• Home
  • Sitemap
  • Contact us
Int J Fire Sci Eng Search

CLOSE


Int J Fire Sci Eng > Volume 40(1); 2026 > Article
Roh, Shon, Woo, and Kong: A Preliminary Study on Machine Learning-Based Prediction of Vapor Pressure and Flash Point of Hazardous Materials

Abstract

As a preliminary studdy, this study applied machine learning (ML) models to predict vapor pressure and flash point—key physicochemical properties of hazardous materials—and compared predictive performance and feature importance across different models. A dataset of 234 hazardous materials was constructed from the National Hazardous Materials Integrated Information System of the Republic of Korea's National Fire Agency using physicochemical properties such as molecular weight, boiling point, and viscosity as input features. Three regression algorithms—Random Forest, Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)—were used, and their predictive performance was evaluated using the coefficient of determination (R²) and root mean square error (RMSE). Among the tested models, XGBoost demonstrated the highest predictive performance for vapor pressure (R² = 0.82, RMSE = 57.74 mmHg), whereas LightGBM achieved superior performance in flash point prediction (R² = 0.94, RMSE = 13.34 °C). By comparison, Random Forest exhibited lower predictive accuracy, particularly for vapor pressure, suggesting limitations in capturing complex nonlinear relationships. Feature importance analysis indicated that while XGBoost relied heavily on a few key features, LightGBM utilized multiple physicochemical properties more evenly. This difference in feature utilization was closely associated with differences in model performance, demonstrating that the appropriate ML model may vary depending on the target property. These findings highlight the significance of model selection and feature interpretation in hazardous material property prediction.

