Method for predicting irradiation chemical damage degree of molecular crystal

Abstract

The invention discloses a method for predicting the irradiation chemical damage degree of a molecular crystal, and the method comprises the steps: carrying out the data preparation, building a training data set, carrying out the preprocessing of the data, constructing a prediction model of the irradiation chemical damage of the molecular crystal through XGBoost after the preprocessing, and carrying out the prediction of the irradiation chemical damage degree of the molecular crystal through the XGBoost. And calculating the contribution degree of each feature to model prediction according to XGBoost, and calculating a feature interaction effect on the contribution of each feature by using an SHAP method to obtain irradiation sensitive molecular group data. The method has the advantages that limitation of single energy scale of a traditional empirical model is broken through, key damage sensitive factors such as molecular groups and bond dissociation energy can be revealed through cross-scale feature fusion and model interpretability analysis, irradiation damage prediction is promoted to be transformed from an empirical formula to an intelligent calculation normal form by the technology, and the method is suitable for large-scale popularization and application. And an interdisciplinary solution is provided for life evaluation of space electronic devices, nuclear reactor material design, energetic molecular structure optimization and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of material irradiation damage simulation, and particularly relates to a method for predicting the degree of irradiation chemical damage of molecular crystals. Background Art

[0002] The irradiation environment is one of the severe application environments faced by energetic materials. Under irradiation conditions, energetic materials will be subjected to the irradiation of high-energy particles, causing molecular bond breaking or knocking out lattice atoms, generating initial damage defects of cascade collisions. Under the action of external conditions such as service temperature and stress, these defects will interact with each other, resulting in damage effects such as color change, accelerated decomposition, and reduced mechanical properties of energetic materials, ultimately causing macroscopic property changes of energetic materials and affecting the service performance of energetic materials. Since energetic materials themselves are high-energy density materials and are relatively sensitive to the external environment, their irradiation damage experiments are difficult, time-consuming, cumbersome, and costly. A small number of targeted experiments combined with computer simulation techniques (especially molecular dynamics simulation) are the main methods for studying the irradiation damage and aging mechanisms of energetic materials.

[0003] The accurate prediction of the degree of irradiation damage is a key scientific issue in the fields of nuclear energy materials, space electronic devices, and energetic molecular crystals. The traditional NRT model is based on a linear empirical formula of the energy of primary knock-on atoms (PKA) and the displacement threshold energy, and can only predict the number of point defects, making it difficult to characterize the multi-scale nonlinear mechanisms of molecular crystal irradiation damage, such as chemical bond breaking, free radical generation, and crystal periodic destruction, resulting in prediction failure in complex molecular systems.

[0004] Based on this background, the present technology proposes a machine learning framework integrating multi-scale data-driven methods. By integrating first-principles calculations, molecular dynamics simulations, and experimental characterization data, an ensemble learning decision tree model is constructed. The physical and chemical properties of molecular crystals, irradiation energy fluctuations, and local chemical environments are encoded as numerical data. Through the coupled analysis of multi-scale damage characteristics of atoms-molecules-crystals, the prediction of irradiation decomposition and the number of displaced atoms is achieved. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a method for predicting the degree of irradiation chemical damage of molecular crystals.

[0006] The technical solution is as follows: A method for predicting the degree of irradiation chemical damage of molecular crystals, which is characterized by including the following steps:

[0007] S1: Data preparation, read the molecular crystal model file of the material to be analyzed, perform molecular dynamics simulation and irradiation simulation using the lammps software package, obtain data and construct a training dataset, where the number of displaced atoms and the number of molecular decompositions are used as target variables to be predicted;

[0008] S2: Data preprocessing, dealing with abnormal data, augmenting the dataset through data enhancement, expanding the feature set through feature engineering, and screening feature quantities;

[0009] S3: Building a prediction model and model evaluation, building a prediction model for the radiation chemical damage of molecular crystals based on XGBoost, splitting the preprocessed dataset in step S2 into a training set and a validation set according to a ratio of 8:2, adopting a 5-fold cross-validation method, and evaluating this prediction model;

[0010] S4: Sensitivity factor analysis, using XGBoost in step S3 to calculate the contribution degree of each feature to the model prediction, and using the SHAP method to calculate the feature interaction effect of the contribution of each feature, obtaining radiation-sensitive molecular group data.

