Generative molecule reverse design system based on reinforcement learning

Abstract

The invention relates to a generative molecule reverse design system based on reinforcement learning, which comprises a data set construction module, a multi-target performance prediction model establishment module, a pre-training module, a reward function construction module and an optimization module, and is characterized in that the data set construction module is used for constructing and screening to obtain a molecular structure performance data set; the multi-target performance prediction model establishment module is used for establishing a multi-target performance prediction model based on the constructed molecular structure performance data set; the pre-training module is used for pre-training a molecular generation model by using the screened molecular structure data; the reward function construction module is used for constructing a layered multi-target reward function; and the optimization module is used for rapidly evaluating key indexes by using a performance prediction model by adopting a reinforcement learning method, and carrying out optimization adjustment on the molecular generation model through a layered multi-target reward function. According to the invention, efficient and systematic reverse design of lithium metal negative electrode interface self-assembly molecules can be realized.

Description

Technical Field

[0001] This invention relates to a generative molecular reverse design system based on reinforcement learning. Background Technology

[0002] With the escalating global energy crisis, the search for new, efficient, and environmentally friendly energy storage technologies has become a hot topic in the field of science and technology. Electrochemical batteries, due to their high energy density, fast response speed, and flexible structure, have become a key energy storage unit in new energy systems. As a representative technology, lithium metal batteries boast extremely high theoretical specific capacity (up to 3860 mAh·g). -1 With its low negative electrode potential (-3.04 V vs. standard hydrogen electrode), lithium metal is considered an ideal choice for next-generation high-energy-density batteries, and the performance of the lithium metal anode plays a decisive role in the safety and cycle life of the entire battery system.

[0003] Traditional functional molecular design for lithium metal anode interfaces relies heavily on empirical judgment and trial-and-error. In practice, it often uses a single performance indicator as the design basis, such as focusing solely on film-forming ability or reduction stability, making it difficult to consider other key factors affecting interfacial performance, such as molecular stability, mechanical strength, and electronic structure. This approach is not only time-consuming and inefficient but also prone to insufficient performance trade-offs, hindering the systematic optimization of molecular performance. For example, while reduction potential significantly influences interfacial reaction behavior, this indicator alone cannot comprehensively reflect the molecule's overall performance in terms of film-forming ability, electrochemical window, and interfacial stability. Therefore, a novel design strategy that comprehensively considers material properties across multiple dimensions and supports automated exploration is urgently needed to overcome the limitations of traditional methods.

[0004] To improve screening efficiency, methods combining high-throughput virtual screening with machine learning prediction models have emerged in recent years. These methods construct structure-property mapping relationships to achieve rapid performance prediction and initial screening of existing molecular structures. While these methods offer significant improvements in efficiency and screening scale compared to traditional trial-and-error designs, they remain essentially limited to screening existing molecular libraries, lacking the ability to actively explore structural space and failing to achieve reverse generation and innovation of molecular structures based on target performance. Furthermore, under multi-objective performance requirements, these methods often employ weighted scoring or post-processing strategies, making it difficult to achieve dynamic trade-offs and synergistic optimization among performance indicators. This results in final screening results that fail to fully meet the comprehensive performance requirements of functional molecules in complex interface environments. Summary of the Invention

[0005] In view of this, it is necessary to provide a reinforcement learning-based generative molecular reverse design system that can provide a novel strategy that takes into account multiple performance indicators, improves the design efficiency of self-assembled molecules, and is applicable to precise interface control, so as to achieve efficient and systematic reverse design of self-assembled molecules at the lithium metal anode interface.

[0006] This invention provides a generative molecular reverse design system based on reinforcement learning. The system includes a dataset construction module, a multi-objective performance prediction model building module, a pre-training module, a reward function construction module, and an optimization module. Specifically: the dataset construction module constructs and filters a molecular structure performance dataset based on existing literature, databases, and experimental and computational results; the multi-objective performance prediction model building module builds a multi-objective performance prediction model based on the constructed molecular structure performance dataset; the pre-training module pre-trains the molecular generation model using the filtered molecular structure data; the reward function construction module constructs a hierarchical multi-objective reward function by introducing configurational constraints and performance indicators layer by layer, based on structural legality; and the optimization module uses reinforcement learning methods to quickly evaluate key indicators using the performance prediction model and optimizes the molecular generation model using the hierarchical multi-objective reward function.

