Construction method of asphalt molecular database and reverse design method

Abstract

The application belongs to the technical field of material science and artificial intelligence, and particularly relates to a construction method and a reverse design method of an asphalt molecule database.A series of initial molecular weights are randomly generated in a preset molecular weight range; average molecular structure parameters of a target asphalt sample are calculated by using a Brown-Ladner method according to element composition and density of the target asphalt sample and the initial molecular weights, and an approximate molecular formula corresponding to each initial molecular weight is derived; the number of atoms in each approximate molecular formula is fine-tuned according to the principle that the number of atoms of each element is an integer, and the accurate molecular weight of each molecular formula after fine-tuning is calculated; the accurate molecular weight is taken as a new input value, iteration is performed until the molecular formula and the corresponding average molecular structure parameters are stably matched, and a series of candidate molecular formulas are obtained; an initial candidate molecular structure is generated according to the candidate molecular formulas and the corresponding average molecular structure parameters by using an RDKit tool, abnormal molecular structures are removed, and an asphalt molecule database corresponding to the target asphalt sample is obtained.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of materials science and artificial intelligence, specifically relating to a method for constructing an asphalt molecule database and a reverse design method. Background Technology

[0002] As the primary binder of asphalt pavements, the performance degradation caused by aging and the efficiency of regeneration directly determine the service life and maintenance costs of asphalt pavements. A deep understanding of the "structure-performance" relationship mechanism of asphalt materials at the molecular scale is fundamental to solving this key scientific problem. However, traditional asphalt molecular research relies heavily on experimental extraction and single simulation methods, which have inherent limitations such as low molecular structure coverage and long performance prediction cycles, making it difficult to support a systematic explanation of aging mechanisms and the efficient development of regenerators.

[0003] Specifically, existing molecular research on asphalt suffers from the following technical shortcomings: Insufficient representativeness of molecular structures: Traditional pitch molecular model construction often relies on analytical methods such as Brown-Ladner, which are typically based on elemental analysis and proton nuclear magnetic resonance spectroscopy (NMR). 1 This method relies heavily on prior assumptions about molecular weight, leading to inconsistencies between the constructed molecular structure parameters and the actual molecular weight. Furthermore, the generated molecular structures are limited and cannot cover the thousands of different molecular configurations contained in the complex heavy oil system of asphalt, thus lacking structural breadth and diversity.

[0004] Molecular generation is highly random, making it difficult to guarantee physical consistency: In recent years, although molecular generation methods based on deep learning (such as recurrent neural networks and generative adversarial networks) have emerged, most of these methods only start from the grammatical rules of SMILES strings, and the generation process lacks constraints on the specific physicochemical properties of the target material (such as asphalt) (such as elemental composition, aromaticity, condensation index, etc.). Therefore, although a large number of generated molecules are grammatically valid, they are physically unreasonable (for example, the elemental ratios deviate significantly from experimental values, and the ring system structure is too anomalous), and cannot be used to simulate the behavior of real materials.

[0005] Performance prediction is inefficient, and reverse design capabilities are lacking: Performance evaluation of newly generated molecules traditionally relies on density functional theory (DFT) or molecular dynamics (MD) simulations. These methods offer high computational accuracy but are extremely expensive, failing to meet the demands of high-throughput screening of thousands of candidate molecules. Furthermore, current technologies lack a closed-loop design framework that can guide molecule generation and perform multi-objective optimization based on target performance.

[0006] Therefore, developing a new method for asphalt molecular design is of significant theoretical and engineering value for accelerating the research and development of asphalt materials. Summary of the Invention

[0007] The purpose of this invention is to provide a method for constructing an asphalt molecular database and a reverse design method, which generates an asphalt molecular database with novel structure and physical rationality through molecular weight iterative inversion and Brown-Ladner structural parameter matching.

[0008] To achieve the above objectives, the present invention provides a method for constructing an asphalt molecule database, comprising the following steps: S1. Randomly generate a series of initial molecular weights within a preset molecular weight range; S2. Based on the elemental composition and density of the target asphalt sample and the initial molecular weight, the average molecular structure parameters of the target asphalt sample are calculated using the Brown-Ladner molecular structure analysis method, and an approximate molecular formula corresponding to each initial molecular weight is derived. S3. Fine-tune the number of atoms in each approximate molecular formula according to the principle that the number of atoms of each element is an integer, and calculate the precise molecular weight of each molecular formula after fine-tuning. S4. Using the precise molecular weight as a new input value, repeat steps S2 and S3 for iteration until each molecular formula is stably matched with its corresponding average molecular structure parameter, thus obtaining a series of candidate molecular formulas and their corresponding average molecular structure parameters. S5. Using the RDKit tool, an initial candidate molecular structure is generated based on the candidate molecular formula and the corresponding average molecular structure parameters. After removing abnormal molecular structures, an asphalt molecular database corresponding to the target asphalt sample is constructed.

[0009] Furthermore, in step S1, the preset molecular weight range is 400~1000 g / mol; In step S2, the elemental composition includes the mass content of carbon, hydrogen, oxygen, nitrogen, and sulfur; the average molecular structure parameters include the hydrogen-to-carbon atomic ratio, corrected molar volume, aromaticity, condensation index, total number of carbon atoms, total number of rings, number of carbon atoms in aromatic rings, number of aromatic rings, number of cycloalkane rings, total number of hydrogen atoms, number of carbon atoms in cycloalkane portions, number of carbon atoms in alkane chains, carbon percentage of cycloalkanes, and carbon percentage of alkanes.