[0011] Preferably: In step S1, the lammps software package is used for molecular dynamics simulation, and the characteristic data of this molecular crystal model is specified;

[0012] The characteristic data includes: model density, unit cell parameters, model volume, total number of atoms, number of various types of atoms, molar mass ratio of various types of atoms, initial potential energy of the system, initial temperature of the system, cohesive energy, primary displacement atom energy, type of primary displacement atom (PKA);

[0013] After obtaining the characteristic data of the molecular crystal model, further calculate the molecular chemical formula in the molecular crystal model, simplify the molecular linear input specification characters, number of molecules, molecular mass, number of various chemical bonds, bond energy of the trigger bond in the molecule;

[0014] After irradiating the molecular crystal model using the lammps software package, calculate the number of displacement atoms and the number of molecular decompositions during the irradiation process through the simulated trajectory data file.

[0015] Preferably: Integrate the above data to form a training dataset with fields including model density, unit cell parameters, model volume, total number of atoms, number of various types of atoms, molar mass ratio of various types of atoms, initial potential energy of the system, initial temperature of the system, cohesive energy, primary displacement atom energy, type of primary displacement atom, molecular chemical formula, simplified molecular linear input specification characters, number of molecules, molecular mass, number of various chemical bonds, bond energy of the trigger bond in the molecule, number of displacement atoms, and number of molecular decompositions.

[0016] Preferably: The abnormal data in step S2 includes outliers and missing values, and the processing is to fill them with the mode in the corresponding numerical values in the field.

[0017] Preferably, in the step S2, the data augmentation to expand the data set is carried out according to the grid interpolation and Monte Carlo methods. A physical model for predicting NRT irradiation defects with a temperature correction term is artificially constructed and carried out under the constraints of the physical model.

[0018] Preferably, the formula of the physical model for predicting NRT irradiation defects with a temperature correction term is as follows:

[0019]

[0020] Where N d is the number of displaced atoms, E PKA is the energy of the primary displaced atom, E d is the displacement threshold energy, k B is the Boltzmann constant, T is the temperature. The influence of temperature on the displacement threshold energy of the material is considered in the construction of this formula, and the atomic thermal activation effect when the PKA energy is lower than the displacement threshold energy is described using the Boltzmann factor.

[0021] Preferably, the feature engineering is to numerically process three types of non-numerical features in the data set, namely, the types of primary displaced atoms, molecular chemical formulas, and simplified molecular linear input specification characters. Among them, the types of primary displaced atoms adopt one-hot encoding, the molecular chemical formulas adopt the number of atoms of different types, and the simplified molecular linear input specification characters adopt Morgan molecular fingerprints to calculate their structural information;

[0022] And train a random forest model to judge the importance of each feature for the target variable and conduct a correlation matrix analysis, delete features with a correlation coefficient > 0.95, and retain features with an importance greater than the average value.

[0023] Preferably, in the step S3, to prevent the model from overfitting, if the performance of the validation set does not improve in 50 consecutive iterations, stop training and return the model with the best performance on the validation set.

[0024] Preferably, the best performance model uses the Bayesian optimization method to systematically optimize the key hyperparameters of XGBoost; by constructing a probability model of the objective function and guiding the hyperparameter search based on the observed performance.

[0025] Preferably, in the step S4, according to the molecular fingerprint data in the important features, use the Rdkit library to reverse-infer the structure in the molecule that is more sensitive to damage under irradiation load, and obtain the data of irradiation-sensitive molecular groups, providing a data-driven basis for molecular irradiation chemical reactions.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] 1) Break through the limitation of the single energy scale of traditional empirical models, and be able to accurately capture the non-linear correlation between irradiation damage and molecular configuration, crystal physical and chemical properties, and irradiation conditions;

[0028] 2) Through cross-scale feature fusion and model interpretability analysis, key damage-sensitive factors such as molecular groups and bond dissociation energies can be revealed;

[0029] 3) The trained model can achieve multi-target prediction of displaced atoms and molecules, and can quickly and accurately predict the irradiation damage degree of different molecular crystals compared with the traditional NRT model.