[0007] Specifically, the dataset construction module is used for: To construct the relevant property data required for the prediction model, experimental test data and calculation results related to functional molecules at the lithium metal anode interface were extracted by searching existing public literature, journal articles and patent literature. By extracting and organizing the structural and performance data from the literature, and combining structural calculation and format unification conversion methods, the data is standardized to construct a structural performance mapping dataset. Construct a high-quality molecular structure dataset for pre-training.

[0008] The construction of the high-quality molecular structure dataset for pre-training includes: Small and medium-sized molecules with complete structures and elemental compositions that match the application scenarios of lithium metal battery interfaces are selected from public databases. Organic small molecules containing O, N, F, S, C, and H elements, with moderate molecular weight and diverse structures are given priority as the source of basic data. During the screening process, the RDKit toolkit is used to perform structural normalization and preliminary filtering of the original molecules to remove molecules that do not meet the design requirements. The rationality of the basic structure of the molecules is further limited by embedding 3D configuration, calculating linearity and complexity.

[0009] The multi-objective performance prediction model building module is specifically used for: The collected structural performance data is processed to ensure the consistency and reliability of the input data; the preprocessing operations include: handling missing values, removing duplicate structures, and detecting and cleaning outliers; Normalization or standardization methods are used to process numerical features to avoid training bias caused by different units of measurement.

[0010] The pre-training module is specifically used for: Select the filtered molecular structure data as training samples, standardize the molecular representation format, and remove entries with syntax errors, structural abnormalities, or atom types that do not meet the design requirements. We employ a deep generative model suitable for sequence modeling, and design the training based on the characteristics of chemical structure sequences. The training process aims to maximize the likelihood function of the molecular structure sequence or minimize the reconstruction error. We use cross-entropy loss function and optimizer to iteratively update parameters, and combine early stopping mechanism with validation set performance to dynamically adjust the number of training rounds. Structural perturbation enhancement and dropout regularization are introduced during training to ensure that the model maintains structural compliance and a certain degree of innovation when faced with different chemical substructures and skeleton combinations.

[0011] The deep generative models include, but are not limited to: recurrent neural networks, long short-term memory networks, variational autoencoders, and autoregressive structures based on Transformer.

[0012] The reward function construction module is specifically used for: In the early stages of training, the reward function prioritizes evaluating the structural rationality of molecules; Introducing configurational features as structural constraints enhances the syntheticity and engineering applicability of the generated molecules; After the basic structural conditions are met, performance-oriented items are gradually introduced, and rapid evaluation is carried out through the performance prediction model. Corresponding reward values ​​are assigned to guide the model to optimize the generation of results.

[0013] The optimization module is specifically used for: The pre-trained molecular generation model is used as the initial policy network, and a hierarchical multi-objective reward function is combined to evaluate each molecular structure generated by the molecular generation model. The evaluation results are fed back to the model in the form of numerical rewards to update its generation strategy; The reinforcement learning method uses an optimization method suitable for sequence generation tasks, and continuously adjusts the model parameters through iterative training to improve the probability of generating polymer structures. During the training process, the molecules generated in each round are first screened for legitimacy and structural rationality. Then, the key indicators are quickly evaluated based on the performance prediction model, and the comprehensive reward value is calculated according to the preset objective function.

[0014] This application constructs a multi-dimensional evaluation system and combines a design framework that links generative and predictive models to achieve synergistic optimization of molecular structure and target performance using large-scale molecular structure data and key performance indicators. This method can effectively guide the structural construction and performance regulation of self-assembled molecules, thereby realizing automated, efficient, and target-oriented reverse design of self-assembled molecular layers at lithium anode interfaces. This significantly improves screening efficiency and reduces development costs, demonstrating broad application prospects and promotional value. Specifically, This application employs a hierarchical multi-objective reward function design to achieve layer-by-layer constraints and comprehensive optimization of molecular structure rationality and multiple performance indicators, overcoming the limitations of traditional single-indicator screening methods. By deeply coupling the generative model and the multi-objective performance prediction model, and combining a closed-loop optimization mechanism of reinforcement learning, this application can automatically generate and optimize molecular structures starting from the target performance, avoiding dependence on existing molecular libraries, greatly expanding the explorable chemical space, and improving the innovation and efficiency of molecular design. Simultaneously, the combination of first-principles calculations and multi-source data support from a large-scale database ensures the accuracy and generalization ability of the performance prediction model, significantly improving the reliability and practicality of the overall scheme. In summary, this invention can effectively shorten the design cycle of functional molecules for lithium metal anode interfaces, reduce R&D costs, improve design efficiency and overall molecular performance, and significantly promote the development and application of high-performance lithium metal battery interface materials. Attached Figure Description

