Liquid crystal polyimide molecule design method based on molecule descriptor and machine learning
By combining molecular descriptors with machine learning, a predictive model with physical constraints is constructed to achieve targeted optimization of liquid crystal polyimide molecular design. This solves the problems of low efficiency and poor reliability in traditional design and supports the industrialization of liquid crystal polyimide.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 烟台国工智能科技有限公司
- Filing Date
- 2026-01-14
- Publication Date
- 2026-04-28
AI Technical Summary
Traditional liquid crystal polyimide molecule design relies on experience accumulation, has a long development cycle, low efficiency, high computational complexity, and makes it difficult to achieve rapid screening and optimization of large-scale molecular structures. Furthermore, the design process lacks multi-source data fusion and physical constraints, resulting in poor syntheticability of the designed molecular structures and performance that does not meet actual needs.
By combining molecular descriptors with machine learning, a predictive model with physical constraints is constructed. Through forward prediction and reverse generation models, performance-oriented directed molecular design is achieved. Combining reaction templates and chemical knowledge graphs, synthetic routes and experimental conditions are output.
It significantly shortens the design cycle, improves the targeting and accuracy of the design, ensures that the model prediction results conform to physical laws, reduces design risks, enhances the reliability and practicality of the design results, and supports the industrial development of liquid crystal polyimide.
Smart Images
Figure CN121938495A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of molecular intelligent design and process guidance technology, specifically to a molecular design method for liquid crystal polyimide based on molecular descriptors and machine learning. Background Technology
[0002] Liquid crystal polyimides, as an advanced functional material that combines the unique orderliness of liquid crystal materials with the excellent comprehensive properties of polyimides, have irreplaceable application value in high-end fields such as aerospace, microelectronics, and flexible displays. Their properties are not only closely related to the chemical structure of the molecular chains, but also influenced by the coupling of multi-scale factors such as the molecular aggregate structure and thermodynamic properties, resulting in a complex nonlinear relationship between molecular structure and properties.
[0003] Traditional liquid crystal polyimide molecular design relies heavily on researchers' experience and repeated trial-and-error experiments, resulting in long development cycles, low efficiency, and high costs. While theoretical methods such as molecular dynamics simulations and quantum chemical calculations have been applied in materials design, these methods are often computationally complex and time-consuming, making it difficult to achieve rapid screening and optimization of large-scale molecular structures. Furthermore, current design processes lack effective fusion of multi-source data and do not fully consider the impact of physical constraints on molecular properties, potentially leading to poor syntheticability and performance that does not meet practical application requirements.
[0004] Furthermore, traditional design methods are mostly forward exploration modes, which make it difficult to accurately deduce the optimal molecular structure from a clear target performance, severely restricting the targeted development and industrialization process of liquid crystal polyimide materials. Summary of the Invention
[0005] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: This application provides a method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning, including: The molecular descriptor extraction module automatically reads the input liquid crystal polyimide molecular structure file and calculates the molecular descriptor. Thermodynamic and molecular simulation data of liquid crystal polyimide systems are calculated using molecular dynamics or quantum chemical methods; Based on the molecular descriptor and the thermodynamic and molecular simulation data, combined with online sensor data, a machine learning prediction model with physical constraints is constructed. After using the machine learning model to make a positive prediction of the performance of liquid crystal polyimide, based on the prediction results, a reverse generation model is used to generate candidate molecular structures that meet the target performance from the set target performance. By using reaction templates and chemical knowledge graphs, synthetic routes are deduced from the candidate molecular structures, and feasible synthetic routes and experimental conditions are output.
[0006] In one possible implementation, the molecular descriptor includes a topological descriptor, an electronic structure descriptor, intermolecular interaction parameters, and anisotropy parameters. The topological descriptor includes molecular weight, branching coefficient, polar surface area, and aromatic ring density. The electronic structure descriptor includes the HOMO-LUMO band gap and dipole moment. The intermolecular interaction parameters include potential energy, hydrogen bond density, and van der Waals energy. The anisotropy parameters include optical anisotropy parameters.
[0007] In one possible implementation, the formula for calculating molecular weight is: in, For atoms atomic mass; The formula for calculating optical anisotropy parameters is: in, For optical anisotropy parameters, Represents parallel direction. Represents the vertical direction; The formula for calculating the dipole moment is: in, It is the dipole moment vector. For atomic charge, For atomic coordinates, is the absolute length of the dipole moment vector.