1. Introduction

Fire behavior is governed by the combined effects of the physicochemical properties of fuel, ventilation conditions, and the spatial distribution of combustible materials. The complex interactions among these factors may lead to rapid fire spread [1,2]. In the case of hazardous materials, fire behavior becomes considerably more complex, as combustion and propagation characteristics are highly sensitive to material properties such as volatility, flammability, and thermal behavior. Consequently, hazardous material fires may escalate rapidly, resulting in severe human and property losses within a short period. Therefore, the ability to predict fire development and establish appropriate response strategies are essential for effective fire management. As urbanization intensifies and industrial facilities continue to expand in scale and complexity, increasingly diverse hazardous substances are being stored and processed. As the types of hazardous substances being stored and handled become increasingly diverse, there is a growing demand for quantitative techniques to evaluate and predict combustion behavior during hazardous material fires. Among the key physicochemical properties, vapor pressure and flash point are particularly important because they provide essential information regarding ignition conditions and fire propagation potential [3,4]. However, experimental determination of these properties often requires substantial time and financial resources, making the acquisition of complete datasets for all materials impractical. To address these limitations, various empirical and theoretical correlations have been developed for predicting vapor pressure and flash point. Widely used approaches include the Clausius-Clapeyron, Antoine, and Wagner equations, which describe temperature-vapor pressure relationships with reasonable accuracy. Despite their practical utility, these methods depend heavily on experimentally derived parameters and exhibit limited generalizability owing to temperature-dependent coefficients [5]. Furthermore, although high predictive performance is often achieved within calibrated data ranges, reliability tends to deteriorate in extrapolated regions [6]. Similar limitations have been reported for flash point prediction models based on boiling point, composition, molecular structure, and vapor pressure, where predictive accuracy varies substantially across substances and operating conditions, and no universally applicable model has yet been established [7]. This implies that conventional property-prediction approaches, while effective under certain conditions, may exhibit limited generalizability across diverse materials and varying conditions. These limitations have motivated growing interest in data-driven methodologies, particularly machine learning (ML) techniques capable of capturing complex nonlinear interactions among features. Such approaches have emerged as promising tools for supplementing experimental methods and estimating physicochemical properties for previously uncharacterized hazardous substances.
Recent studies have proposed various ML-based models for predicting physicochemical properties based on molecular structure and physicochemical properties. Krüger et al. [8] proposed the GC2NN model, which integrates a graph convolutional neural network with the group contribution concept for vapor pressure prediction. The proposed model achieved a high correlation with experimental data (coefficient of determination, R² = 0.94), outperforming conventional empirical and group contribution-based methods. Similarly, Santana et al. [9] proposed the PUFFIN model, in which physical relationships, such as the Antoine equation, were integrated into a neural network framework, demonstrating that the incorporation of physics-based information can improve vapor pressure prediction accuracy. In addition, Li et al. [10] proposed an ML-QSPR framework capable of simultaneously predicting multiple fuel properties, including vapor pressure, flash point, and boiling point. Their study demonstrated that diverse molecular descriptors extracted from molecular structures can effectively support multi-property prediction. Likewise, Ullah et al. [11] employed various boosting algorithms using molecular descriptors to develop physicochemical property prediction models, achieving high predictive performance (R² ≥ 0.74). Through feature importance analysis, the authors identified key factors influencing physicochemical properties. These studies indicate that ML-based physicochemical property prediction is shifting toward the integration of molecular structure information and physicochemical descriptors for estimating diverse material properties. Such approaches have emerged as promising complements to conventional correlation-based methods.
Flash point and vapor pressure are key physicochemical properties related to the fire hazard characteristics of hazardous materials. Flash point is associated with ignition potential, while vapor pressure is closely linked to the formation of flammable vapors. Accurate prediction of these properties provides essential baseline data required for preliminary fire risk assessments, particularly when experimental data are limited. To address this need, this study aims to develop ML models for predicting vapor pressure and flash point using the physicochemical properties of hazardous materials and to quantitatively compare their predictive performance. To address the limitations of conventional correlation-based approaches, ensemble-based ML models capable of capturing nonlinear interactions among physicochemical properties were employed. Particular emphasis was placed on identifying model structures that can provide reliable predictions under limited data conditions. To achieve this objective, a dataset was established using Class 4 hazardous materials obtained from the National Hazardous Materials Integrated Information System of the Republic of Korea’s National Fire Agency [12]. Physicochemical properties, including molecular weight, boiling point, melting point, vapor density, specific gravity, and viscosity, were used as input features. Random Forest, Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) models were used to predict vapor pressure and flash point, and predictive performance was evaluated using R² and root mean square error (RMSE). In addition, feature importance analysis was conducted to identify key features affecting property prediction and to evaluate their physical interpretability.

2. Methods

2.1. Dataset construction

To develop ML models for predicting vapor pressure and flash point, critical physicochemical properties relevant to hazardous material fire response, a dataset was established using information obtained from the National Hazardous Materials Integrated Information System of the National Fire Agency. The dataset focused on flammable liquids (Class 4 hazardous materials), and a total of 2859 raw data records were collected. For each substance, eight physicochemical properties were compiled, including molecular weight, melting point, boiling point, vapor density, specific gravity, viscosity, vapor pressure, and flash point. Only records with complete information for all features were retained, while records containing missing values were excluded. Following the data screening process, a final dataset consisting of 234 samples was obtained. To ensure data consistency and improve model stability and predictive performance, preprocessing was performed to standardize the units of physicochemical properties across the dataset. The physicochemical properties of the hazardous materials were used as input features, while vapor pressure and flash point were assigned as output features, formulating the prediction task as a multi-output regression problem.