[0015] Figure 1 This is a hardware architecture diagram of the generative molecular inverse design system based on reinforcement learning of this invention; Figure 2 This is a schematic diagram illustrating the technical route of the generative molecular reverse design method based on reinforcement learning in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0017] See Figure 1 The diagram shown is a hardware architecture diagram of the reinforcement learning-based generative molecular reverse design system 10 of this invention. Please refer to it as well. Figure 2 The system includes: a dataset construction module 101, a multi-objective performance prediction model building module 102, a pre-training module 103, a reward function construction module 104, and an optimization module 105.

[0018] The dataset construction module 101 is used to construct and screen molecular structure performance datasets based on existing literature, databases, and experimental and computational results. That is: The dataset construction module 101 constructs a structured performance dataset for multi-objective prediction modeling and generative model pre-training, which mainly includes the following two types of data: On the one hand, the data required by the dataset construction module 101 to construct the performance prediction model mainly comes from existing literature, databases, and experimental and computational results, covering key structural and performance parameters of interface functional molecules, such as reduction potential, HOMO-LUMO energy level difference, Young's modulus, molecular adsorption energy, and ion diffusion coefficient. Relevant data is obtained through a combination of literature review, structural database extraction, and computational simulation, and the collected data is standardized. On the other hand, to build a high-quality structural dataset for pre-trained generative models, a set of interface functional molecules with reasonable structure, matching element composition, and certain chemical diversity and representativeness is selected, thereby improving the generalization ability of model training and the structural legitimacy of generated molecules.

[0019] In addition, for performance parameters that are missing or not yet published in some literature, the dataset construction module 101 performs supplementary calculations using first-principles calculation methods, serving as an important source of reinforcement for the structural performance dataset.

[0020] Specifically: First, relevant property data required for constructing the prediction model are obtained by searching existing public literature, journal articles and patent literature, and extracting experimental test data and calculation results related to functional molecules at the lithium metal anode interface, including but not limited to performance indicators such as molecular reduction potential, HOMO-LUMO energy level difference, Young's modulus, molecular adsorption energy, and ion diffusion coefficient.

[0021] Secondly, molecular data containing SMILES structures were downloaded from open-source molecular databases such as PubChem, Materials Project, and QM9, and their basic physicochemical properties were automatically extracted using a program. For performance parameters not publicly available in literature or databases, first-principles calculations, such as density functional theory, were employed. Quantum chemistry software, including VASP and Gaussian, along with common functionals such as B3LYP and PBE, were used to optimize the geometric structure, calculate charge distribution, and perform energy analysis of the target molecules under a reasonable basis set to obtain key parameters such as reduction potential and adsorption energy. After data acquisition, a standardized format, such as CSV, was used to standardize the structure, including normalizing atomic order and removing enantiomers. Feature extraction algorithms, such as RDKit, were then used to convert the molecular structure into a vector form acceptable to the model, providing stable and reliable input data for subsequent performance prediction model training. By extracting and organizing structure and performance data from the literature, and combining structural calculations with standardized format conversion methods, the data was standardized to construct a structure-performance mapping dataset.

[0022] Next, to enhance the diversity, legitimacy, and adaptability of the generative model during structure generation, the dataset construction module 101 also constructs a high-quality molecular structure dataset for pre-training. The goal of this dataset is to cover a wide range of structure types, elemental compositions, and molecular configuration features, ensuring that the generative model can fully learn representative structural syntax and molecular construction rules. Specifically, this includes: screening small and medium-sized molecules with complete structures and elemental compositions consistent with lithium metal battery interface application scenarios from public databases such as PubChem, prioritizing organic small molecules containing elements such as O, N, F, S, C, and H, with moderate molecular weights and diverse structures as the basic data source. During the screening process, the RDKit toolkit is used to perform structural normalization and preliminary filtering of the original molecules, eliminating molecules containing unstable ring systems, heavy metal elements, extreme charge structures, etc., that do not meet the design requirements; and further limiting the rationality of the basic molecular structure by embedding 3D configurations, calculating linearity, and complexity. To improve the model's adaptability to lithium anode interface scenarios, this embodiment introduces some existing functional molecular structures from target application scenarios as templates, enhancing the model's ability to memorize key functional groups and skeletal features. The processed molecular data are uniformly converted into structural representation formats such as SMILES or SELFIES, which can be used for sequence modeling, and used as training input for the generative model. This enables the full learning of basic molecular structural construction patterns and the establishment of a foundation for modeling. This part of the dataset provides stable and reliable initial structural semantic support for the subsequent effective training and reinforcement learning optimization of the generative model.

