Machine learning based method for predicting natural toxicity of gaseous fire extinguishing agents
By using a machine learning model based on the ICE database, the problem of inaccurate prediction of the toxicity of gaseous fire extinguishing agents in existing technologies has been solved. This enables rapid and accurate toxicity assessment of gaseous fire extinguishing agent molecules, improves the efficiency of fire extinguishing agent R&D and data quality, and provides a unified database and standard.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CIVIL AVIATION UNIV OF CHINA
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-16
AI Technical Summary
Existing machine learning models suffer from insufficient data adaptability and systematicity in predicting the toxicity of gaseous fire extinguishing agents, resulting in inaccurate predictions of the toxicity of specific compounds. Furthermore, the lack of a unified and comprehensive database fails to meet the demand for accurate predictions.
A machine learning model based on the ICE database was established. Through feature engineering and optimization using various algorithms, a natural toxicity prediction model for gaseous fire extinguishing agents was constructed. This included data preprocessing, feature selection and dimensionality reduction, model training and optimization. The XGBR algorithm was used for optimization, and Bayesian optimization and grid search were combined to adjust hyperparameters, enabling rapid and accurate prediction of gaseous fire extinguishing agent molecules.
It improves the accuracy and efficiency of predicting the toxicity of gaseous fire extinguishing agents, shortens the research and development cycle, provides a unified data standard and format, ensures the generalization ability and prediction accuracy of the model, and is applicable to the toxicity assessment of specific compounds.
Smart Images
Figure CN122224329A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of toxicity assessment and safety performance research of gaseous fire extinguishing agents. Specifically, it provides a machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents, which helps to comprehensively and deeply understand the toxicological characteristics of fire extinguishing agents and provides technical support and guarantee for promoting the safe and green development of the civil aviation industry. Background Technology
[0002] Currently, the development of halon alternative fire extinguishing agents mainly focuses on halogenated hydrocarbon compounds (such as pentafluoroethane, hexafluoropropane, heptafluoropropane, trifluoroiodomethane, perfluorohexanone, etc.). These compounds can rapidly cover the fire scene after release, achieving total flooding. However, during this process, nearby personnel may directly come into contact with the extinguishing agent or inhale it, posing certain safety risks. Therefore, the toxicity of halon alternatives must be effectively assessed when screening them. When screening and evaluating halon alternatives, not only their impact on the atmospheric environment should be considered, but also their potential threats to human health and safety. Therefore, toxicological performance evaluation is a crucial step in the research, evaluation, and promotion of next-generation fire extinguishing agents.
[0003] The toxicity assessment of gaseous fire extinguishing agents is complex, involving multiple aspects, including the highest no-toxicity level (NOAEL), the lowest volume fraction with a toxic effect (LOAEL), and the median lethal concentration (LC50). 50 ), median lethal dose (LD50) 50 These toxicity assessments typically involve acute toxicity tests on animals (common experimental animals include mice, rabbits, and dogs) and cardiac sensitization tests on beagles, which have drawbacks such as high cost, long testing cycles, and high workload.
[0004] Given the complexity of experimental assessments, researchers have established quantitative or qualitative relationships between the molecular structure of compounds and their biological activity or toxicity to predict and assess the toxicity of unknown compounds. For example, based on quantitative structure-activity relationship (QSAR) toxicology, Grzyll et al. provided a model-building method that correlated mouse inhalation toxicity data of fire extinguishing agents with exposure time and constructed a predictive model applicable to chemical gas fire extinguishing agents by further refining parameters. Back et al. provided another method for constructing predictive models, establishing a correlation model between the logarithm of the highest harmless dose (NOAEL) and descriptors such as the boiling point and carpa index of the fire extinguishing agent based on toxicological data of cardiosensitive toxicity of more than 30 chemical gas fire extinguishing agents.