[0008] In one possible implementation, the calculation of thermodynamic and molecular simulation data of the liquid crystal polyimide system using molecular dynamics or quantum chemical methods includes: The free energy, axial sequence parameter, radial distribution function, and entropy change of the liquid crystal polyimide system were calculated using molecular dynamics or quantum chemical methods. The calculation formulas are as follows: in, Let ΔH be the free energy, ΔS be the enthalpy change of the reaction, ΔS be the entropy of the reaction, T be the temperature, and θ be the angle between the principal axis of the molecule and the average orientation axis. For sequence parameters, It is a radial distribution function. Radial distance, Number density, For the Dirac function, Let i be the distance between particle i and particle j. For entropy change, Here, denoted by Boltzmann, and det, is the determinant. Let e be the reduced Planck constant, and e be the base of the natural logarithm. Let M be the covariance matrix and M be the mass matrix.
[0009] In one possible implementation, a physically constrained machine learning prediction model is constructed based on the molecular descriptor and the thermodynamic and molecular simulation data, combined with online sensor data, including: The acquired molecular descriptors, thermodynamic and molecular simulation data, and processed online sensor data are fused to construct a multi-source dataset, which is then divided into a training set, a validation set, and a test set according to a preset ratio. Determine the physical constraints and transform them into regularization terms or constraint expressions for the machine learning model; A graph neural network is selected to construct the prediction model. The prediction model is trained using a training set, and the hyperparameters of the prediction model are adjusted using a validation set. The model prediction accuracy was verified using a test set, and physical constraints were introduced to correct the model output to ensure that the model prediction results conformed to physical laws. The trained model is evaluated for performance. When the evaluation metrics meet the preset requirements, it is determined as the final machine learning prediction model.
[0010] In one possible implementation, the loss function of the machine learning prediction model is calculated as follows: Where y is the true value. Let λ be the predicted value, and λ be the weighting coefficient. For the calculated free energy change, This refers to the free energy change measured experimentally.
[0011] In one possible implementation, after forward prediction of the properties of the liquid crystal polyimide using the machine learning model, a reverse generation model is used to generate candidate molecular structures that satisfy the set target properties, based on the prediction results, including: The molecular descriptor of the liquid crystal polyimide to be predicted is input into the machine learning prediction model for positive prediction, and the corresponding performance prediction results are obtained. Based on the prediction results, and with the target performance as input, a search is performed in the molecular space constrained by chemical validity using reinforcement learning or generative adversarial networks. New molecular structures are constructed atom-by-atom or functional group through a policy network until the reward function is maximized, thus obtaining candidate molecular structures that meet the requirements.
[0012] In one possible implementation, the reward function is calculated as follows: Where R is the reward function, , , These are the weighting coefficients. Let be the objective function. Due to the difficulty of synthesizing energy or conducting experiments, It represents the absolute value of the change in free energy.
[0013] In one possible implementation, the step of reasoning about the candidate molecular structure using reaction templates and chemical knowledge graphs to output feasible synthetic routes and experimental conditions includes: The candidate molecular structure was disassembled to obtain simple chemical structure fragments; The disassembled structural fragments were matched with reaction templates and chemical knowledge graphs respectively to output feasible synthetic routes and experimental conditions. The calculation formulas are as follows: Where Route represents the synthesis route. As a reaction template, Represents a set of molecular fragments; in, , , These are the weighting coefficients. Number of steps; To estimate total cost or risk level; The availability of starting reagents.
[0014] Compared with the prior art, the beneficial effects of this application are as follows: This application constructs a machine learning prediction model with physical constraints by integrating molecular descriptors, thermodynamic and molecular simulation data, and online sensor data. This model can accurately predict the performance of liquid crystal polyimide. At the same time, by using a reverse generation model, it can reverse-engineer candidate molecular structures from the set target performance. This breaks the traditional experience-driven and trial-and-error exploration design mode, realizes performance-oriented directional molecular design, and greatly improves the targeting and accuracy of the design.
[0015] The automatic extraction and batch calculation of molecular descriptors, efficient simulation analysis of thermodynamic data, and the rapid prediction capabilities of machine learning models in this application significantly shorten the molecular design cycle and avoid the large amount of repetitive experimental and computational work in traditional methods. Meanwhile, the introduction of physical constraints ensures that the model prediction results conform to basic physical laws, and the application of reaction templates and chemical knowledge graphs guarantees the syntheticability of candidate molecules, effectively reducing design risks and improving the reliability and practicality of the design results.