[0023] The multi-objective performance prediction model building module 102 is used to build a multi-objective performance prediction model based on the constructed molecular structure performance dataset. That is: The multi-objective performance prediction model building module 102 preprocesses the collected structural performance data, including data cleaning, outlier removal, normalization, and feature selection, to ensure data quality meets modeling requirements. The preprocessed data covers the molecular structure, electronic, mechanical, and interface-related features, used to predict key performance indicators such as reduction potential. For different types of performance objectives, regression or classification models are used for modeling. Modeling methods include, but are not limited to, random forest (RF), gradient boosting decision tree (GBDT), support vector machine (SVM), and neural network (NN), with the specific model selection determined based on the distribution characteristics of the performance data and the needs of the prediction task. During model training, model performance is evaluated through cross-validation, and hyperparameters are tuned using grid search, Bayesian optimization, and other methods to obtain the optimal model structure and parameter combination. For parallel prediction of multiple performance indicators, a multi-task learning framework or multi-model collaborative strategy is employed to achieve rapid and accurate evaluation of the comprehensive performance of candidate molecules, providing effective feedback for optimizing the generated model.

[0024] Specifically: The multi-objective performance prediction model building module 102 constructs a structure-performance mapping model to quickly and accurately predict key performance indicators of functional molecules at the lithium metal anode interface, thereby providing data support for reward function calculation and structure selection in reinforcement learning. To ensure the model's generalization ability and prediction accuracy, the constructed dataset undergoes systematic processing and reasonable modeling, specifically including: First, the multi-objective performance prediction model building module 102 performs preprocessing operations on the collected structural performance data, including missing value processing, duplicate structure removal, outlier detection and cleaning, etc., to ensure the consistency and reliability of the input data. Secondly, the multi-objective performance prediction model building module 102 uses normalization or standardization methods to process numerical features, avoiding training bias caused by different dimensions. For molecular structure description, multi-dimensional molecular descriptors such as topological structure features, electronic structure indices, and physicochemical properties are extracted based on cheminformatics tools such as RDKit and Mordred to construct an input feature matrix for model training.

[0025] To address the prediction needs of different performance indicators, a multi-model combination approach is adopted to construct regression or classification models. For continuous targets such as reduction potential, regression algorithms such as Support Vector Regression (SVR), Random Forest Regression (RF), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting (XGBoost) are used; for categorical attributes, classification models such as Logistic Regression, Support Vector Machine (SVM), or Multilayer Perceptron (MLP) are employed. During model training, cross-validation is introduced to evaluate model stability, and metrics such as mean squared error, mean absolute error, and coefficient of determination are used to measure model performance. Hyperparameters are adjusted using grid search or Bayesian optimization methods to improve the model's prediction accuracy and generalization ability. In this embodiment, to simultaneously meet the requirements of multi-objective collaborative optimization, a multi-task learning architecture is adopted, using multiple performance indicators as joint training objectives. Based on a shared structural feature extractor, the prediction outputs of each performance indicator are optimized separately. This ensures the compactness of the model structure while enhancing the model's ability to learn the potential coupling relationships between performance indicators, effectively improving the prediction accuracy of the overall performance.

[0026] The final performance prediction model serves as a key module for constructing the reward function, evaluating the performance of various objectives of the generated molecule, and providing real-time feedback during the reinforcement learning optimization process to guide the molecular structure to evolve towards a better multi-objective direction.