2.2. Machine learning models

To predict the physicochemical properties of hazardous materials, this study employed three widely used tree-based ensemble learning models: Random Forest, XGBoost, and LightGBM. These models were selected owing to their capability to capture nonlinear relationships among features and their reliable predictive performance under limited data conditions.
Random Forest is a bagging-based ensemble model that generates multiple decision trees and derives the final prediction by averaging the outputs of individual trees. During model training, multiple training subsets are created from the original dataset through bootstrap sampling with replacement, and each decision tree is independently trained using a different sampled dataset. To enhance model diversity and reduce inter-tree correlation, a subset of features is randomly selected from the total feature set at each node split to determine the optimal partition. This dual-randomization mechanism contributes to reducing overfitting and improving model generalizability while effectively capturing complex nonlinear interactions among features [13]. The overall architecture of the Random Forest model is illustrated in Figure 1.
XGBoost is a gradient boosting-based algorithm that adopts an additive modeling framework, in which predictive performance is progressively improved by sequentially learning the residuals of previous stages. At each stage, a new decision tree is constructed to minimize the prediction errors of the previous model, and the final prediction is generated by summing the outputs of all individual trees. A distinguishing feature of XGBoost is its optimization strategy, which incorporates both the first- and second-order derivatives (gradient and Hessian) of the loss function to optimize the objective function. In addition, parallel processing, tree pruning, and column subsampling techniques are employed to simultaneously improve training efficiency and predictive performance [14]. The overall architecture of the XGBoost model is illustrated in Figure 2.
Similar to XGBoost, LightGBM is a gradient boosting decision tree-based algorithm designed to improve computational efficiency through multiple optimization strategies. One of its key features is histogram-based learning, in which continuous features are converted into discrete bins, substantially reducing split-finding costs and memory consumption during model training. Unlike conventional level-wise tree expansion methods, LightGBM adopts a leaf-wise (best-first) tree growth strategy, where the leaf node associated with the greatest reduction in loss is preferentially expanded. This mechanism enables the construction of deeper tree structures and often improves predictive performance. In addition, gradient-based one-sided sampling is employed to accelerate training by retaining data points with large gradients while selectively sampling less informative data. To further enhance efficiency in high-dimensional feature spaces, exclusive feature bundling groups sparse features that are mutually exclusive to reduce dimensionality and computational complexity. These structural characteristics enable LightGBM to achieve fast training speeds and high predictive accuracy across large-scale and sparse datasets. The overall architecture of the LightGBM model is illustrated in Figure 3.

2.3. ML model training

The compiled dataset consisting of 234 samples was used for ML model training and performance evaluation. Of the total dataset, 50 samples were used as the test set, while the remaining 184 samples were used for model training. The training data were subsequently divided into training and validation subsets at an 8:2 ratio to support model fitting and validation. To ensure reproducibility of the data partitioning process, a fixed random seed was applied.
Prior to model training, standardization was performed using the scikit-learn StandardScaler as a preprocessing step to ensure consistent numerical scaling across physicochemical properties. Each feature was standardized to have a mean of 0 and variance of 1, preventing disproportionate influence from features with different units and numerical ranges among physicochemical properties. Because vapor pressure exhibited a broad value range and nonlinear distribution behavior, a logarithmic transformation was applied to reduce data skewness and improve model stability during pattern learning. After prediction, the estimated vapor pressures were converted back to their original physical units using an inverse exponential transformation. For the categorical feature petroleum class, one-hot encoding was employed to transform discrete categories into numerical vectors. In addition, grid search-based hyperparameter optimization was conducted for each model using predefined search spaces; the detailed hyperparameter ranges are presented in Table 1.

2.4. Data analysis method

The predictive performance of the ML-based hazardous material property prediction models was quantitatively evaluated using R² and RMSE. R² measures the extent to which a model explains the variance in observed data and is widely used to assess the agreement between predicted and actual values. In general, values closer to 1 indicate stronger explanatory capability, whereas negative values may be obtained when the residual sum of squares exceeds the total sum of squares. R² was calculated using the ratio between the residual sum of squares and total sum of squares, as expressed in Eq. (1).
(1)
R2=1-∑i=1N(yi-y^i)2∑i=1N(yi-y−i)2
Here, yi represents the actual value, ŷi represents the predicted value, and N is the total number of samples.
Additionally, RMSE was used to quantify the model's prediction errors. RMSE is defined as the square root of the average squared difference between predicted and actual values, representing the magnitude of absolute prediction error. Lower RMSE values indicate better predictive performance and are calculated as shown in Eq. (2).
(2)
RMSE=1N∑i=1N(yi-y^i)2