[0005] Foreign scholars have used a database containing over 700 epoxidation reactions to construct an epoxidation site prediction model based on deep convolutional neural networks, enabling rapid screening of potentially toxic molecules among drug candidates. Other scholars have successfully predicted the chronic toxicity of over 500 compounds using a k-nearest neighbor classification model. These studies demonstrate that machine learning (ML) can significantly shorten molecule screening time. Applying it to the toxicity prediction of "halon" alternatives is expected to improve prediction accuracy and efficiency, providing scientific support for the development of new fire extinguishing agents.
[0006] Furthermore, some toxicity assessment software, such as TEST, PkCSM, and Pred-Skin, employ various machine learning algorithms and feature extraction methods to achieve rapid assessment of compound toxicity, greatly facilitating material screening. Meanwhile, pharmaceutical toxicity prediction platforms such as VenomPred and ADMETlab3.0 incorporate a wider range of machine learning model combinations, improving consensus machine learning strategies to identify toxicological prediction features, significantly increasing prediction efficiency, reducing research costs, and providing valuable experience for fire extinguishing agent toxicity prediction. Unfortunately, most current machine learning toxicity prediction models are based on extensive compound databases; due to the diverse range of molecules included, toxicity predictions for certain specific mammalian toxicities related to fire extinguishing agents are inaccurate.
[0007] In summary, previous research has achieved a certain level of toxicity assessment for compounds, improving prediction efficiency and reducing research costs. However, current machine learning models still have many shortcomings in toxicity prediction. Most existing machine learning models are based on extensive compound databases, and the complexity of molecular types leads to inaccurate toxicity predictions for specific compounds (such as chemical fire extinguishing agents). Furthermore, machine learning models lack a unified and comprehensive database for specific fields such as chemical fire extinguishing agents, resulting in insufficient adaptability and systematicity of relevant toxicity data. Moreover, they are not adequately adapted to the toxicological characteristics of specific compounds, and the data quality does not meet the requirements for accurate prediction. Summary of the Invention
[0008] Based on the above situation, this invention develops a machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents, which can more accurately predict the natural toxicity of potential fire extinguishing agent molecules; filling the gap in the limited and imprecise toxicity indicators of existing similar fire extinguishing agent molecular structures.
[0009] It should be noted that, based on the EPA18 toxicity prediction software developed by the U.S. Environmental Protection Agency, the experimental and predicted toxicities of common fire extinguishing agent molecules were fitted, revealing a significant discrepancy between experimental and theoretical predictions. For the toxicological analysis of chemical gaseous fire extinguishing agents, there is an urgent need to establish a unified and comprehensive database, improve data quality, and further refine prediction models to achieve accurate predictions of the natural toxicity of fire extinguishing agent molecules. This invention, relying on the ICE database and using the chemical composition and physical properties of existing alternative fire extinguishing agents as characteristics, with experimental toxicity values as the target value, innovatively utilizes machine learning methods to develop a toxicity prediction model for rapidly predicting the natural toxicity of gaseous fire extinguishing agent molecules. This invention contributes to a comprehensive and in-depth understanding of the toxicological characteristics of fire extinguishing agents, providing technical support and assurance for promoting the safe and green development of the civil aviation sector.
[0010] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: A machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents includes the following steps: Step 1: Obtain natural toxicity data of molecules with similar structures to gaseous fire extinguishing agents from the database; Step 2: Preprocess the collected dataset to standardize the toxicity scale; Step 3: Calculate the molecular descriptor of the preprocessed data using software such as PaDEL-Descriptor, E-Dragon, RDKit, and CDK Descriptor, and then perform feature engineering. Step 4: Divide the feature dataset into training and testing sets, build a machine learning model, and use the training set to train the model to find the best machine learning algorithm; Step 5: Input the test set data into the model for prediction, select evaluation indicators to evaluate the model, and select the best-performing combination as the natural toxicity prediction model for gas extinguishing agents according to different division ratios. Step 6: Based on the best-performing machine learning model selected, predict the natural toxicity of gaseous fire extinguishing agents, and then adjust and optimize the model.