[0016] This application incorporates key factors such as target performance, free energy change, and synthetic difficulty into a reward function, achieving a dynamic balance between molecular performance optimization and synthetic feasibility during the reverse design process. This ensures that the designed molecules meet the performance requirements of the target application scenario while reducing the difficulty and cost of industrial production. Furthermore, the automated reasoning of the synthetic route and the explicit output of experimental conditions provide clear and reliable technical guidance for subsequent experimental verification and industrial production, accelerating the transformation process from molecular design to product commercialization.
[0017] This application can achieve targeted optimization of performance while maintaining molecular syntheticability, providing an automated and intelligent design platform for the industrial development of liquid crystal polyimide, and has significant innovation, practicality and promotion value. Attached Figure Description
[0018] Figure 1 A flowchart illustrating a method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning, provided for embodiments of this application; Figure 2 A schematic diagram of the liquid crystal polyimide molecular structure generated by the molecular descriptor extraction module provided in this application embodiment, which automatically reads the input liquid crystal polyimide molecular structure file. Figure 3 The molecular weight distribution histogram generated by the multi-scale thermodynamics and molecular simulation module provided in the embodiments of this application; Figure 4 The performance effect diagram generated by the simplified performance prediction and reverse engineering module provided in the embodiments of this application. Detailed Implementation
[0019] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.
[0020] Figure 1 A flowchart illustrating a molecular design method for liquid crystal polyimide based on molecular descriptors and machine learning, provided for embodiments of this application, is shown below. Figure 1 This embodiment presents a molecular design method for liquid crystal polyimide based on molecular descriptors and machine learning, comprising: S101, the molecular descriptor extraction module automatically reads the input liquid crystal polyimide molecular structure file and calculates the molecular descriptor.
[0021] See Figure 2 and Figure 3 In this embodiment, several liquid crystal polyimide molecular structures are input into the molecular descriptor extraction module in the form of SMILES. The molecular descriptor extraction module automatically identifies the file type and processes multiple molecules in batches. Specifically, the polyimide SMILES form input into the molecular descriptor extraction module is as follows: O=C1NC(=O)c2cc(Oc3cccc4c(Oc5ccc6c(c5)C(=O)N(c5ccccc5)C6=O)cccc34)ccc21. The module automatically calculates the topological descriptor, electrical descriptor, intermolecular interaction parameters, and anisotropy parameters, and displays the specific gridded two-dimensional molecular structure and descriptor distribution histogram through a visualization interface.
[0022] In this embodiment, the topological descriptor includes molecular weight, branching coefficient, polar surface area, and aromatic ring density; the electronic structure descriptor includes HOMO-LUMO band gap and dipole moment; the intermolecular interaction parameters include potential energy, hydrogen bond density, and van der Waals energy; and the anisotropy parameters include optical anisotropy parameters.
[0023] The formula for calculating molecular weight is: in, For atoms atomic mass; The formula for calculating optical anisotropy parameters is: in, For optical anisotropy parameters, Represents parallel direction. Represents the vertical direction; The formula for calculating the dipole moment is: in, It is the dipole moment vector. For atomic charge, For atomic coordinates, is the absolute length of the dipole moment vector.
[0024] The expressions for HOMO / LUMO energy and bandgap are: In the formula The values are obtained from DFT / TD-DFT / quantum computing.
[0025] S102 uses molecular dynamics or quantum chemistry methods to calculate thermodynamic and molecular simulation data for liquid crystal polyimide systems.
[0026] In this embodiment, the free energy, axial sequence parameter, radial distribution function, and entropy change of the liquid crystal polyimide system are calculated using molecular dynamics or quantum chemistry methods. The calculation formulas are as follows: in, Let ΔH be the free energy, ΔS be the enthalpy change of the reaction, ΔS be the entropy of the reaction, T be the temperature, and θ be the angle between the principal axis of the molecule and the average orientation axis. For sequence parameters, When the liquid crystal orientation is higher, S is closer to 1. It is a radial distribution function. Radial distance, Number density, Let i be the distance between particle i and particle j. For entropy change, Here, denoted by Boltzmann, and det, is the determinant. Let e be the reduced Planck constant, and e be the base of the natural logarithm. Let M be the covariance matrix and M be the mass matrix.
[0027] S103, based on molecular descriptors and thermodynamic and molecular simulation data, combined with online sensor data, constructs a machine learning prediction model with physical constraints.