3. Results and Discussion

3.1. Predictive performance of ml models for hazardous material properties

The predictive capabilities of Random Forest, XGBoost, and LightGBM models for estimating the vapor pressure and flash point of hazardous materials were comparatively investigated. Model performance was quantitatively evaluated using R² and RMSE with the test dataset. The prediction results for vapor pressure and flash point are compared in Figures 4(a) and 4(b), respectively.
For vapor pressure prediction, substantial differences in predictive performance were observed among the evaluated models. The Random Forest model demonstrated the weakest performance, with an R² value of -0.0006 and the highest RMSE (84.72 mmHg), indicating limited capability in explaining the variability of the observed data. Negative R² values suggest that model predictions may perform worse than a simple mean-based estimation. As shown in Figure 4(a), the predicted values were concentrated within a narrow range, indicating that the model was unable to adequately capture the broad distribution of the observed vapor pressure values. This result is likely attributable to the large numerical range and strong nonlinear characteristics of vapor pressure, combined with the relatively limited representation of high-vapor-pressure samples in the dataset. Because the Random Forest model relies on averaging predictions from multiple decision trees, its reduced sensitivity to extreme values may cause predictions to converge toward an average range. By comparison, XGBoost exhibited the strongest predictive performance, with an R² of 0.82 and reduced RMSE of 57.74 mmHg. As shown in Figure 4(a), the predicted values were relatively evenly distributed around the identity line, indicating that the model effectively captured the nonlinear behavior of vapor pressure. Although LightGBM also showed moderate predictive performance (R² = 0.67), its predictions displayed greater dispersion than those of XGBoost, accompanied by a comparatively higher RMSE (74.09 mmHg). This behavior may be related to the leaf-wise growth mechanism of LightGBM, which can promote excessive partitioning in certain regions of the feature space and increase prediction variability. These findings indicate that gradient boosting-based approaches may be more suitable than bagging-based methods for predicting target variables with broad ranges and pronounced nonlinear behavior, such as vapor pressure, where the sequential error-correction framework of XGBoost appeared to improve the representation of complex patterns within the data.
In contrast to the vapor pressure results, the flash point prediction results exhibited a different pattern of model performance. The Random Forest model demonstrated relatively strong predictive performance, with an R² value of 0.85 and RMSE of 17.55 °C. As illustrated in Figure 4(b), most predicted values were distributed near the identity line; however, partial underestimation of actual values occurred in certain regions. This tendency is likely related to the averaging mechanism of Random Forest, which reduces sensitivity to extreme values during prediction. Compared with Random Forest, XGBoost demonstrated improved predictive performance, with an R² value of 0.91 and lower RMSE of 14.82 °C. Although predictions were generally concentrated near the identity line, slight underestimation was still observed in the higher-flash-point region, as illustrated in Figure 4(b). Among the evaluated models, LightGBM exhibited the best predictive performance, with the highest R² value of 0.94 and lowest RMSE of 13.34 °C. The flash point dataset ranged from -46 °C to 187 °C, with a mean of 32.71 °C and standard deviation of 57.38 °C. The RMSE of the LightGBM model corresponds to approximately 5.7% of the total data range and 23.2% of the dataset standard deviation. As shown in Figure 4(b), predictions generated by LightGBM were distributed most closely around the identity line across the entire range and exhibited comparatively lower dispersion, indicating stable and reliable predictive performance. The superior performance of LightGBM may be attributable to its leaf-wise growth mechanism, which preferentially expands leaves associated with large reductions in loss, enabling the model to focus on regions that contribute more to loss reduction and capture subtle variations within the dataset. The formation of relatively deeper tree structures may have further contributed to improved learning of complex nonlinear relationships. Because flash point is a critical property used in hazardous material classification and safety management, prediction errors should be interpreted in the context of practical applications and proximity to relevant threshold values. Therefore, the findings of this study may provide useful preliminary information for evaluating the feasibility of ML-based hazardous material property prediction, particularly for substances with limited experimental data.