[0011] Furthermore, the toxicity data indicators selected in step 1 include the rat inhalation toxicity LC50. 50 The data for this modeling were obtained from at least one of the following: the maximum no adverse reaction level (NOAEL) and the minimum adverse reaction level (LOAEL); the data were obtained from at least one of the ICE database, the ECHA CHEM database, and the ECOTOX database. The ICE database integrates data from multiple authoritative agencies, such as the U.S. Environmental Protection Agency (EPA) and the European Chemicals Agency (ECHA).
[0012] Furthermore, step 2 specifically includes: a systematic search of the ICE toxicity database to identify compounds that meet the initial screening criteria, and data cleaning to remove duplicates or errors, ultimately identifying over 200 compounds with LC... 50 The potential extinguishing agent molecules from the experiment were categorized and organized according to different time periods, such as 2 hours and 4 hours, and a dataset of potential gaseous fire extinguishing agent toxicity was successfully extracted, totaling at least one hundred LC values. 50 (4h) Toxicity data.
[0013] Furthermore, in step 2, based on the Pubchem database, code is written to automatically filter molecules in batches by limiting the carbon chain length, the number of C atoms, and the types of molecular elements, and the data is cleaned to remove duplicates or errors.
[0014] Furthermore, step 3 specifically includes: processing and transforming these raw data, removing redundant data, and retaining data that has a significant impact on the target toxicity value, thereby reducing noise interference and ensuring the final model maintains excellent performance. This invention is based on the molecular structures in the fire extinguishing agent toxicity database and uses one of the molecular descriptor calculation software programs, such as PaDEL-Descriptor, E-Dragon, RDKit, or CDK Descriptor, to calculate over a thousand 1D and 2D molecular descriptors.
[0015] Furthermore, in step 3, this invention employs a systematic feature selection and dimensionality reduction strategy to reduce data dimensionality while retaining key information, thereby improving the training efficiency and generalization ability of the model. First, considering that missing values may lead to incomplete information and potential biases during model training, features with missing values are removed. A low-variance filtering method is used for initial feature screening, setting a variance threshold between 0.05 and 0.3 to remove features with small variations and lack of discriminative power in the samples. Next, this invention introduces methods including feature correlation analysis. By calculating parameters such as the Pearson correlation coefficient between features, highly correlated feature pairs were accurately located. Based on factors such as the chemical meaning of the features, their close association with toxicity, and data distribution, highly correlated features with high redundancy were removed. At least one method, including the correlation coefficient method, chi-square test, and recursive feature elimination method, was used to further optimize the feature subset. At least one algorithm, including random forest (RF), decision tree (DT), and support vector machine (SVM), was used to train the current feature set. Based on the feature weights and importance scores given by the model, features with low scores were removed. Then, the model was retrained and the feature importance scores were updated.
[0016] Furthermore, in step 4, several classic machine learning regression models were selected, including at least one of the following regression models: Random Forest (RF), Decision Tree (DT), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting Algorithm (XGBR), and Lasso Regression (LASSO).
[0017] Furthermore, in step 5, the feature dataset is randomly sampled into training set and test set according to the ratios of 9:1, 8:2, 7:3, 6:4, and 5:5, respectively. The test set is divided into internal validation set and external validation set. The model performs cross-validation on the training set to find the optimal hyperparameters.
[0018] It should be noted that the 9:1 ratio means that 90% of the feature dataset is randomly sampled and used as the training set; 10% of the random sample is used as the test set, and the other ratios are similar.
[0019] Furthermore, the evaluation indicators in step 5 include the correlation coefficient (R²). 2 ), Root Mean Square Error (RMSE), Overfitting Risk, and Mean Square Error (MSE).
