaningfully increase burn time; its main function remains structural reinforcement.

-

Flame retardant (Mg(OH)₂) substantially extends burn time through endothermic decomposition and MgO formation, which is consistent with established flame retardant mechanisms [12,13].

-

Catalyst and accelerator ratios affect curing and composite homogeneity, indirectly influencing fire behavior.

-

Machine learning models improve prediction compared to linear regression; however, the DNN’s modest R² underlines the need for larger datasets and cross-validation to enhance model robustness.

Overall, the results demonstrate the viability of integrating targeted flame-retardant dosing with machine learning-driven formulation design to develop sustainable, fire-safe composite materials for the construction sector.

4. Discussion

The integrated experimental and machine learning analysis presented in this study provides new insights into the fire performance of sisal fiber-reinforced polyester composites modified with magnesium hydroxide as a flame retardant. This approach aligns with recent trends in combining data-driven tools with materials engineering to accelerate sustainable design cycles [8,9].

The exploratory data analysis and linear regression confirmed that fiber content has no statistically significant effect on burn time, which is a finding that is consistent with previous studies that natural fibers alone do not meaningfully improve flame resistance without proper flame-retardant loading [17,18]. This supports the established understanding that sisal fiber’s primary role is to provide mechanical reinforcement and sustainability benefits rather than direct contribution to thermal resistance. Similar observations have been reported by Tran et al. (2024) and Jafari et al. (2024) for various natural fiber composites, underscoring that flame-retardant loading is critical for meeting fire safety standards [8,9].

In contrast, flame-retardant content emerged as the primary factor influencing burn time extension. The machine learning models consistently ranked magnesium hydroxide content as the dominant predictor across Random Forest, SVR, and DNN architectures (Figure 12, Figure 13 and Figure 14). This aligns with established flame-retardant theory: magnesium hydroxide decomposes endothermically, absorbs heat, releases water vapor that dilutes flammable gases, and generates a protective MgO layer that insulates the polymer matrix [12,13]. Recent works have demonstrated that combining Mg(OH)₂ with other inorganic fillers such as ammonium polyphosphate (APP) or nanoclays can further enhance char yield and flame suppression, suggesting that future improvements could be achieved by hybrid flame-retardant systems tailored for bio-based FRP panels [14,15].

Catalyst and accelerator contents showed moderate influence, which is consistent with their effect on polymer crosslinking kinetics and the resulting composite microstructure. Optimal curing conditions affect void content and resin homogeneity, indirectly modifying thermal degradation and flame spread. This observation is supported by recent advances in nano-filler incorporation that show well-dispersed nanoparticles can improve thermal stability without compromising mechanical integrity [14,15].

From a modeling perspective, the Deep Neural Network outperformed Random Forest and SVR, achieving an R² of 0.334. This modest value highlights two key points: first, the combustion of polymer composites is governed by complex multiphase phenomena—pyrolysis, char formation, and gas-phase reactions—which cannot be fully captured with a limited number of bulk composition parameters; second, the experimental dataset size (n = 43) constrains the capacity of any data-driven model to generalize. Similar challenges are reported in the literature for the predictive modeling of fire performance with small experimental sets [22]. It should be noted that the sample size used in this study is relatively small for training and validating complex machine learning models such as the Deep Neural Network. This limited dataset constrains the model’s ability to generalize beyond the observed data and increases the risk of overfitting, as reflected by the modest R² value obtained. To mitigate this limitation and ensure more robust and unbiased predictive performance, future studies should employ cross-validation techniques or resampling methods such as bootstrapping. Expanding the dataset with additional experiments and incorporating a wider range of physical and thermal parameters will further enhance the reliability and applicability of the machine learning predictions for real-world fire resistance design.

The improvement of nonlinear models over linear regression demonstrates the potential of machine learning as a complementary tool for rapid formulation screening, reducing the experimental workload and guiding more targeted testing. This is consistent with emerging practices in materials informatics, where hybrid data-driven and physics-based models provide more accurate and scalable fire safety predictions [8,30].

A comparison with similar studies shows that our DNN model’s performance (R² = 0.334) is within the range reported for small-dataset applications in composite fire testing. Yan et al. (2023) achieved R² values of 0.87 for limiting oxygen index prediction with 163 samples [22]. Our results demonstrate the feasibility of machine learning approaches even with limited experimental data, though they underscore the importance of dataset expansion for improved predictive accuracy.

The custom small-scale fire tunnel employed here proved technically robust and cost-effective for controlled flame impingement tests, delivering reproducible measurements under realistic thermal loads. However, the inherent scale limitations imply that results should be cautiously extrapolated to full-scale structural elements. Large-scale standardized fire resistance tests (ISO 834, ASTM E119) remain necessary for regulatory approval. In addition, computational fluid dynamics validation using FireFOAM or FDS can bridge the gap between small-scale experiments and real building-scale fire dynamics [23,24].

While this study provides a practical and affordable framework for integrating experimental testing with machine learning prediction, its applicability is currently limited to laboratory-scale panels. Future work should expand the dataset, explore hybrid flame-retardant systems (e.g., Mg(OH)₂ + APP), and include multi-physics CFD simulations to fully validate fire resistance performance under realistic service conditions.

5. Conclusions

This study combined systematic small-scale fire resistance testing with machine learning modeling to evaluate the fire performance of polyester resin composites reinforced with sisal fiber and treated with magnesium hydroxide. The integrated experimental–machine learning approach demonstrated here provides a practical and affordable framework for designing sustainable, fire-safe composite panels for building applications.

Key conclusions are as follows:

• Sisal fiber content does not significantly affect burn time. Its role remains primarily mechanical reinforcement and sustainability enhancement.
• Magnesium hydroxide effectively extends burn time by leveraging endothermic decomposition and protective MgO layer formation, validating its use as a low-cost flame retardant for bio-based composites.
• Catalyst and accelerator ratios have moderate impact, highlighting the importance of optimal curing for consistent thermal behavior.
• Machine learning models, particularly the Deep Neural Network, outperformed simple linear regression, capturing nonlinear interactions within the small dataset. However, the predictive power (R² = 0.334) is modest, underscoring the need for larger datasets and further parameter inclusion.

Future research should focus on expanding the dataset, incorporating additional physical and thermal parameters (e.g., thermal conductivity, char residue characteristics), and validating the findings through full-scale standardized fire resistance tests to confirm applicability to real-world construction scenarios. The study provides an exploratory framework for assessing the fire performance of natural fiber-based composites using a hybrid experimental and machine learning approach that contributes to emerging literature advocating for compact yet information-rich data pipelines in flame-retardant materials research.