3.2. Feature importance analysis

To investigate how each ML model utilized input features for hazardous-material property prediction, feature importance analysis was conducted for each model. The resulting feature importance distributions were comparatively evaluated for vapor pressure and flash point prediction.
For vapor pressure prediction, distinct differences in feature utilization were observed among the evaluated models (Figure 5). In the Random Forest model, boiling point and petroleum class emerged as the most influential features, with importance values of approximately 0.336 and 0.306, respectively. This finding suggests that vapor pressure prediction was strongly associated with thermal properties such as boiling point, while petroleum class may have served as an indirect descriptor of overall material properties. Other properties, including molecular weight and melting point, also contributed to model prediction, indicating a relatively balanced use of multiple physicochemical properties. By comparison, the XGBoost model exhibited a markedly concentrated feature importance distribution, with petroleum class accounting for an importance value of 0.749. This pattern may reflect the tendency of boosting-based models to establish strong branching structures around specific regions during the data partitioning process, increasing the reliance on specific features. By contrast, the LightGBM model displayed a more evenly distributed feature importance pattern across multiple physicochemical properties. Notably, viscosity and boiling point exhibited similar importance values (~0.180), followed by melting point (0.157), specific gravity (0.130), and molecular weight (0.129). The similar contributions of these features suggest that LightGBM incorporated diverse physicochemical properties in a balanced manner, leading to more stable feature utilization. Meanwhile, despite its strong dependence on a limited number of features, XGBoost maintained a comparatively strong predictive performance; this may be associated with its iterative residual-correction mechanism, which effectively captures complex nonlinear relationships.
For flash point prediction, feature importance distributions also differed substantially among the evaluated models, exhibiting trends similar to those observed for vapor pressure prediction (Figure 6). In the Random Forest model, petroleum class and boiling point emerged as the dominant predictive features, with importance values of approximately 0.386 and 0.286, respectively. This observation suggests that flash point prediction may be associated with both material-classification characteristics and thermophysical behavior. In the XGBoost model, the importance value of petroleum class was 0.986, indicating strong reliance on this categorical feature. This pattern may reflect the tendency of the model to establish strong branching structures around specific regions during the data partitioning process, resulting in a feature importance distribution concentrated on a limited subset of features. By contrast, the LightGBM model exhibited a more balanced distribution of feature importance across multiple inputs, consistent with the vapor pressure prediction results. Viscosity and boiling point showed the highest importance values (0.188 and 0.187, respectively), followed by specific gravity (0.149), melting point (0.148), and molecular weight (0.133). Overall, the model exhibited lower dependence on any single input and incorporated multiple physicochemical properties in a more distributed manner. Notably, the feature importance patterns for flash point prediction varied more substantially across models than those observed for vapor pressure. While some models relied heavily on a single dominant feature, others incorporated broader physicochemical properties more evenly. These differences likely reflect variations in feature utilization strategies inherent to model architecture and may have contributed to differences in predictive performance and stability.

4. Conclusion

This study comparatively evaluated the performance of Random Forest, XGBoost, and LightGBM models for predicting vapor pressure and flash point, two key physicochemical properties relevant to hazardous material fire response. Using a dataset obtained from the National Hazardous Materials Integrated Information System of the National Fire Agency, model performance was evaluated using R² and RMSE. The main findings are summarized as follows:
- For vapor pressure prediction, XGBoost demonstrated the best performance among the evaluated models, with an R² value of 0.82 and an RMSE of 57.74 mmHg, while the Random Forest model showed relatively poor predictive performance, for the vapor pressure dataset, which exhibited a broad numerical range and nonlinear characteristics. These findings suggest that gradient boosting-based approaches may be more suitable for predicting physicochemical properties characterized by broad numerical ranges and strong nonlinear behavior.
- For flash point prediction, all evaluated models showed generally stable performance, with LightGBM achieving the highest predictive accuracy (R² = 0.94, RMSE = 13.34 °C). This result indicates that differences in model architecture and learning strategies may substantially influence predictive performance for flash point estimation.
- Feature importance analysis further revealed that variable importance differed according to both model type and target property. While XGBoost relied heavily on petroleum class, LightGBM incorporated multiple physicochemical properties in a comparatively balanced manner. These findings imply that distinct feature utilization strategies employed by different models may affect not only prediction accuracy but also stability and interpretability.
Future work should focus on expanding dataset size and incorporating additional physicochemical descriptors to improve predictive reliability and extend model applicability to a broader range of hazardous materials.