[0020] Furthermore, this invention delves into model optimization. Based on determining the machine learning algorithm and the ratio of test set to training set, this invention innovatively employs at least one method, including Bayesian optimization and grid search, for automatic parameter tuning and fine-tuning of the model to improve its predictive performance.
[0021] Furthermore, this invention applies the fully trained machine learning model to the toxicity prediction task of some common gaseous fire extinguishing agents and compares it with known experimental data to prove the feasibility of the model.
[0022] The innovation of this invention compared to existing technologies lies in: (1) Compared with large molecular databases such as PubChem, the special fire extinguishing agent molecular database of the present invention has a significant advantage in data quality. The database effectively eliminates redundant information through strict screening. (2) By using unified data standards and formats, the accuracy of data analysis and mining can be greatly improved. This not only accelerates the rapid identification of potential fire extinguishing agent molecules, but also shortens the research and development cycle and improves research and development efficiency; (3) This invention innovatively attempts to apply machine learning technology to the toxicity prediction of halon substitutes for the first time, so that the toxicity of compounds that meet the conditions can be predicted quickly, which greatly improves the R&D efficiency of gaseous fire extinguishing agents, and at the same time supplements and improves the toxicity index values of similar molecular structures of fire extinguishing agents. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0024] Figure 1 This is a research roadmap for the implementation of this invention.
[0025] Figure 2 This is a diagram of the feature selection process.
[0026] Figure 3 This diagram illustrates the feature dimensionality reduction process based on the RFE method.
[0027] Figure 4 The graph shows the performance of the prediction models obtained by different machine learning algorithms.
[0028] Figure 5 A comparison chart of key indicators for each model.
[0029] Figure 6 The diagram shows the fire extinguishing agent toxicity prediction model established using the XGBR algorithm with different division ratios.
[0030] Figure 7 This is a trend chart showing the impact of feature changes on model evaluation indicators.
[0031] Figure 8 This is a heatmap of the Pearson correlation coefficients for the final feature set.
[0032] Figure 9 This is the final feature weight ranking graph.
[0033] Figure 10 This is a graph showing the final performance of the toxicity prediction model. Detailed Implementation
[0034] This invention discloses a method for predicting the natural toxicity of gaseous fire extinguishing agents based on machine learning.
[0035] To better understand the present invention, the following embodiments are provided for further detailed description of the present invention, but they should not be construed as limiting the present invention. Any non-essential improvements and adjustments made by those skilled in the art based on the above-described invention are also considered to fall within the protection scope of the present invention.
[0036] The technical solution of the present invention will be further described below with reference to specific embodiments. Example
[0037] This embodiment provides a method for predicting the natural toxicity of gaseous fire extinguishing agents based on machine learning, such as... Figure 1 As shown, the method includes the following steps: Step 1: Obtain the natural toxicity data for all molecules from the ICE database.
[0038] Given the widespread use and potential toxicity risks of chemical gaseous fire extinguishing agents, international standards have established toxicity requirements for Halon alternatives: LC50 inhalation toxicity in rats. 50 (4h) ≥ 5000ppm (0.5%), cardiac NOAEL ≥ minimum extinguishing concentration, CSNOAEL > twice the minimum extinguishing concentration. Due to the high cost and scarcity of publicly available data for cardiotoxicity testing, this invention selects rat acute inhalation toxicity LC-15 related to the human respiratory system. 50 (4h) was used as the prediction endpoint. The data for this modeling was obtained from the ICE database. The data selected for this invention is the latest ICE data released by the U.S. National Toxicology Department (NTP) in July 2024.
[0039] Step 2: Preprocess the collected dataset to standardize the toxicity scale.
[0040] Based on publicly available compound molecules in the Pubchem database, code was used to remove molecules containing isotopes, ions, and free radicals, narrowing down the search. Simultaneously, based on the characteristics of ideal gaseous fire extinguishing agents, thresholds or ranges were set for carbon chain length, total number of atoms, and element types. Code was written to limit the total number of atoms in the molecule (to within 20 atoms), the number of carbon atoms (less than or equal to 6), and the types of elements in the molecule (only C, H, F, Cl, Br, I, and other non-metallic elements). This enabled automated batch screening of molecules, and preliminarily obtained information on potential compounds that meet the structural requirements of "Halon" alternative fire extinguishing agents, including molecular formula, molecular structure, and physicochemical properties.