[0027] The pre-training module 103 is used to pre-train a molecular generation model using screened molecular structure data. That is: The pre-training module 103 uses selected molecular structure data to pre-train the pre-trained molecular generation model, enabling the model to learn the basic structural rules and chemical grammar of molecules, thereby improving the effectiveness and diversity of generated molecules. The dataset used includes interface functional molecules with reasonable chemical structures and elemental compositions that conform to application scenarios, expressed in molecular sequence formats such as SMILES or SELFIES. During the pre-training phase, sequence-based generative models, such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or variational autoencoders (VAEs), are selected to perform unsupervised learning on a large amount of molecular structure data, enabling the model to generate complete molecules from noise or partial structures. The training process uses a cross-entropy loss function to evaluate generation accuracy and improves model convergence efficiency and generalization ability by adjusting model structure and optimizer parameters. After pre-training, the model is used to generate candidate molecules that conform to basic chemical structure rules, providing initial parameters and structural distribution basis for subsequent reinforcement learning optimization.

[0028] Specifically: To enhance the generative model's ability to learn chemical structure rules and construct a molecular generation framework with basic structural legitimacy and expressive diversity, the pre-training module 103 first pre-trains the generative model. This stage does not directly introduce performance-oriented approaches but instead focuses on structural grammar learning to provide a reasonable structural initialization and parameter foundation for subsequent reinforcement learning stages. Its core lies in enabling the generative model to possess the modeling capabilities for molecular skeleton construction, functional group arrangement, and sequence expression through training with large-scale, high-quality molecular structure data.

[0029] First, the pre-training module 103 selects filtered molecular structure data as training samples, uniformly adopting molecular representation formats such as SMILES or SELFIES, and removing entries with syntax errors, structural abnormalities, or atom types that do not meet design requirements. This dataset covers small and medium-sized organic molecules related to lithium metal battery interface applications, possessing good structural representativeness and compositional diversity.

[0030] Secondly, regarding the model architecture, a deep generative model suitable for sequence modeling is adopted, including but not limited to recurrent neural networks (RNNs), long short-term memory networks (LSTMs), variational autoencoders (VAEs), or autoregressive structures based on Transformers, and the training design is combined with the characteristics of chemical structure sequences. The training process aims to maximize the likelihood function of the molecular structure sequence or minimize the reconstruction error, and iteratively updates the parameters through the cross-entropy loss function and optimizer, while dynamically adjusting the number of training rounds in conjunction with an early stopping mechanism and validation set performance.

[0031] Finally, to improve the generalization ability and stability of the model, this embodiment introduces techniques such as structural perturbation enhancement and dropout regularization during training to ensure that the model maintains structural compliance and a certain degree of innovation when facing different chemical substructures and skeleton combinations.

[0032] Through the above, the pre-trained molecular generation model learns and masters the basic construction rules of molecular structures under unsupervised conditions, and can generate diverse and structurally reasonable molecular sequences, providing an effective structural initialization foundation and generation strategy framework for subsequent reinforcement learning optimization based on performance feedback.

[0033] The reward function construction module 104 is used to construct a hierarchical multi-objective reward function by introducing configuration constraints and performance indicators layer by layer, based on the premise of structural legality. That is: The reward function construction module 104 constructs a hierarchical multi-objective reward function to guide the generative model in gradually optimizing the structural features and target performance of molecules during reinforcement learning. This reward function first evaluates the basic legitimacy of the generated molecules, then sequentially introduces structural constraints such as structural linearity and molecular complexity, and provides negative feedback to structures that do not meet the conditions, thus filtering out molecules that do not meet the design requirements in advance. After satisfying the basic structural conditions, a comprehensive score is then calculated based on key performance indicators such as reduction potential, driving the model towards optimization towards structurally reasonable and high-performance directions. This effectively ensures the practicality of the generated molecules and the coordination of multi-objective performance, improving the accuracy and effectiveness of molecular design.

[0034] Specifically: The reward function construction module 104 employs a hierarchical multi-objective reward function mechanism to guide the generative model in generating structurally sound and high-performance interface functional molecules during reinforcement learning. This reward function, based on structural legality, introduces configurational constraints and performance indicators layer by layer, ensuring that the generation process, while meeting basic chemical rules, gradually approaches the comprehensive performance requirements of the target application. In the initial training phase, the reward function prioritizes evaluating the structural rationality of the molecule, including basic grammatical constraints such as bonding legality and elemental valence correctness. Subsequently, configurational features such as structural linearity, molecular complexity, and atomic composition are introduced as structural constraints to improve the synthesizability and engineering applicability of the generated molecules. After meeting the basic structural conditions, performance-oriented terms are gradually introduced, including but not limited to key indicators such as reduction potential, HOMO-LUMO energy level difference, and Young's modulus. These are rapidly evaluated through a performance prediction model, and corresponding reward values ​​are assigned to guide the model in optimizing the generated results. This reward mechanism sets the weights of each indicator according to the actual application, supports multi-objective collaborative optimization, effectively improves the practicality, target relevance, and structural diversity of molecules during reinforcement learning training, and provides a stable and efficient optimization direction for the automated design of lithium metal anode interface molecules.