Notes

Author Contributions

Conceptualization and methodology, J.R. and M.K.; formal analysis and investigation, J.R.; writing - original draft preparation, J.R. and J.W.; writing - review and editing, M.K. and S.S.; supervision, M.K.; project administration, M.K.; All authors have read and agreed to the published version of the manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

Figure 1.
Schematic of random forest model.
KIFSE-b784d9bcf1.jpg
Figure 2.
Schematic of XGBoost model.
KIFSE-b784d9bcf2.jpg
Figure 3.
Schematic of LightGBM model.
KIFSE-b784d9bcf3.jpg
Figure 4.
Comparison of predicted and actual values for (a) vapor pressure and (b) flash point using ML models.
KIFSE-b784d9bcf4.jpg
Figure 5.
Comparison of feature importance for vapor pressure prediction across ML models (Random forest, X GBoost, LightGBM).
KIFSE-b784d9bcf5.jpg
Figure 6.
Comparison of feature importance for flash point prediction across ML models (Random forest, XGBoost, LightGBM).
KIFSE-b784d9bcf6.jpg
Table 1.
Hyperparameter Search Space for ML Models
Model Hyperparameter Description Search Range
Random forest n_estimators Number of trees 50, 100, 200, 300
max_depth Maximum tree depth None, 5, 10, 15, 20, 30
min_samples_split Minimum samples to split node 2, 4, 6, 10
min_samples_leaf Minimum samples at leaf node 1, 2, 4, 8
max_features Number of features for split sqrt, log2, none
bootstrap Bootstrap sampling True, False
Criterion Split quality metric squared_error, absolute_error
XGBoost n_estimators Number of trees 50, 100, 200, 300
max_depth Maximum tree depth 3, 5, 7, 10
learning_rate Learning rate 0.01, 0.03, 0.05, 0.1, 0.2
subsample Row sampling ratio 0.6, 0.8, 1.0
colsample_bytree Feature sampling ratio 0.6, 0.8, 1.0
gamma Minimum loss reduction 0, 0.1, 0.3, 0.5
min_child_weight Minimum child weight 1, 3, 5, 7
reg_alpha L1 regularization 0, 0.1, 0.5, 1.0
reg_lambda L2 regularization 1, 3, 5, 10
LightGBM n_estimators Number of trees 30, 50, 100, 200
learning_rate Learning rate 0.03, 0.05, 0.1, 0.2
num_leaves Number of leaves 15, 31, 63
max_depth Maximum tree depth -1, 3, 5, 7, 10
subsample Row sampling ratio 0.7, 0.85, 1.0
colsample_bytree Feature sampling ratio 0.7, 0.85, 1.0
reg_alpha L1 regularization 0.0, 0.1, 0.5
reg_lambda L2 regularization 0.0, 1.0, 3.0
min_child_samples Minimum samples per leaf 5, 10, 20, 30

References

1. M. J. Hurley, D. Gottuk, J. R. Hall, K. Harada, E. Kuligowski, M. Puchovsky, J. Torero, J. M. Watts et al. "SFPE Handbook of Fire Protection Engineering", 5th ed., Springer, New York (2016). https://doi.org/10.1007/978-1-4939-2565-0.