[0041] To achieve accurate prediction of fire extinguishing agent toxicity, this invention systematically searched the ICE toxicity database, identifying compounds that met the initial screening criteria. The data was then cleaned to remove duplicates or errors, ultimately identifying 278 compounds with LC... 50 The experimental values of potential fire extinguishing agent molecules were analyzed. Based on this, the toxicity data of these molecules were categorized and organized according to different time periods, such as 2 hours and 4 hours, and a potential gaseous fire extinguishing agent toxicity dataset was successfully extracted, totaling 169 LC values. 50 (4h) toxicity data provides reliable support for the construction of subsequent toxicity prediction models.
[0042] Step 3: Use PaDEL-Descriptor software to calculate the 1D and 2D molecular descriptors of the preprocessed data and perform feature engineering.
[0043] Feature Extraction: Once sufficient training data is available, the raw data needs to be processed and transformed to remove redundant data and retain data that significantly impacts the target toxicity value. This reduces noise interference and ensures the final model maintains excellent performance. Understanding the target features and the algorithm used greatly aids in feature selection. Generally, the selected features are related to structure or physicochemical properties, such as molecular weight, number of carbon atoms, number of double bonds, number of hydrogen bond acceptors, molecular volume, oil-water partition coefficient, molar refractivity, dipole moment, and polarizability. Based on the molecular structures in the fire extinguishing agent toxicity database, this invention uses PaDEL-Descriptor software to calculate 1444 1D and 2D molecular descriptors, providing rich input features for subsequent machine learning model construction.
[0044] Feature selection and dimensionality reduction: In machine learning model development, the proper control of the number of features is crucial. Too many features can lead to features being redundant, while too few features cannot comprehensively describe the prediction target and cannot guarantee the accuracy of the final prediction model. Therefore, before formally starting model training, the obtained features need to be evaluated and selected.
[0045] To optimize the training feature set, this invention employs a systematic feature selection and dimensionality reduction strategy to reduce data dimensionality while retaining key information, thereby improving the model's training efficiency and generalization ability. First, features with missing values are removed, laying the foundation for subsequent processing. Then, a low-variance filtering method is used for initial feature screening, setting a reasonable variance threshold (≤ 0.1) to remove features with small variations and lack of discriminative power, effectively eliminating some redundant information.
[0046] Subsequently, since high correlations between features can increase model complexity and potentially lead to multicollinearity, thus affecting model stability and predictive performance, this invention introduces feature correlation analysis. By calculating the Pearson correlation coefficient between features, highly correlated feature pairs are precisely identified. Based on factors such as the chemical significance of the features, their strong association with toxicity, and data distribution, highly correlated features with high redundancy are removed, further simplifying the feature dimensions and making the feature set more refined and efficient. The changes in the number of features during feature selection are summarized in... Figure 2 .
[0047] Building upon this foundation, to further optimize the feature subset and make it more aligned with the model's prediction requirements, this invention employs Recursive Feature Elimination (RFE). Based on the feature importance evaluation mechanism of a machine learning model, it iteratively eliminates the least important features. In each iteration, the current feature set is trained using a random forest algorithm. Based on the feature weights and importance scores provided by the model, features with low scores are eliminated. Then, the model is retrained and the feature importance scores are updated. Figure 3 The process of feature dimensionality reduction based on the RFE method is presented. As the number of iterations increases, some features that are not representative are filtered out. Finally, through multiple rounds of iteration, this invention finally selects 15 core features that are most representative and predictive. These features cover key information closely related to toxicity prediction and can provide strong support for efficient training and accurate prediction of the model.