[0010] Furthermore, in step S5, the step of generating initial candidate molecules using the RDKit tool specifically includes: S501. Based on the candidate molecular formula and the corresponding average molecular structure parameters, first introduce fused ring aromatics and sulfur-containing heterocycles. S502. Based on the number of cycloalkane rings and the number of carbon atoms in the alkane chain, cyclohexane and alkyl side chains with branching characteristics are randomly introduced. S503. Based on the nitrogen and oxygen content, amino and hydroxyl groups are introduced; S504. Use the RDKit toolkit to complete the hydrogen atoms and use the UFF force field to minimize the energy to obtain the initial candidate molecular structure.

[0011] Furthermore, in step S5, the removal of abnormal molecular structures is achieved by constructing N isolated trees using the isolated forest algorithm, calculating the average path length of each molecular structure in each isolated tree and converting it into anomaly probability, and removing the abnormal molecular structures according to a preset anomaly ratio.

[0012] Furthermore, it also includes: S6. Using the RDKit tool, convert the molecular structures in the asphalt molecular database obtained by any of the above-described methods into SMILES sequences. S7. Randomly select a preset number of SMILES sequences from the SMILES sequences, and randomly extract a string from each selected SMILES sequence, input them into the pre-trained molecular generation model respectively, and output new SMILES sequences to generate new molecular structures. S8. The average structural parameters of the generated new molecular structure are calculated using the Brown-Ladner molecular structure analysis method. The new molecular structure is retained only when it matches its corresponding average structural parameters. Finally, the new molecular structure after screening is added to the asphalt molecular database to obtain the expanded asphalt molecular database.

[0013] Furthermore, in step S7, after the string is input into the model, the probability distribution of the next character output by the molecular generation model is adjusted and sampled based on the temperature sampling strategy. The sampled character is then concatenated to the end of the input prefix. The above process is repeated until a preset stopping condition is met to obtain the new SMILES sequence.

[0014] Furthermore, the molecular generation model is either a molecular generation model based on a long short-term memory network or a lightweight molecular language model based on a shallow Transformer encoder architecture. The pre-training of the molecular generation model based on the long short-term memory network is carried out by constructing a dataset using the asphalt molecule database, and then training is performed using the Teacher Forcing mechanism with cross-entropy loss as the loss function. The construction of the dataset includes: extracting a continuous string from the initial position of the SMILES sequence corresponding to each molecular structure in the asphalt molecular database, and using it and the complete SMILES sequence as input and output labels respectively to construct the dataset; The lightweight molecular language model based on the shallow Transformer encoder architecture consists of four Transformer encoders with an embedding dimension of 128. Each layer adopts a 4-head self-attention mechanism and a 512-dimensional feedforward network. The pre-training adopts a hybrid task consisting of masked language modeling and causal language modeling. It is first unsupervised pre-trained on the PubChem database and then fine-tuned using the asphalt molecule dataset. The masking language modeling is used to randomly mask part of the characters in the input SMILES sequence and predict its original value, while the causal language modeling is used to input the prefix of the SMILES sequence and predict the next character in sequence.

[0015] This invention also provides a method for reverse design of asphalt molecules, comprising the following steps: S9. Using the RDKit tool, extract the Morgan fingerprint from each molecule in the asphalt molecule database obtained by the above-described method of constructing the asphalt molecule database, and convert it into a numerical vector. S10. Input the numerical vector into the pre-trained performance index prediction model, and then output the predicted value of the corresponding performance index. S11. Compare the predicted values ​​of each performance index with the target values ​​of the performance index of the target asphalt sample, and select the molecules that are closest to the target values ​​as the final molecular structure of the target asphalt sample.

[0016] Furthermore, in step S11, the following multi-objective optimization scoring function is used for screening:

[0017] in, P i Indicates the first i Predicted values ​​of the performance indicators T i For the first i The target value of the performance indicator ω i Here, n represents the total number of performance indicators, and the lower the comprehensive score calculated by the multi-objective optimization scoring function, the closer the predicted value of the performance indicator is to the target value.

[0018] Furthermore, the performance metric prediction model is a random forest model or an XGBoost model; The input label of the dataset used for pre-training is the numerical vector corresponding to the asphalt molecule, and the output label is the performance index value corresponding to each asphalt molecule obtained by the GFN2-xTB method. The performance metrics include one or more of the following: predicted total energy, dipole moment, HOMO energy, LUMO energy, and HOMO-LUMO bandgap.

[0019] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages: 1. The method for constructing the asphalt molecular database provided by this invention generates a series of candidate molecular formulas based on the structural parameters generated by Brown-Ladner and through a molecular weight iterative inversion strategy. This effectively overcomes the assumption bias problem existing in classical methods and enriches the molecular weight distribution and structural diversity. Then, the RDKit tool is used to generate corresponding molecular structures based on the candidate molecular formulas and structural parameter constraints, demonstrating good reliability, diversity, and representativeness.

[0020] 2. The generated molecules are highly novel and have strong structural exploration capabilities: Using generative models such as LSTM and lightweight Transformer, asphalt molecules with a chemical similarity of 0.874, a diversity of 0.459, and a novelty of 0.996 with the training set were successfully generated. This means that more than 99% of the generated molecules are completely new structures, providing a powerful tool for exploring unknown chemical spaces.