2. J. G. Quintiere, "Principles of Fire Behavior", 2nd ed., CRC Press, Boca Raton, FL (2016).

3. X. Liu and Z. Liu, “Research Progress on Flash Point Prediction”, Journal of Chemical & Engineering Data, Vol. 55, No. 9, pp. 2943-2950 (2010), https://doi.org/10.1021/je1003143.
crossref
4. A. Delle Site, “The Vapor Pressure of Environmentally Significant Organic Chemicals: A Review of Methods and Data at Ambient Temperature”, Journal of Physical and Chemical Reference Data, Vol. 26, No. 1, pp. 157-193 (1997), https://doi.org/10.1063/1.556006.
crossref
5. S. Ohe, “A Prediction Method of Vapor Pressures for Pure Substances”, Fluid Phase Equilibria, Vol. 561, pp. 113474(2022), https://doi.org/10.1016/j.fluid.2022.113474.
crossref
6. K. Qing, E. Wang, Q. Song, Z. Yang and Y. Duan, “A New Vapor Pressure Equation with Accurate Extrapolation Capabilities”, Fluid Phase Equilibria, Vol. 601, pp. 114576(2026), https://doi.org/10.1016/j.fluid.2025.114576.
crossref
7. L. Y. Phoon, A. A. Mustaffa, H. Hashim and R. Mat, “A Review of Flash Point Prediction Models for Flammable Liquid Mixtures”, Industrial & Engineering Chemistry Research, Vol. 53, No. 32, pp. 12553-12565 (2014), https://doi.org/10.1021/ie501233g.
crossref
8. M. Krüger, T. Galeazzo, I. Eremets, B. Schmidt, U. Pöschl, M. Shiraiwa and T. Berkemeier, “Improved Vapor Pressure Predictions using Group Contribution-Assisted Graph Convolutional Neural Networks (GC2NN)”, Geoscientific Model Development, Vol. 18, No. 20, pp. 7357-7371 (2025), https://doi.org/10.5194/gmd-18-7357-2025.
crossref
9. V. V. Santana, C. M. Rebello, L. P. Queiroz, A. M. Ribeiro, N. Shardt and I. B. R. Nogueira, “PUFFIN: A Path-Unifying Feed-Forward Interfaced Network for Vapor Pressure Prediction”, Chemical Engineering Science, Vol. 286, pp. 119623(2024), https://doi.org/10.1016/j.ces.2023.119623.
crossref
10. R. Li, J. M. Herreros, A. Tsolakis and W. Yang, “Integrated Machine Learning-Quantitative Structure Property Relationship (ML-QSPR) and Chemical Kinetics for High Throughput Fuel Screening toward Internal Combustion Engine”, Fuel, Vol. 307, pp. 121908(2022), https://doi.org/10.1016/j.fuel.2021.121908.
crossref
11. A. Ullah, M. Shaheryar and H.-J. Lim, “Machine Learning Approach for the Estimation of Henry’s Law Constant Based on Molecular Descriptors”, Atmosphere, Vol. 15, No. 6, pp. 706(2024), https://doi.org/10.3390/atmos15060706.
crossref
12. National Fire Agency, "Hazardous Materials Information System", https://hazmat.nfa.go.kr/ (accessed n.d.).

13. L. Breiman, “Random Forests”, Machine Learning, Vol. 45, pp. 5-32 (2001), https://doi.org/10.1023/A:1010933404324.
crossref
14. T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System”, Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 785-794 (2016), https://doi.org/10.1145/2939672.2939785.
crossref
15. G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye and T.-Y. Liu, “LightGBM: A Highly Efficient Gradient Boosting Decision Tree”, Proceedings of the 31st Conference on Neural Information Processing Systems (NeurIPS 2017), pp. 3146-3154 (2017).



ABOUT
BROWSE ARTICLES
EDITORIAL POLICY
AUTHOR INFORMATION
Editorial Office
Room 906, The Korea Science Technology Center The first building, 22, Teheran-ro 7 Gil, Gangnam-gu, Seoul, Republic of Korea
Tel: +82-2-555-2450/+82-2-555-2452    Fax: +82-2-3453-5855    E-mail: kifse@hanmail.net                

Copyright © 2026 by Korean Institute of Fire Science and Engineering.

Developed in M2PI

Close layer
prev next