[0048] Step 4: Divide the feature dataset into training and test sets, construct four classic machine learning models, and use the training set to train the models to find the best machine learning algorithm.
[0049] The choice of machine learning algorithm. Even using the same data and feature set, different machine learning algorithms will produce models with varying accuracy. The initial training set used in this invention includes LC-NMR data for 169 potential fire extinguishing agent molecules. 50 (4h) data, combined with the 15 multivariate features selected above, were used for model training. This invention selected four classic machine learning regression models, including Extreme Gradient Boosting (XGBR), Random Forest Regression (RF), Lasso Regression (LASSO), and Gradient Boosting Decision Model (LightGBM), aiming to select the most suitable model for the dataset and prediction task of this invention through extensive model comparison. The training results are as follows: Figure 4 As shown. Figure 5 This is a performance evaluation comparison of prediction models obtained from four different machine learning algorithms. The XGBR model significantly outperforms the other algorithms in prediction accuracy. The LightGBM model's accuracy is close to that of the XGBR model, but still lags behind. The RF and LASSO models both fall short in prediction accuracy. This may be because the XGBR model has advantages in handling complex datasets and capturing nonlinear relationships, enabling it to exhibit higher accuracy and stability in predicting fire extinguishing agent toxicity. Considering all indicators, the XGBR algorithm demonstrates the best accuracy (R²). 2 = 0.765, RMSE = 0.7127 mg / L), which was ultimately used to construct a predictive model for the toxicity of chemical gas fire extinguishing agents.
[0050] Step 5: Input the test set data into the model for prediction, select evaluation indicators to evaluate the model, and select the best-performing combination as the natural toxicity prediction model for gas extinguishing agents based on different division ratios.
[0051] A good machine learning model needs strong predictive power, which means not only matching known data but also accurately predicting unknown data. Therefore, model evaluation is essential. To determine a model's accuracy, datasets are often divided into two parts: a training set and a test set. The training set is used to train the model, and its output consists of known data from the dataset. The test set is used to evaluate the model, providing a final, one-time assessment of the trained model's generalization performance and actual effectiveness on unseen data. Since the ratio of training to test set partitioning directly affects the model's learning performance and the objectivity of the final evaluation, this invention overcomes the limitations of a single partitioning scheme and conducts multi-ratio experimental comparative analysis. Specifically, multiple random partitioning operations were performed for each of several industry-standard and theoretically sound partitioning ratios to reduce random bias from a single partition, using ratios of 9:1, 8:2, 7:3, 6:4, and 5:5. Furthermore, this invention incorporates cross-validation technology, further dividing the dataset into multiple subsets for alternating training and testing. This allows for a more comprehensive evaluation of model performance and avoids overfitting or underfitting issues arising from different data partitioning methods. This ensures the model exhibits strong robustness and generalization ability across different data subsets.
[0052] Figure 6 This paper presents the performance of the XGBR model under different partition ratios. By observing and analyzing the fluctuations of performance indicators (Root Mean Square Error, R²) under different partition ratios, it is possible to effectively identify the data utilization strategy most favorable to the characteristics of the current fire extinguishing agent toxicity dataset and the model architecture, ensuring that the model can both fully learn data patterns and obtain reliable independent evaluations. This invention ultimately determined that the training set and test set should be partitioned at an 8:2 ratio. Image comparison reveals that the model's prediction accuracy reaches its optimal level under this condition.
[0053] Step 6: Based on the best-performing machine learning model selected, predict the natural toxicity of gaseous fire extinguishing agents, and then tune and optimize the model.