[0021] 3. Breakthrough in both prediction efficiency and accuracy: The performance prediction module, built upon ensemble learning models such as Random Forest and XGBoost, improves prediction efficiency and accuracy for total energy, dipole moment, and HOMO. The model demonstrates high prediction accuracy for key properties such as the LUMO band gap (e.g., XGBoost achieves an R² of 0.954 for the predicted dipole moment, compared to only 0.093 Debye for MAE), and its computational speed is significantly faster than DFT, meeting the requirements for high-throughput screening. Combined with SHAP interpretability analysis, the influence mechanism of molecular structure features on performance is revealed, enhancing the model's credibility.

[0022] 4. The selected typical average molecular configurations exhibit excellent comprehensive performance: Several typical average molecular configurations selected through a multi-objective scoring function have been verified by molecular dynamics simulations. Their predictions for key properties such as density, solubility parameters, glass transition temperature, viscosity, and flow activation energy all show a high degree of agreement with experimental data. Their prediction accuracy surpasses that of mainstream models reported in existing literature, and their molecular structures are more simplified, providing a reliable molecular model for the study of asphalt aging mechanisms and the development of rejuvenators.

[0023] 5. This invention pioneers an end-to-end AI-driven reverse design framework for asphalt molecules, deeply integrating generative artificial intelligence, physical constraints, ensemble learning performance prediction, and multi-objective optimization screening for the first time, constructing a closed-loop reverse design platform for asphalt molecular systems. Compared to traditional methods relying on experimental trial and error or single simulations, this invention enables the reverse generation and screening of optimal molecular structures starting from target performance, significantly improving the efficiency and intelligence level of new material discovery, while taking into account molecular structure creativity, physical rationality, high-throughput screening efficiency, and reverse design capabilities. Attached Figure Description

[0024] Figure 2 This is a flowchart of the method for constructing the asphalt molecule database of the present invention; Figure 1 This is an overall flowchart of the method for reverse design of asphalt molecules and database construction based on generative artificial intelligence, as per the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0026] like Figure 1 and 2 As shown, this invention provides a method for constructing an asphalt molecule database to achieve efficient and intelligent screening of asphalt molecules for target performance, comprising the following steps: S1. Randomly generate a series of initial molecular weights within a preset molecular weight range; S2. Based on the elemental composition and density of the target asphalt sample and the initial molecular weight. M rand The average molecular structure parameters of the target asphalt sample were calculated using the Brown-Ladner molecular structure analysis method, and an approximate molecular formula corresponding to each initial molecular weight was derived. S3. Fine-tune the number of atoms in each approximate molecular formula according to the principle that the number of atoms of each element is an integer, and calculate the precise molecular weight of each fine-tuned molecular formula. M calc ; S4, The precise molecular weight M calcAs new input values, the average structural parameters are recalculated and the molecular formula is updated. Steps S2 and S3 are repeated iteratively until each molecular formula is stably matched with its corresponding average molecular structure parameters, resulting in a series of candidate molecular formulas and their corresponding average molecular structure parameters. S5. Using the RDKit tool, an initial candidate molecular structure is generated based on the candidate molecular formula and the corresponding average molecular structure parameters. After removing abnormal molecular structures, an asphalt molecular database corresponding to the target asphalt sample is constructed.

[0027] In step S1, the preset molecular weight range is 400~1000 g / mol.

[0028] In step S2, the elemental composition includes the mass content of carbon (C), hydrogen (H), oxygen (O), nitrogen (N), and sulfur (S); the average molecular structure parameters include the hydrogen-to-carbon atomic ratio, corrected molar volume, aromaticity, condensation index, total number of carbon atoms, total number of rings, number of carbon atoms in aromatic rings, number of aromatic rings, number of cycloalkane rings, total number of hydrogen atoms, number of carbon atoms in cycloalkane portions, number of carbon atoms in alkane chains, carbon percentage of cycloalkanes, and carbon percentage of alkanes.

[0029] The core parameters of the Brown-Ladner method are calculated as follows: Hydrogen-to-carbon atomic ratio: (1); Corrected molar volume: (2); Aromaticity: (3); Condensation index: (4). Wherein, H / C represents the ratio of hydrogen atoms to carbon atoms in the average molecular structure. d This represents the density of the asphalt sample at 20℃. M c / d ) c This is the corrected molar volume based on each carbon atom. C% and H% refer to the mass percentage of carbon and hydrogen, respectively. f A is the aromaticity, and CI is the condensation index.

[0030] Furthermore, based on average molecular weight M More structural parameters can be calculated: Total number of carbon atoms: (5); Total number of rings: (6); Number of carbon atoms in the aromatic ring: (7); Number of aromatic rings: (8); Cycloalkanes: (9); Total number of hydrogen atoms: (10); Number of carbon atoms in the cycloalkane portion: (11); Number of carbon atoms in the alkane chain: (12); Cycloalkanes carbon fraction: (13); Alkane carbon content: (14). In the above formula, M C represents the average molecular weight of the asphalt sample. T The number of carbon atoms in the average molecular structure, C A It is the number of carbon atoms in the aromatic ring, R T R represents the total number of rings. A R represents the number of aromatic rings. N H represents the ring number of the cycloalkanes. T C represents the total number of hydrogen atoms. N It is the number of carbon atoms in the cycloalkane moiety, C P It represents the number of carbon atoms in the alkane chain.

[0031] The present invention effectively overcomes the assumption bias problem existing in the classical method through the molecular weight iterative inversion strategy in steps S2 to S4, and enriches the molecular weight distribution and structural diversity.