[0030] This technology not only promotes the transformation of irradiation damage prediction from empirical formulas to intelligent computing paradigms, but also provides interdisciplinary solutions for the needs of space electronic device life assessment, nuclear reactor material design, energetic molecule structure optimization, etc. Brief Description of the Drawings

[0031] Figure 1 It is a schematic flow chart of the present invention. Detailed Embodiments

[0032] The present invention will be further described below in conjunction with embodiments.

[0033] As Figure 1 shown, a method for predicting the degree of irradiation chemical damage of molecular crystals, characterized in that it includes the following steps:

[0034] S1: Data preparation, read the molecular crystal model file of the material to be analyzed, use the lammps software package for molecular dynamics simulation and irradiation simulation, obtain data and construct a training data set, where the number of displaced atoms and the number of molecular decompositions are used as target variables to be predicted;

[0035] S2: Data preprocessing, process abnormal data, augment the data set by data enhancement, expand the feature set by feature engineering, and screen feature quantities;

[0036] S3: Construct a prediction model and model evaluation, construct a prediction model for the irradiation chemical damage of molecular crystals based on XGBoost, divide the data set preprocessed in step S2 into a training set and a validation set according to a ratio of 8:2, adopt a 5-fold cross-validation method, and evaluate the prediction model;

[0037] S4: Sensitivity factor analysis, use XGBoost in step S3 to calculate the contribution degree of each feature to the model prediction, and use the SHAP method to calculate the feature interaction effect of the contribution of each feature to obtain irradiation-sensitive molecular group data.

[0038] Embodiment

[0039] 1. Data preparation

[0040] By reading the molecular crystal model file of the material to be analyzed, the lammps software package is used to perform molecular dynamics simulation, and the following characteristic data of the molecular crystal model are specified for calculation: model density, unit cell parameters, model volume, total number of atoms, number of each type of atoms, molar mass ratio of each type of atoms, initial potential energy of the system, initial temperature of the system, cohesive energy, primary detached atom energy, and primary detached atom (PKA) type;

[0041] Calculate the molecular chemical formula in the molecular crystal model, simplify the molecular linear input standard characters, molecular number, molecular mass, the number of various chemical bonds, and the bond energy of the trigger bond in the molecule;

[0042] After using lammps to simulate the irradiation of the molecular crystal model, the number of atoms displaced and the number of molecular decompositions during the irradiation process were calculated using the simulation trajectory data file. These two features were used as target variables to be predicted.

[0043] All the above data are organized to form a training data set with the fields of model density, unit cell parameters, model volume, total number of atoms, number of each type of atoms, molar mass ratio of each type of atoms, initial potential energy of the system, initial temperature of the system, cohesive energy, energy of primary detached atoms, type of primary detached atoms, molecular chemical formula, simplified molecular linear input standard characters, number of molecules, molecular mass, number of each type of chemical bonds, bond energy of trigger bonds in molecules, number of detached atoms and number of molecular decompositions.

[0044] 2. Data preprocessing

[0045] Abnormal data processing: For abnormal values and missing values, the mode of the corresponding values of the field is used to fill them.

[0046] Data enhancement: For the molecular data corresponding to each chemical formula, a grid is constructed in the PKA energy-temperature space. The number of dislocated atoms and the number of decomposed molecules are interpolated within the grid using K-nearest neighbor interpolation. Each interpolation is judged using the NRT irradiation defect prediction physical model with a temperature correction term that predicts the number of irradiated point defects. If the number of dislocated atoms generated by the interpolation deviates from the NRT model by more than 20%, the interpolation is repeated. After completing the interpolation of the number of dislocated atoms, a grid is constructed in the space between the number of dislocated atoms and the number of decomposed molecules, and linear interpolation is performed to generate the number of decomposed molecules data.