[0054] Model tuning and optimization. To further improve the accuracy and stability of the fire extinguishing agent toxicity prediction model, this invention has conducted in-depth model optimization work. Based on determining the machine learning algorithm and the ratio of test to training set partitioning, this invention employs two commonly used automatic hyperparameter tuning methods: Bayesian optimization and grid search, to fine-tune the XGBR model and improve its predictive performance. Bayesian optimization predicts the optimal parameter combination by constructing a probabilistic model, efficiently exploring the parameter space, while grid search ensures comprehensiveness by exhaustively searching predefined parameter combinations. By combining these two methods, this invention can systematically adjust key hyperparameters of the XGBR model, such as maximum tree depth (max_depth), learning rate (eta), gamma, number of trees (n_estimators), and subsample ratio. Furthermore, the model's accuracy is closely related to feature selection; indiscriminate use of all initially selected complex features may result in wasted computational resources due to insufficient simplification. Considering that some complex features may have good independence, while others may contribute similarly to the model as simple features, features with poor independence and low weight can be removed. To obtain the most concise feature set, this invention adopts a last-place elimination strategy. First, the initial 15 features are sorted by relative importance using the XGBR algorithm. Then, the least important feature (i.e., the 15th most important feature) is removed, and the remaining 14 features form a new dataset for the next feature selection step. This process is repeated, removing one least important feature at a time, until the optimal feature subset is found. This invention combines automatic parameter tuning technology with a feature optimization scheme, ultimately effectively improving the model's accuracy. The specific parameter tuning and optimization results are summarized in Table 1.
[0055] To more intuitively understand the degree of influence of features on the model, this invention focuses on R, an important evaluation metric for machine learning models. 2 The process by which RMSE and overfitting risk scores change with features was summarized, such as... Figure 7 As shown.
[0056] Combining the above analysis Figure 7It is evident that if a model is built using only a simple feature set with good independence, the model's accuracy may actually decrease. For example, when the number of features is reduced to 5, the model performs poorly on the test set, with an R² of only 0.681 and an RMSE of 0.8308 mg / L. This may be due to insufficient features, resulting in an incomplete description of the target properties. In summary, this invention removes unimportant and poorly independent features, obtaining a concise training feature set that fully describes the target characteristics. The model trained based on this final training feature set maintains higher accuracy than the initial training set. Finally, after a series of parameter optimizations and secondary evaluation and selection of features, it was determined that a toxicity prediction model with 14 features performs optimally. The contribution of the 14 selected features to the toxicity index of the fire extinguishing agent was analyzed using a Pearson correlation coefficient heatmap and feature importance ranking. Figure 8 , Figure 9 Under these conditions, the performance of the XGBR model was significantly improved (R0). 2 = 0.794, RMSE = 0.6668 mg / L), the model prediction is as follows: Figure 10 As shown, the model's overfitting risk has also been improved accordingly, remaining within an acceptable range. This indicates that the model has a stronger ability to explain data variability and predict the toxicity of new samples.
[0057] Model Validation: To validate the practical application potential of the constructed fire extinguishing agent toxicity prediction model, this invention applied the fully trained XGBR model to the toxicity prediction task of a number of common gaseous fire extinguishing agents. These fire extinguishing agents cover several typical gases that have received much attention in current research and practical applications, such as pentafluoroethane, heptafluoropropane, and carbon tetrachloride (Halon 104). These compounds not only represent mainstream fire extinguishing agent types, but their toxicity data also have high availability and acceptance, providing a solid benchmark for model validation. The comparative analysis of the model prediction results and known experimental data is shown in Table 2.
[0058]
[0059] The results showed a high degree of agreement between the model's predicted values and actual toxicity data, strongly demonstrating the model's excellent quantitative prediction capabilities. Furthermore, the model provided stable and reasonable predictions for fire extinguishing agents with different chemical structures and mechanisms of action, without systematic bias or serious misjudgments of specific types of compounds, further validating the model's practicality and reliability. In fact, after constructing the machine learning prediction model, only simple features needed to be obtained using PaDEL-Descriptor software, and performance prediction could be completed in less than a minute, a significant improvement in speed compared to traditional step-by-step screening methods. This finding lays a solid foundation for the rapid identification of novel, low-toxicity, and highly effective fire extinguishing agents, and also provides clear feedback directions and reliable validation criteria for further iterative optimization of the model.