[0032] In step S5, the step of generating initial candidate molecules using the RDKit tool specifically includes: S501. Based on the candidate molecular formulas and corresponding average molecular structure parameters, first introduce polycyclic aromatic hydrocarbons (naphthalene, phenanthrene, pyrene, etc.) and sulfur-containing heterocycles (benzothiophene). S502, according to the number of cycloalkane rings R N and the number of carbon atoms in alkanes (C) P , cyclohexane and alkyl side chains with branching characteristics are randomly introduced; S503. Based on the nitrogen and oxygen content, amino and hydroxyl groups are introduced; S504. Use the RDKit toolkit to complete the hydrogen atoms and use the UFF force field to minimize the energy to obtain the initial candidate molecular structure.

[0033] Thus, a three-dimensional molecule is constructed using the modular assembly strategy in step S5.

[0034] In step S5, the removal of anomalous molecular structures involves identifying molecules with abnormal structures or properties using the Isolation Forest algorithm. By constructing N isolated trees, the average path length of each molecular structure within each isolated tree is calculated and converted into anomaly probability. The anomalous molecular structures are then removed according to a preset anomalous proportion (e.g., 3%).

[0035] In particular, this invention also performs unsupervised clustering of molecular descriptors using k-means and Gaussian mixture models (GMM), and determines the optimal number of clusters using silhouette coefficients, thereby classifying the database.

[0036] Furthermore, database construction methods also include: S6. Using the RDKit tool, the molecular structures in the asphalt molecular database obtained by any of the above-described methods are efficiently converted into novel asphalt molecule SMILES sequences that conform to chemical rules and satisfy physical consistency. S7. Randomly select a preset number of SMILES sequences from the SMILES sequences, and randomly extract a string from each selected SMILES sequence, input them into the pre-trained molecular generation model respectively, and output new SMILES sequences to generate new molecular structures. S8. The average structural parameters of the generated new molecular structure are calculated using the Brown-Ladner molecular structure analysis method. The new molecular structure is retained only when it matches its corresponding average structural parameters. Finally, the new molecular structure after screening is added to the asphalt molecular database to obtain the expanded asphalt molecular database.

[0037] Specifically, the molecular generation model is either a molecular generation model based on a long short-term memory network or a lightweight molecular language model based on a shallow Transformer encoder architecture.

[0038] The structure of the molecular generation model based on long short-term memory networks is as follows: Embedding layer: 128-dimensional, maps SMILES characters to continuous vectors.

[0039] Two-layer LSTM backbone network: hidden layer dimension 256, dropout rate 0.2, used to learn long-term dependencies of sequences.

[0040] Fully connected output layer: outputs the probability distribution of the next character.

[0041] Pre-training was performed using the asphalt molecule database to construct a dataset, and then the Teacher Forcing mechanism was used for training, with the cross-entropy loss function as the loss function; the optimizer was Adam (initial learning rate 0.001).

[0042] The construction of the dataset includes: extracting a continuous string from the initial position of the SMILES sequence corresponding to each molecular structure in the asphalt molecular database, and using it and the complete SMILES sequence as input and output labels respectively to construct the dataset.

[0043] The lightweight molecular language model based on a shallow Transformer encoder architecture consists of a four-layer Transformer encoder with an embedding dimension of 128. Each layer employs a 4-head self-attention mechanism and a 512-dimensional feedforward network. Pre-training employs a hybrid task of masked language modeling (MLM) and causal language modeling (CLM) at a ratio of 8:2. Unsupervised pre-training is first performed on the PubChem database (>150 million molecules), followed by fine-tuning on the asphalt molecule dataset. Masked language modeling is used to randomly mask parts of the input SMILES sequence and predict their original values, while causal language modeling is used to input the prefix of the SMILES sequence and predict the next character sequentially. Data loading: A data sharding storage strategy (1 million sequences per shard) and multi-threaded queue transfer are used to address memory bottlenecks.

[0044] Specifically, in the pre-training phase, 150 million unlabeled SMILES sequences from the PubChem database were used as training corpus to perform unsupervised pre-training on the shallow Transformer encoder architecture. The pre-training task employed a hybrid approach combining masked language modeling (MLM) and causal language modeling (CLM): the former randomly masked parts of the input sequence and predicted their original values, enabling the model to learn the bidirectional contextual representation of molecular sequences; the latter predicted the next character by using the prefix of the input sequence, enabling the model to learn the left-to-right generation pattern of molecular SMILES sequences. The losses from the MLM and CLM tasks were weighted in an 8:2 ratio to form a joint loss function, which was used to jointly optimize the model parameters through backpropagation. After this phase, the model possesses general molecular sequence understanding and generation capabilities.

[0045] After pre-training, given the significant differences in structural characteristics and distribution patterns between asphalt molecules and general molecular databases, directly applying the pre-trained model to the asphalt molecule generation task would have limited effectiveness. Therefore, task-adaptive fine-tuning is necessary. During the fine-tuning phase, asphalt molecule SMILES sequence data was collected, with a scale on the order of one-thousandth of the pre-training data. Some low-level coding layer parameters in the pre-trained model were fixed, and only high-level parameters were fine-tuned. Training continued on the asphalt molecule dataset using a hybrid MLM and CLM task, with the MLM and CLM loss ratio adjusted to 1:1 and the learning rate set to one-tenth of the pre-training learning rate, until the joint loss function converged, resulting in a dedicated asphalt molecule generation model.