[0035] The optimization module 105 is used to employ reinforcement learning methods, utilize a performance prediction model to quickly evaluate key indicators, and optimize and adjust the molecular generation model through a hierarchical multi-objective reward function. That is: The optimization module 105, based on a pre-trained molecular generation model, employs reinforcement learning to optimize and adjust the model through a hierarchical multi-objective reward function. Specifically, the pre-trained molecular generation model serves as the initial policy network, and the performance of each generated molecular structure is evaluated using the reward function, with reward signals fed back based on the evaluation results. The reinforcement learning process continuously adjusts the generation model parameters using methods such as policy gradient algorithms or proximal policy optimization (PPO), gradually biasing it towards generating molecules that satisfy structural constraints while also considering multi-objective performance. This achieves reverse molecular structure design starting from target performance, overcoming the limitations of traditional screening methods. It enables the active exploration and generation of self-assembling functional molecules at the lithium metal anode interface with excellent electrochemical and mechanical properties, while ensuring molecular legitimacy and reasonable structure, achieving efficient and automated multi-objective molecular reverse design.

[0036] Specifically: After completing the pre-training of the molecular generation model and constructing the hierarchical multi-objective reward function, the optimization module 105 employs a reinforcement learning strategy to optimize and adjust the pre-trained molecular generation model, achieving molecular structure design based on the reverse of target performance. This process, through policy optimization methods in reinforcement learning, enables the molecular generation model to gradually learn how to generate interface functional molecules that better meet performance requirements. Specifically, the optimization module 105 uses the pre-trained molecular generation model as the initial policy network and, combined with the hierarchical multi-objective reward function, evaluates each molecular structure generated by the molecular generation model. The evaluation results are fed back to the model in the form of numerical rewards to update its generation strategy. The reinforcement learning algorithm uses policy gradient, proximal policy optimization, or other optimization methods suitable for sequence generation tasks, continuously adjusting model parameters through iterative training to improve the probability of generating polymer structures. During training, each round of generated molecules is first screened for legality and structural rationality, then key indicators are quickly evaluated based on the performance prediction model, and a comprehensive reward value is calculated according to the preset objective function. The reward is used to update the generative model, strengthening its evolution towards "structural feasibility + excellent performance", thereby actively exploring and outputting novel interface molecules that meet multiple performance requirements in the structural space.

[0037] Through the above-mentioned reinforcement learning optimization process, this invention realizes a reverse design path starting from the goal orientation, effectively overcoming the problems of existing technologies that rely on existing structure libraries and cannot carry out structural innovation, improving the diversity, practicality and goal consistency of molecular generation, significantly enhancing design efficiency and accuracy, and enabling automated generation and optimization of molecular structures under multi-goal performance orientation.

[0038] This invention constructs a reinforcement learning optimization framework that couples generative and predictive models, combined with a multi-performance target-oriented optimization strategy. This framework guides structure generation from the perspective of target performance, breaking the dependence of existing methods on molecular libraries and guiding the automatic generation and screening of molecular structures. It enables proactive exploration of the structural space, overcoming the limitation of traditional high-throughput screening methods that "can only select, not create," and thus possessing structural innovation capabilities. Simultaneously, by constructing a multi-performance synergistic optimization design framework, it introduces multiple key performance indicators such as reduction potential, HOMO-LUMO energy level difference, Young's modulus, molecular adsorption energy, and ion diffusion coefficient to achieve comprehensive evaluation and target-driven optimization of molecular layer interface performance, thereby improving the accuracy and practicality of interface functional molecule design. This invention can significantly improve design efficiency, reduce experimental costs, and provide new ideas and technical paths for the efficient control of lithium metal anode interface structures, providing strong support for the development and application of lithium metal batteries.