[0060] This invention is the first to incorporate systematic toxicity indicators into the core architecture, accurately supplementing and improving the data information of traditional databases in the toxicity dimension. It not only provides standardized data support for the comprehensive evaluation of the multiple properties of fire extinguishing agent molecules, but also lays a solid data foundation for subsequent research work such as the optimization of toxicity prediction models.
[0061] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents, characterized in that, Includes the following steps: Step 1: Obtain natural toxicity data of molecules with similar structures to gaseous fire extinguishing agents from the database; Step 2: Preprocess the collected dataset to standardize the toxicity scale; Step 3: Use software to calculate the molecular descriptors of the preprocessed data and perform feature engineering. Step 4: Divide the feature dataset into training and testing sets, build a machine learning model, and use the training set to train the model to find the best machine learning algorithm; Step 5: Input the test set data into the model for prediction, select evaluation indicators to evaluate the model, and select the best-performing combination as the natural toxicity prediction model for gas extinguishing agents according to different division ratios. Step 6: Based on the best-performing machine learning model selected, predict the natural toxicity of gaseous fire extinguishing agents, and then adjust and optimize the model.
2. The machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents according to claim 1, characterized in that, The toxicity data indicators selected in step 1 include rat inhalation toxicity LC50. 50 At least one of the following: maximum non-toxic response dose (NOAEL) and minimum toxic response dose (LOAEL); the data for this modeling were obtained from at least one of the following databases: ICE (Integrated Chemical Environmental Database), ECHA CHEM (European Chemicals Agency Chemical Database), and ECOTOX (Ecotoxicology Knowledge Database).
3. The machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents according to claim 1, characterized in that, Step 2 specifically includes: a systematic search of the toxicity database to identify compounds that meet the initial screening criteria, data cleaning to remove duplicates or errors, and finally determining compounds with LC... 50 Potential extinguishing agent molecules as measured by experiments; By classifying and organizing molecules according to different time periods, a dataset of potential gaseous fire extinguishing agent toxicity was successfully extracted.
4. The machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents according to claim 3, characterized in that, In step 2, based on the Pubchem database, code is written to automatically filter molecules in batches by limiting the carbon chain length, the number of C atoms, and the types of molecular elements, and the data is cleaned to remove duplicates or errors.
5. The machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents according to claim 1, characterized in that, Step 3 specifically includes: processing and transforming the original data, removing redundant data, and retaining data that has a significant impact on the target toxicity value, thereby reducing noise interference and ensuring that the final model maintains excellent performance; and adopting systematic feature selection and dimensionality reduction strategies to reduce data dimensionality while retaining key information, thereby improving the training efficiency and generalization ability of the model.
6. The machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents according to claim 1, characterized in that, The machine learning model constructed in step 4 includes at least one of Random Forest (RF), Decision Tree (DT), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting Algorithm (XGBR), and Lasso Regression (LASSO).
7. The machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents according to claim 1, characterized in that, In step 5, the feature dataset is randomly sampled into training set and test set according to the ratios of 9:1, 8:2, 7:3, 6:4, and 5:5, respectively. The test set is divided into internal validation set and external validation set. The model performs cross-validation on the training set to find the optimal hyperparameters.
8. The machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents according to claim 1, characterized in that, The evaluation indicators in step 5 include the correlation coefficient (R²). 2 ), Root Mean Square Error (RMSE), Overfitting Risk, and Mean Square Error (MSE).
9. The machine learning-based method for predicting the natural toxicity of gaseous fire extinguishing agents according to claim 1, characterized in that, Step 6 involves model tuning and optimization, specifically including two commonly used automatic parameter tuning methods: Bayesian optimization and grid search. These methods refine the model's parameters to improve its predictive performance.