[0046] In step S7, after the string is input into the model, the probability distribution of the next character output by the molecular generation model is adjusted and sampled based on the temperature sampling strategy (temperature=0.8, adjusting generation diversity). The sampled character is then concatenated to the end of the input prefix. The above process is repeated until the preset stopping condition is met to obtain the new SMILES sequence.

[0047] In step S8, the key structural parameters (such as aromaticity) calculated by the Brown-Ladner method are... f A Condensation index (CI), total carbon number (C) T Total number of rings R T (e.g., as hard or soft constraints) in the generative model. For example, during the generation process, a molecule is only retained if the generated molecule matches the calculated structural parameters.

[0048] This invention also provides a method for reverse design of asphalt molecules, comprising the following steps: S9. Using the RDKit tool, extract the Morgan fingerprint from each molecule in the asphalt molecule database obtained by the above-described method of constructing the asphalt molecule database, and convert it into a numerical vector. S10. Input the numerical vector into the pre-trained performance index prediction model, and then output the predicted value of the corresponding performance index. S11. Compare the predicted values ​​of each performance index with the target values ​​of the performance index of the target asphalt sample, and select the molecules that are closest to the target values ​​as the final molecular structure of the target asphalt sample.

[0049] Specifically, in step S9, the RDKit toolkit is used to extract structural descriptors from each valid molecule. For example, a 64-dimensional binary vector is generated using the Morgan fingerprint (ECPF2) with a radius of 2. All numerical features are Z-score normalized.

[0050] In step S10, the performance index prediction model is either a random forest model or an XGBoost model; the random forest model is used to predict the total energy, dipole moment, HOMO energy, LUMO energy, and HOMO-LUMO bandgap. E gap Model performance is measured by mean absolute error (MAE) and coefficient of determination (R²). 2 The root mean square error (RMSE) is used for evaluation. The Extreme Gradient Boosting (XGBoost) model is also used for predicting the above properties, and its prediction accuracy is better than that of Random Forest.

[0051] The pre-training dataset uses numerical vectors corresponding to asphalt molecules as input labels and performance index values ​​for each asphalt molecule obtained through the GFN2-xTB method as output labels. The GFN2-xTB method is used to optimize the geometry and calculate single-point energies of the generated molecules to achieve efficient electronic structure evaluation in the high-throughput screening process. All calculations are performed automatically by calling the xTB program.

[0052] In step S11, the following multi-objective optimization scoring function is used for screening:

[0053] in, P i Indicates the first i Predicted values ​​of the performance indicators T i For the first i The target value of the performance indicator ω i Here, n represents the total number of performance indicators, and the lower the comprehensive score calculated by the multi-objective optimization scoring function, the closer the predicted value of the performance indicator is to the target value.

[0054] The SHapley Additive exPlanations (SHAP) method is used to explain the decision-making mechanism of the prediction model. SHAP values ​​quantify the positive or negative contribution of each molecular feature (such as MoILogP, molecular weight, and specific fingerprint sites) to the prediction results. SHAP force maps are plotted to visually demonstrate the decision-making process for individual molecule predictions.

[0055] Finally, the performance of the selected typical average molecular configurations was verified using classical molecular dynamics (MD) simulations and semi-empirical quantum chemical methods.

[0056] 1. Classical MD Simulation Method Software and Force Fields: LAMMPS software was used, with COMPASS force fields employed.

[0057] Model construction: 40 asphalt molecules were placed in an initial density of 1.0 g / cm³. 3 In the cube simulation box.

[0058] Simulation conditions: First, relax for 5 ns under the NVT ensemble, then simulate for 105 ns under the NPT ensemble (1 atm). Temperature range: 273.15 K to 458.15 K, time step: 1 fs.

[0059] Glass transition temperature ( T gMeasurement: The specific volume-temperature curve was measured by a stepwise cooling process (600K to 80K, cooling rate 5K / ns) under the NPT ensemble. T g It is defined as the intersection of the fitted lines of the glassy and rubbery regions.

[0060] Viscosity calculation: Viscosity is predicted by combining nonequilibrium molecular dynamics (NEMD), the Einstein-Stokes equation, and the Rouse model.

[0061] Einstein-Stokes equations: (16) Rouse model: (17) in, η Viscosity, D Where is the diffusion coefficient. T For temperature, R g Radius of gyration k B Boltzmann's constant, R Let be the ideal gas constant. M It is the molar mass.

[0062] Radius of gyration: (18) Where N represents the number of atoms in the molecule. r i Let represent the position vector of the i-th atom, and r cm This represents the position of the center of mass of the asphalt molecule.

[0063] Diffusion coefficient (based on mean square displacement MSD): (19) in, r i (0) and r i (t) Let represent the position vectors of the i-th molecule at the initial time and time t, respectively, where N is the total number of molecules in the system and D is the diffusion coefficient.

[0064] This invention constructs a molecular property prediction model based on ensemble learning, thereby enabling rapid and accurate assessment of the key physicochemical properties of newly generated molecules.

[0065] In summary, this invention constructs an asphalt molecule generation module that integrates physical consistency constraints, a molecular performance prediction module based on ensemble learning, and a multi-objective high-throughput screening module; by driving molecule generation and screening from the perspective of target performance, a closed-loop design process is formed.

[0066] This invention uses No. 70 base asphalt as the target material and elaborates on the implementation process and effects of the invention.