[0039] Although the present invention has been described with reference to the present preferred embodiments, those skilled in the art should understand that the above preferred embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., 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 generative molecular inverse design system based on reinforcement learning, characterized in that, The system includes a dataset construction module, a multi-objective performance prediction model building module, a pre-training module, a reward function construction module, and an optimization module, among which: The dataset construction module is used to construct and screen molecular structure performance datasets based on existing literature, databases, and experimental and computational results. The multi-objective performance prediction model building module is used to build a multi-objective performance prediction model based on the constructed molecular structure performance dataset; The pre-training module is used to pre-train a molecular generation model using screened molecular structure data; The reward function construction module is used to construct a hierarchical multi-objective reward function by introducing configuration constraints and performance indicators layer by layer, based on the premise of structural legality. The optimization module is used to employ reinforcement learning methods, utilize performance prediction models to quickly evaluate key indicators, and optimize and adjust the molecular generation model through a hierarchical multi-objective reward function.

2. The system as described in claim 1, characterized in that, The dataset construction module is specifically used for: To construct the relevant property data required for the prediction model, experimental test data and calculation results related to functional molecules at the lithium metal anode interface were extracted by searching existing public literature, journal articles and patent literature. By extracting and organizing the structural and performance data from the literature, and combining structural calculation and format unification conversion methods, the data is standardized to construct a structural performance mapping dataset. Construct a high-quality molecular structure dataset for pre-training.

3. The system as described in claim 2, characterized in that, The construction of the high-quality molecular structure dataset for pre-training includes: Small and medium-sized molecules with complete structures and elemental compositions that match the application scenarios of lithium metal battery interfaces are selected from public databases. Organic small molecules containing O, N, F, S, C, and H elements, with moderate molecular weight and diverse structures are given priority as the source of basic data. During the screening process, the RDKit toolkit is used to perform structural normalization and preliminary filtering of the original molecules to remove molecules that do not meet the design requirements. The rationality of the basic structure of the molecules is further limited by embedding 3D configuration, calculating linearity and complexity.

4. The system as described in claim 3, characterized in that, The multi-objective performance prediction model building module is specifically used for: The collected structural performance data is processed to ensure the consistency and reliability of the input data; the preprocessing operations include: handling missing values, removing duplicate structures, and detecting and cleaning outliers; Normalization or standardization methods are used to process numerical features to avoid training bias caused by different units of measurement.

5. The system as described in claim 4, characterized in that, The pre-training module is specifically used for: Select the filtered molecular structure data as training samples, standardize the molecular representation format, and remove entries with syntax errors, structural abnormalities, or atom types that do not meet the design requirements. We employ a deep generative model suitable for sequence modeling, and design the training based on the characteristics of chemical structure sequences. The training process aims to maximize the likelihood function of the molecular structure sequence or minimize the reconstruction error. We use cross-entropy loss function and optimizer to iteratively update parameters, and combine early stopping mechanism with validation set performance to dynamically adjust the number of training rounds. Structural perturbation enhancement and dropout regularization are introduced during training to ensure that the model maintains structural compliance and a certain degree of innovation when faced with different chemical substructures and skeleton combinations.

6. The system as described in claim 5, characterized in that, The deep generative models include, but are not limited to: recurrent neural networks, long short-term memory networks, variational autoencoders, and autoregressive structures based on Transformer.

7. The system as described in claim 6, characterized in that, The reward function construction module is specifically used for: In the early stages of training, the reward function prioritizes evaluating the structural rationality of molecules; Introducing configurational features as structural constraints enhances the syntheticity and engineering applicability of the generated molecules; After the basic structural conditions are met, performance-oriented items are gradually introduced, and rapid evaluation is carried out through the performance prediction model. Corresponding reward values ​​are assigned to guide the model to optimize the generation of results.

8. The system as described in claim 7, characterized in that, The optimization module is specifically used for: The pre-trained molecular generation model is used as the initial policy network, and a hierarchical multi-objective reward function is combined to evaluate each molecular structure generated by the molecular generation model. The evaluation results are fed back to the model in the form of numerical rewards to update its generation strategy; The reinforcement learning method uses an optimization method suitable for sequence generation tasks, and continuously adjusts the model parameters through iterative training to improve the probability of generating polymer structures. During the training process, the molecules generated in each round are first screened for legitimacy and structural rationality. Then, the key indicators are quickly evaluated based on the performance prediction model, and the comprehensive reward value is calculated according to the preset objective function.

Images