[0028] In this embodiment, the acquired molecular descriptors, the thermodynamic and molecular simulation data, and the processed online sensor data are fused to construct a multi-source dataset, which is then divided into a training set, a validation set, and a test set according to a preset ratio. Determine the physical constraints and transform them into regularization terms or constraint expressions for the machine learning model; A graph neural network is selected to construct the prediction model. The prediction model is trained using a training set, and the hyperparameters of the prediction model are adjusted using a validation set. The model prediction accuracy was verified using a test set, and physical constraints were introduced to correct the model output to ensure that the model prediction results conformed to physical laws. The trained model is then evaluated for performance. When the evaluation metrics meet preset requirements, it is selected as the final machine learning prediction model. The formula for calculating the loss function of the machine learning prediction model is: Where y is the true value. Let λ be the predicted value, and λ be the weighting coefficient. For the calculated free energy change, This refers to the free energy change measured experimentally.
[0029] S104. After using a machine learning model to make a positive prediction of the performance of liquid crystal polyimide, based on the prediction results, a reverse generation model is used to generate candidate molecular structures that meet the set target performance.
[0030] In this embodiment, the molecular descriptor of the liquid crystal polyimide to be predicted is input into the machine learning prediction model for positive prediction to obtain the corresponding performance prediction result. Based on the prediction results, and with the target performance as input, a search is performed in the molecular space constrained by chemical validity using reinforcement learning or generative adversarial networks. New molecular structures are constructed atom-by-atom or functional group using a policy network until the reward function is maximized, yielding candidate molecular structures that meet the requirements. The formula for calculating the reward function is: Where R is the reward function, , , These are the weighting coefficients. Let be the objective function. Due to the difficulty of synthesizing energy or conducting experiments, It represents the absolute value of the change in free energy.
[0031] S105, using reaction templates and chemical knowledge graphs, performs synthetic route reasoning on the candidate molecular structure and outputs feasible synthetic routes and experimental conditions.
[0032] The candidate molecular structure was disassembled to obtain simple chemical structure fragments; The disassembled structural fragments are matched with reaction templates and chemical knowledge graphs, rigorously verifying reaction feasibility, including functional group compatibility, reaction center geometry, and stereochemical constraints. Finally, in the output stage, the system successfully generates one or more feasible synthetic routes, outputting the feasible routes and experimental conditions. These routes detail the complete reaction sequence from starting materials dianhydrides and diamines to the target product, liquid crystal polyimide, including all intermediates, required reagents, and reaction conditions. This provides chemists with clear and reliable theoretical guidance for experimental design. The calculation formulas are as follows: Where Route represents the synthesis route. As a reaction template, Represents a set of molecular fragments; in, , , These are the weighting coefficients. Number of steps; To estimate total cost or risk level; The availability of starting reagents.
[0033] See Figure 4 This embodiment also provides a simplified performance prediction and reverse design AI module for devices without GPUs. This embodiment uses traditional machine learning methods instead of deep learning and simplifies AI molecular design. As shown in the figure, the system can predict the thermal properties of polyimide, such as glass transition temperature, thermal decomposition temperature, and exponential thermal stability score, as well as mechanical properties such as Young's modulus, tensile strength, and hardness. Analysis of the graph shows that the predicted modulus and hardness of the liquid crystal polyimide are relatively close, with the hardness being slightly higher but fluctuating less, generally exhibiting a trend of high strength and high hardness with small differences. Simultaneously, the system constructs multidimensional performance relationships between different properties; for example, high Tg is associated with high modulus and high stability; strength and hardness are positively correlated, but the thermal decomposition temperature remains at a relatively high level. Through model prediction and reverse design algorithms, the model can recommend five molecules with the best performance and list their glass transition temperature and other information.
[0034] Specifically, the simplified machine learning predictor uses random forests instead of graph neural networks, genetic algorithms instead of reinforcement learning, and RDKit descriptors plus Morgan fingerprints as features. This method is suitable for small to medium-sized datasets.
[0035] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.
[0036] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0037] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.
Claims
1. A method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning, characterized in that, include: The molecular descriptor extraction module automatically reads the input liquid crystal polyimide molecular structure file and calculates the molecular descriptor. Thermodynamic and molecular simulation data of liquid crystal polyimide systems are calculated using molecular dynamics or quantum chemical methods; Based on the molecular descriptor and the thermodynamic and molecular simulation data, combined with online sensor data, a machine learning prediction model with physical constraints is constructed. After making a positive prediction of the performance of liquid crystal polyimide using the machine learning model, a candidate molecular structure that meets the target performance is generated by reverse generation based on the prediction results, starting from the set target performance. By using reaction templates and chemical knowledge graphs, synthetic routes are deduced from the candidate molecular structures, and feasible synthetic routes and experimental conditions are output.