[0067] Example 1 (Method of the present invention, including physical constraints) Step 1: Experimental Data Acquisition and Target Parameter Setting First, the No. 70 base asphalt underwent routine performance tests, elemental composition analysis, and characterization using 1H NMR and IR spectroscopy. The measured elemental composition (carbon 83.78%, hydrogen 9.23%, sulfur 6.02%) and density at 20℃ (1.041 g / cm³) were used. 3 As input, the target performance is set as follows: total energy approximately -101.2 Hartree, HOMO-LUMO bandgap approximately 2.18 eV, and dipole moment approximately 2.09 Debye.

[0068] Step 2: Molecular generation and database construction based on physical consistency The Molecular Generator method of this invention is employed. First, the Brown-Ladner method (Equations 1 to 14) and molecular weight iterative inversion strategy (steps S1 to S4) are applied to obtain a series of candidate molecules and their corresponding average molecular structure parameters, including a total carbon number of 32, a total ring number of 5, an aromatic ring number of 3, a cycloalkane ring number of 2, and an alkane carbon number of 14. Using these parameters as targets, a three-dimensional molecular structure is generated using the modular assembly strategy (step S5). After energy minimization, the initial candidate molecules are obtained.

[0069] A total of 50,000 representative molecules were generated, forming the initial database. Subsequently, the Isolation Forest algorithm was used for anomaly detection, identifying approximately 1,500 potential anomalous molecules (about 3% of the total), demonstrating the overall reliability of the database. K-means and Gaussian mixture models were then used for cluster analysis of the database. The optimal number of clusters was determined to be 3 using the silhouette coefficient, clearly dividing the molecules into three clusters.

[0070] Step 3: Molecular generation and uncertainty analysis based on LSTM An LSTM molecular generation model was constructed and trained using an original dataset containing 49,820 asphalt molecules. A continuous string was then randomly extracted from these 49,820 asphalt molecules and input into the model, successfully generating 3,811 structurally valid and novel molecules. The generated molecules showed a chemical similarity of 0.874, a diversity of 0.459, and a novelty of 0.996 compared to the original dataset. Brown-Ladner physical constraints were then applied (i.e., molecules were validated using parameters such as aromaticity and condensation index; only molecules meeting preset physical parameter ranges were retained).

[0071] The uncertainty of the generation process was quantified using Shannon entropy. The proportion of uncertain characters for the vast majority of molecules was concentrated between 8% and 12%, with an average of 10.42%, indicating that the model generation process was stable. Analysis of a few high-uncertainty steps showed that the model might exhibit decision hesitation near complex ring structures, but its overall reliability was high.

[0072] Step 4: Molecular performance prediction and model interpretability Based on the GFN2-xTB calculation results of the generated molecules, random forest and XGBoost prediction models were constructed. Taking total energy prediction as an example, the random forest model had a mean absolute error of 0.422 eV and a root mean square error of 0.599 eV on the test set, with a coefficient of determination close to 1, and the predicted values ​​were in high agreement with the DFT calculation values. The XGBoost model performed better in dipole moment prediction, with a coefficient of determination of 0.954 and a mean absolute error of only 0.093 Debye.

[0073] The SHAP method was used to interpret the bandgap prediction model. Analysis showed that the lipid-water partition coefficient, specific molecular fingerprint sites, and molecular weight are the most important features affecting bandgap prediction. SHAP further revealed the positive and negative contributions of each feature in the prediction results of individual molecules.

[0074] Step 5: Multi-target high-throughput screening A multi-objective scoring function was used to screen the generated molecules. The target total energy was set to -101.2 Hartree, the band gap to 2.18 eV, and the dipole moment to 2.09 Debye, with corresponding weights assigned. The 10 molecules with the lowest scores (closest to the target) were selected, with the first and second ranked molecules having scores of 0.060 and 0.072, respectively, and their predicted properties highly consistent with the target.

[0075] Step 6: Simulation verification of typical molecular configurations The three molecules with the lowest scores were named LYM1, LYM2, and LYM3. Their performance was verified using molecular dynamics simulations and GFN2-xTB calculations.

[0076] Density and solubility parameters: The density predictions of the three models differ from the experimental values ​​(1.01-1.04 g / cm³) in the low-temperature region (250-350 K). 3 The results show a high degree of agreement. In the high-temperature region, the model prediction of this invention (approximately 0.93 g / cm³) is highly consistent. 3 It is closer to the experimental values ​​than existing literature models. The solubility parameter shows a reasonable trend with temperature change.

[0077] Glass transition temperatures: The glass transition temperatures of LYM1, LYM2, and LYM3 are 262.27 K, 251.93 K, and 258.50 K, respectively. Among them, the glass transition temperature of LYM2 is closest to the experimental value (251.57 K), showing higher prediction accuracy.

[0078] Viscosity and flow activation energy: The viscosity value of LYM2 calculated based on the Rouse model showed the best agreement with experimental values ​​across the entire temperature range. The flow activation energies of all three models were within the experimental range (30-80 kJ / mol), with LYM1 and LYM2 exhibiting the best performance.

[0079] Example 2 (Method of the present invention: Lightweight Transformer + Physical Constraints) This embodiment is essentially the same as Embodiment 1, except that the molecular generation model uses a lightweight molecular language model (based on a 4-layer Transformer encoder) and also incorporates Brown-Ladner physical constraints. The model is first pre-trained on the PubChem database to learn the general grammar rules of molecules, and then fine-tuned on the bituminous molecule dataset. This model can also generate bituminous molecules with high novelty (greater than 0.99), and due to the use of a more efficient attention mechanism, its generation speed and stability when processing long sequences of shorts are superior to the LSTM model. The optimal molecule selected at the end performs comparably to that of Embodiment 1, but with an efficiency improvement of approximately 30%.