[0047] The formula for the NRT irradiation defect prediction physical model with a temperature correction term is as follows. This formula is specifically constructed by introducing a temperature correction term into the classic NRT model for the predetermined data processing goal:

[0048]

[0049] Among them, N d is the number of displaced atoms, E PKA is the energy of the primary displaced atom, E d is the displacement threshold energy, k B is the Boltzmann constant, and T is the temperature. The influence of temperature on the displacement threshold energy of the material is considered in the construction of this formula, and the atomic thermal activation effect when the PKA energy is lower than the displacement threshold energy is described by the Boltzmann factor. After data augmentation, the data volume of the dataset can be expanded from 100 to 1000 within the original dataset space.

[0050] Feature engineering: Numerically process the three non-numerical features of the primary displaced atom type, molecular chemical formula, and simplified molecular linear input specification characters in the dataset. More specifically, the one-hot encoding is used for the primary displaced atom type, the number of different types of atoms is used for the molecular chemical formula, and the Morgan molecular fingerprint is used to calculate its structural information (collection radius is 2, number of bits is 1024) for the simplified molecular linear input specification characters.

[0051] Data screening: Train a random forest model to judge the importance of each feature for the target variable and perform a correlation matrix analysis, delete features with a correlation coefficient > 0.95, and retain features with an importance greater than the average value.

[0052] 3. Model training and evaluation

[0053] Model training: Use XGBoost (eXtreme Gradient Boosting) as the main machine learning model to predict the damage behavior of molecules under irradiation conditions. The preprocessed dataset is split into a training set and a validation set according to a ratio of 8:2, and a 5-fold cross-validation method is adopted. To prevent overfitting, if the performance of the validation set does not improve in 50 consecutive iterations, stop training and return the model with the best performance on the validation set.

[0054] To obtain the best model performance, the Bayesian optimization method is used to systematically optimize the key hyperparameters of XGBoost, where the key hyperparameters include the learning rate, maximum tree depth, minimum child node weight, sample sampling ratio, feature sampling ratio, splitting threshold, L1 regularization parameter, and L2 regularization parameter. By constructing a probability model (Gaussian process) of the objective function, the hyperparameter search is guided based on the observed performance.

[0055] Model evaluation: Use the root mean square error, coefficient of determination, mean absolute error, and mean absolute percentage error as evaluation indicators to evaluate the model.

[0056] 4. Interpretability analysis

[0057] Calculate the contribution of each feature to the model prediction through XGBoost. Further, use the SHAP (SHapley Additive exPlanations) method to calculate the SHAP values and average SHAP values of each feature, and screen out the fingerprint positions with a frequency > 5% and a correlation coefficient with the target variable > 0.3 and their radii. The atomic environment within the range . Further, based on the molecular fingerprint data in the important features, use the Rdkit library to extract the atomic environment information and deduce that the structure important for damage under irradiation load in the molecule is a cyclic structure centered on a C atom with a local environment of N-C-N, a hybridization state of SP3, and a correlation coefficient with the target variable of 0.556. After comparing it with the structural formula, the important group for irradiation damage in the molecule is the C-N ring, providing a data-driven basis for the irradiation chemical reaction of the molecule.

[0058] Among them, the atomic environment information includes the central atom, bond type, cyclic structure, and hybridization state.

[0059] Finally, it should be noted that the above description is only the preferred embodiment of the present invention. Under the inspiration of the present invention, those of ordinary skill in the art can make various similar representations without departing from the purpose and claims of the present invention, and such transformations all fall within the protection scope of the present invention.

Claims

1. A method for predicting the degree of radiation chemical damage of a molecular crystal, characterized in that: It includes the following steps: S1: Data preparation. Read the molecular crystal model file of the material to be analyzed, perform molecular dynamics simulation and irradiation simulation using the LAMMPS software package, obtain data and construct a training dataset, where the number of displaced atoms and the number of molecular decompositions are used as the target variables to be predicted; S2: Data preprocessing. Process abnormal data, augment the dataset by data enhancement, expand the feature set by feature engineering, and screen feature quantities; S3: Build a prediction model and model evaluation. Build a prediction model for the radiation chemical damage of molecular crystals based on XGBoost, divide the preprocessed dataset in step S2 into a training set and a validation set according to a ratio of 8:2, adopt a 5-fold cross-validation method, and evaluate the prediction model; S4: Sensitivity factor analysis. Use XGBoost in step S3 to calculate the contribution degree of each feature to the model prediction, and use the SHAP method to calculate the feature interaction effect of the contribution of each feature to obtain radiation-sensitive molecular group data.