2. The method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning according to claim 1, characterized in that, The molecular descriptor includes a topological descriptor, an electronic structure descriptor, intermolecular interaction parameters, and anisotropy parameters. The topological descriptor includes molecular weight, branching coefficient, polar surface area, and aromatic ring density. The electronic structure descriptor includes the HOMO-LUMO band gap and dipole moment. The intermolecular interaction parameters include potential energy, hydrogen bond density, and van der Waals energy. The anisotropy parameters include optical anisotropy parameters.
3. The method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning according to claim 2, characterized in that, The formula for calculating molecular weight is: in, For atoms atomic mass, Molecular weight; The formula for calculating optical anisotropy parameters is: in, For optical anisotropy parameters, Represents parallel direction. Represents the vertical direction; The formula for calculating the dipole moment is: in, It is the dipole moment vector. For atomic charge, For atomic coordinates, is the absolute length of the dipole moment vector.
4. The method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning according to claim 1, characterized in that, The calculation of thermodynamic and molecular simulation data for the liquid crystal polyimide system using molecular dynamics or quantum chemical methods includes: The free energy, axial sequence parameter, radial distribution function, and entropy change of the liquid crystal polyimide system were calculated using molecular dynamics or quantum chemical methods. The calculation formulas are as follows: in, Let ΔH be the free energy, ΔS be the enthalpy change of the reaction, ΔS be the entropy of the reaction, T be the temperature, and θ be the angle between the principal axis of the molecule and the average orientation axis. For sequence parameters, It is a radial distribution function. Radial distance, Number density, For the Dirac function, Let i be the distance between particle i and particle j. For entropy change, Here, denoted by Boltzmann, and det, is the determinant. Let e be the reduced Planck constant, and e be the base of the natural logarithm. Let M be the covariance matrix and M be the mass matrix.
5. The method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning according to claim 1, characterized in that, Based on the molecular descriptor and the thermodynamic and molecular simulation data, combined with online sensor data, a machine learning prediction model with physical constraints is constructed, including: The acquired molecular descriptors, thermodynamic and molecular simulation data, and processed online sensor data are fused to construct a multi-source dataset, which is then divided into a training set, a validation set, and a test set according to a preset ratio. Determine the physical constraints and transform them into regularization terms or constraint expressions for the machine learning model; A graph neural network is selected to construct the prediction model. The prediction model is trained using a training set, and the hyperparameters of the prediction model are adjusted using a validation set. The model prediction accuracy was verified using a test set, and physical constraints were introduced to correct the model output to ensure that the model prediction results conformed to physical laws. The trained model is evaluated for performance. When the evaluation metrics meet the preset requirements, it is determined as the final machine learning prediction model.
6. The method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning according to claim 5, characterized in that: The formula for calculating the loss function of the machine learning prediction model is as follows: Where y is the true value. Let λ be the predicted value, and λ be the weighting coefficient. For the calculated free energy change, This refers to the experimentally measured change in free energy.
7. The method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning according to claim 1, characterized in that, After forward prediction of the properties of liquid crystal polyimide using the machine learning model, based on the prediction results, a reverse generation model is used to generate candidate molecular structures that meet the set target properties, including: The molecular descriptor of the liquid crystal polyimide to be predicted is input into the machine learning prediction model for positive prediction, and the corresponding performance prediction results are obtained. Based on the prediction results, and with the target performance as input, a search is performed in the molecular space constrained by chemical validity using reinforcement learning or generative adversarial networks. New molecular structures are constructed atom-by-atom or functional group through a policy network until the reward function is maximized, thus obtaining candidate molecular structures that meet the requirements.
8. The method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning according to claim 7, characterized in that, The formula for calculating the reward function is as follows: Where R is the reward function, , , These are the weighting coefficients. Let be the objective function. Due to the difficulty of synthesizing energy or conducting experiments, It represents the absolute value of the change in free energy.
9. The method for designing liquid crystal polyimide molecules based on molecular descriptors and machine learning according to claim 1, characterized in that, The process involves reasoning about the candidate molecule structure using reaction templates and chemical knowledge graphs, outputting feasible synthetic routes and experimental conditions, including: The candidate molecular structure was disassembled to obtain simple chemical structure fragments; The disassembled structural fragments were matched with reaction templates and chemical knowledge graphs respectively to output feasible synthetic routes and experimental conditions. The calculation formulas are as follows: Where Route represents the synthesis route. As a reaction template, Represents a set of molecular fragments; in, , , These are the weighting coefficients. Number of steps; To estimate total cost or risk level; The availability of starting reagents.