[0080] Comparative Example 1 (Ordinary LSTM, without physical constraints) This comparative example uses the same LSTM network structure and training parameters as Example 1, but without any Brown-Ladner physical constraints. That is, the model learns and generates solely based on the grammatical rules of the SMILES string, without imposing restrictions on target parameters such as element composition, aromaticity, or condensation index.

[0081] After training on the same asphalt molecule training set, the model generated 5000 candidate molecules. Statistically, the effective resolution rate of the generated molecules using SMILES was 87.5%, lower than the 99.2% in Example 1. Although the novelty was as high as 0.998, a large number of the generated molecules exhibited significant physical inconsistencies: for example, the average aromaticity was only 0.28 (target value 0.44), the average total carbon number was 23 (target 32), and approximately 35% of the molecules lacked sulfur (while the raw material asphalt contains approximately 6% sulfur). In molecular dynamics simulations, these molecules could not reproduce the experimentally measured density and glass transition temperature, with deviations exceeding 20%.

[0082] Comparative Example 2 (Standard CNN, without physical constraints) This comparative example uses a one-dimensional convolutional neural network as the molecular generation model. The network structure includes an embedding layer, three convolutional blocks (each containing a convolutional layer, a batch normalization layer, and a pooling layer), and a fully connected output layer. No physical constraints are imposed.

[0083] The effective resolution of the generated molecules using SMILES was 82.1%, lower than that of ordinary LSTM. The novelty score was 0.995, but the ring structure distribution of the generated molecules deviated significantly from reality: the average total number of rings was only 2.1 (target 5), and almost no polycyclic aromatic hydrocarbons were present. Approximately 40% of the generated molecules were straight-chain alkanes or simple monocyclic compounds, completely inconsistent with the complex aromatic characteristics of asphalt. This comparative example further demonstrates that purely data-driven models lacking physical constraints are insufficient to capture the deep structural features of asphalt molecules.

[0084] Performance Comparison Summary Table 1 below summarizes the comparison results of Examples 1-2 and Comparative Examples 1-2 on key indicators.

[0085] Table 1. Comparison of molecular generation performance between the examples and the comparative examples.

[0086]

[0087] As shown in Table 1, compared with ordinary LSTM and CNN models without physical constraints, the embodiments of this invention (containing Brown-Ladner physical constraints) significantly improve the effective generation rate of SMILES, physical consistency (the degree of agreement between aromaticity, total carbon number, elemental composition and target values) of the generated molecules while maintaining high novelty and diversity, and greatly reduce the uncertainty of the generation process. The physical rationality pass rate increased from less than 50% to over 96%, demonstrating the effectiveness and superiority of the physical constraint embedding strategy proposed in this invention.

[0088] Comparative Example 3: Comparison with the traditional Brown-Ladner method An average molecular model was constructed using the traditional Brown-Ladner method (without molecular weight iterative inversion, directly assuming a molecular weight of 800 g / mol). The calculated structural parameters deviated significantly from the results after iterative inversion: the uncorrected aromaticity was 0.454, while it stabilized at 0.438 after iterative inversion; the total carbon number decreased from 38 to 32. This indicates that the molecular weight iterative inversion strategy of this invention significantly enhances the physical rationality and internal consistency of the model. Furthermore, traditional methods can only generate one or a few average molecules at a time, failing to construct a large-scale database containing tens of thousands of representative molecules and lacking high-throughput screening capabilities.

[0089] In summary, this invention successfully constructs an AI-driven reverse design platform for asphalt molecules, integrating physically constrained molecular generation, high-performance prediction, high-throughput screening, and multi-scale simulation verification. Through comparative verification using specific embodiments and comparative examples, this method can efficiently generate asphalt molecules with diverse structures, sound physical properties, and high novelty, and rapidly screen candidate molecules whose performance meets expectations. Molecular dynamics simulations have verified that the selected typical average molecular configurations outperform existing literature models in predicting key properties such as density, glass transition temperature, and viscosity, providing a reliable molecular model and efficient tool for asphalt aging mechanism research and rejuvenator development.

[0090] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an asphalt molecular database, characterized in that, Includes the following steps: S1. Randomly generate a series of initial molecular weights within a preset molecular weight range; S2. Based on the elemental composition and density of the target asphalt sample and the initial molecular weight, the average molecular structure parameters of the target asphalt sample are calculated using the Brown-Ladner molecular structure analysis method, and an approximate molecular formula corresponding to each initial molecular weight is derived. S3. Fine-tune the number of atoms in each approximate molecular formula according to the principle that the number of atoms of each element is an integer, and calculate the precise molecular weight of each molecular formula after fine-tuning. S4. Using the precise molecular weight as a new input value, repeat steps S2 and S3 for iteration until each molecular formula is stably matched with its corresponding average molecular structure parameter, thus obtaining a series of candidate molecular formulas and their corresponding average molecular structure parameters. S5. Using the RDKit tool, an initial candidate molecular structure is generated based on the candidate molecular formula and the corresponding average molecular structure parameters. After removing abnormal molecular structures, an asphalt molecular database corresponding to the target asphalt sample is constructed.