2. The method according to claim 1, wherein: In step S1, the LAMMPS software package is used for molecular dynamics simulation, and the characteristic data of the molecular crystal model is specified for calculation; The characteristic data includes: model density, unit cell parameters, model volume, total number of atoms, number of various types of atoms, molar mass ratio of various types of atoms, initial potential energy of the system, initial temperature of the system, cohesive energy, energy of primary displaced atoms, type of primary displaced atoms (PKA); After obtaining the characteristic data of the molecular crystal model, further calculate the molecular chemical formula in the molecular crystal model, simplify the molecular linear input specification characters, number of molecules, molecular mass, number of various chemical bonds, and bond energy of the trigger bond in the molecule; After using the LAMMPS software package to perform irradiation simulation on the molecular crystal model, calculate the number of displaced atoms and the number of molecular decompositions during the irradiation process through the simulation trajectory data file.

3. The method according to claim 2, wherein: Integrate the above data to form a training dataset with fields of model density, unit cell parameters, model volume, total number of atoms, number of various types of atoms, molar mass ratio of various types of atoms, initial potential energy of the system, initial temperature of the system, cohesive energy, energy of primary displaced atoms, type of primary displaced atoms, molecular chemical formula, simplified molecular linear input specification characters, number of molecules, molecular mass, number of various chemical bonds, bond energy of the trigger bond in the molecule, number of displaced atoms and number of molecular decompositions.

4. The method according to claim 1, wherein: The abnormal data in step S2 includes outliers and missing values, and the processing is filled with the mode of the corresponding numerical values in the field.

5. The method according to claim 4, wherein: In step S2, the dataset is augmented by data enhancement based on grid interpolation and Monte Carlo methods, and a physical model for predicting NRT irradiation defects with a temperature correction term is artificially constructed and carried out under the constraints of the physical model.

6. The method according to claim 5, characterized in that: The formula of the physical model for predicting NRT irradiation defects with a temperature correction term is as follows: where N d is the number of displaced atoms, E PKA is the energy of the primary displaced atom, E d is the displacement threshold energy, k B is the Boltzmann constant, T is the temperature. The influence of temperature on the displacement threshold energy of the material is considered in the construction of this formula, and the thermal activation effect of atoms when the PKA energy is lower than the displacement threshold energy is described using the Boltzmann factor.

7. The method according to claim 6, wherein: The feature engineering performs numerical processing on three types of non-numerical features in the dataset, namely the primary off-site atomic species, molecular chemical formula, and simplified molecular linear input specification characters. Among them, the primary off-site atomic species uses one-hot encoding, the molecular chemical formula is transformed into the numerical values of the number of different types of atoms, and the simplified molecular linear input specification characters use Morgan molecular fingerprints to calculate their structural information; And train a random forest model to judge the importance of each feature for the target variable and conduct a correlation matrix analysis, delete the features with a correlation coefficient > 0.95, and retain the features with an importance greater than the average value.

8. The method according to claim 1, wherein: In the S3 step, to prevent model overfitting, if the performance of the validation set does not improve in 50 consecutive iterations, stop training and return the model with the best performance on the validation set.

9. The method according to claim 8, wherein: The best-performance model uses the Bayesian optimization method to systematically optimize the key hyperparameters of XGBoost; by constructing a probability model of the objective function and guiding the hyperparameter search based on the observed performance.

10. The method according to claim 1, wherein: In the S4 step, according to the molecular fingerprint data in the important features, use the Rdkit library to reverse-infer the structures in the molecule that are more sensitive to damage under irradiation load, and obtain the irradiation-sensitive molecular group data, providing a data-driven basis for molecular irradiation chemical reactions.

Images