2. The method for constructing an asphalt molecular database according to claim 1, characterized in that, In step S2, the preset molecular weight range in step S1 is 400~1000 g / mol; In step S2, the elemental composition includes the mass content of carbon, hydrogen, oxygen, nitrogen, and sulfur; the average molecular structure parameters include the hydrogen-to-carbon atomic ratio, corrected molar volume, aromaticity, condensation index, total number of carbon atoms, total number of rings, number of carbon atoms in aromatic rings, number of aromatic rings, number of cycloalkane rings, total number of hydrogen atoms, number of carbon atoms in cycloalkane portions, number of carbon atoms in alkane chains, carbon percentage of cycloalkanes, and carbon percentage of alkanes.

3. The method for constructing an asphalt molecular database according to claim 2, characterized in that, In step S2 and step S5, the steps of generating initial candidate molecules using the RDKit tool specifically include: S501. Based on the candidate molecular formula and the corresponding average molecular structure parameters, first introduce fused ring aromatics and sulfur-containing heterocycles. S502. Based on the number of cycloalkane rings and the number of carbon atoms in the alkane chain, cyclohexane and alkyl side chains with branching characteristics are randomly introduced. S503. Based on the nitrogen and oxygen content, amino and hydroxyl groups are introduced; S504. Use the RDKit toolkit to complete the hydrogen atoms and use the UFF force field to minimize the energy to obtain the initial candidate molecular structure.

4. The method for constructing a bitumen molecular database based on elemental composition and target performance according to claim 1, characterized in that, In step S5, the removal of abnormal molecular structures is achieved by constructing N isolated trees using the isolated forest algorithm, calculating the average path length of each molecular structure in each isolated tree and converting it into anomaly probability, and removing the abnormal molecular structures according to a preset anomaly ratio.

5. The method for constructing a bitumen molecular database based on elemental composition and target performance according to any one of claims 1-4, characterized in that, Also includes: S6. Using the RDKit tool, the molecular structures in the asphalt molecular database obtained by the construction method of the asphalt molecular database according to any one of claims 1-4 are converted into SMILES sequences respectively; S7. Randomly select a preset number of SMILES sequences from the SMILES sequences, and randomly extract a string from each selected SMILES sequence, input them into the pre-trained molecular generation model respectively, and output new SMILES sequences to generate new molecular structures. S8. The average structural parameters of the generated new molecular structure are calculated using the Brown-Ladner molecular structure analysis method. The new molecular structure is retained only when it matches its corresponding average structural parameters. Finally, the new molecular structure after screening is added to the asphalt molecular database to obtain the expanded asphalt molecular database.

6. The method for constructing an asphalt molecular database according to claim 5, characterized in that, In step S7, after the string is input into the model, the probability distribution of the next character output by the molecular generation model is adjusted and sampled based on the temperature sampling strategy. The sampled character is then concatenated to the end of the input prefix. The above process is repeated until the preset stopping condition is met, and the new SMILES sequence is obtained.

7. The method for constructing an asphalt molecular database according to claim 5, characterized in that, The molecular generation model is either a molecular generation model based on a long short-term memory network or a lightweight molecular language model based on a shallow Transformer encoder architecture. The pre-training of the molecular generation model based on the long short-term memory network is carried out by constructing a dataset using the asphalt molecule database, and then training is performed using the Teacher Forcing mechanism with cross-entropy loss as the loss function. The construction of the dataset includes: extracting a continuous string from the initial position of the SMILES sequence corresponding to each molecular structure in the asphalt molecular database, and using it and the complete SMILES sequence as input and output labels respectively to construct the dataset; The lightweight molecular language model based on the shallow Transformer encoder architecture consists of four Transformer encoders with an embedding dimension of 128. Each layer adopts a 4-head self-attention mechanism and a 512-dimensional feedforward network. The pre-training adopts a hybrid task consisting of masked language modeling and causal language modeling. It is first unsupervised pre-trained on the PubChem database and then fine-tuned using the asphalt molecule dataset. The masking language modeling is used to randomly mask part of the characters in the input SMILES sequence and predict its original value, while the causal language modeling is used to input the prefix of the SMILES sequence and predict the next character in sequence.

8. A method for reverse design of asphalt molecules, characterized in that, Includes the following steps: S9. Using the RDKit tool, extract the Morgan fingerprint from each molecule in the asphalt molecule database obtained by the construction method of the asphalt molecule database according to any one of claims 1-7, and convert it into a numerical vector; S10. Input the numerical vector into the pre-trained performance index prediction model, and then output the predicted value of the corresponding performance index. S11. Compare the predicted values ​​of each performance index with the target values ​​of the performance index of the target asphalt sample, and select the molecules that are closest to the target values ​​as the final molecular structure of the target asphalt sample.

9. The reverse design method for asphalt molecules according to claim 8, characterized in that, In step S11, the following multi-objective optimization scoring function is used for screening: in, P i Indicates the first i Predicted values ​​of the performance indicators T i For the first i The target value of the performance indicator ω i Here, n represents the total number of performance indicators, and the lower the comprehensive score calculated by the multi-objective optimization scoring function, the closer the predicted value of the performance indicator is to the target value.

10. The reverse design method for asphalt molecules according to claim 8, characterized in that, The performance index prediction model is either a random forest model or an XGBoost model. The input label of the dataset used for pre-training is the numerical vector corresponding to the asphalt molecule, and the output label is the performance index value corresponding to each asphalt molecule obtained by the GFN2-xTB method. The performance metrics include one or more of the following: predicted total energy, dipole moment, HOMO energy, LUMO energy, and HOMO-LUMO bandgap.

Images