Chemical intelligent design system for high-dimensional parameter global optimization

Abstract

The application discloses a high-dimensional parameter global optimization chemical intelligent design system and relates to the technical field of chemical design. The high-dimensional parameter global optimization chemical intelligent design system defines the core architecture of the high-dimensional parameter global optimization chemical intelligent design system, integrates four core modules of chemical and material database construction, chemical descriptor extraction, high-dimensional structure-activity relationship prediction model training, intelligent design and reverse deduction, and forms a full-process closed-loop technical scheme of data-feature-model-design. The core benefit is that the traditional chemical research and development is broken through the limitation of scattered data, fragmented algorithm and experience-dependent design, and for the first time, the integrated design from data automatic collection to intelligent scheme output is realized, a systematic solution for high-dimensional parameter global optimization is provided, the protection foundation of the whole technical scheme is laid, and the intelligent level and research and development efficiency of chemical research and development are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of chemical design technology, specifically to a high-dimensional parameter global optimization intelligent chemical design system. Background Technology

[0002] Intelligent chemical design under high-dimensional parameter global optimization relies on high-quality structured data support, high-dimensional correlation system global optimization algorithms, predictive models that deeply integrate theoretical and experimental data, and the ability to accurately transform performance requirements into design solutions. However, in actual chemical research and development scenarios, factors such as the complexity of research objects, variable operating conditions, and dispersed data sources objectively exist, and existing technical solutions have revealed the following shortcomings: First, existing technologies lack a systematic approach to constructing high-quality chemical and materials databases. Data is largely scattered across unstructured carriers such as literature and patents, exhibiting chaotic formats and poor correlation, failing to form a standardized, structured data accumulation, and thus unable to provide a unified and reliable data source for high-dimensional parameter optimization. Second, key steps in existing technologies, such as chemical descriptor extraction and structure-activity relationship model training, are independent and lack collaborative design. A global optimization algorithm adapted to high-dimensional relational systems has not been developed, making it difficult for models to uncover deep intrinsic relationships between chemical objects, and high-dimensional parameter optimization is prone to getting trapped in local optima. Finally, existing technologies struggle to accurately translate the performance requirements of target chemicals into feasible technical constraints, lacking an integrated reverse engineering mechanism from performance requirements to formulations, process parameters, and synthetic routes. The design process heavily relies on the experience of researchers, easily leading to low feasibility and long development cycles.

[0003] Therefore, there is an urgent need for an intelligent design system with systematic database construction, collaborative algorithm model training, and precise reverse inference capabilities to solve the above-mentioned technical bottlenecks and achieve global optimization of high-dimensional parameters and efficient and accurate design of chemicals. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a chemical intelligent design system with high-dimensional parameter global optimization, which solves the problems of difficulty in high-dimensional parameter global optimization, fragmented and inconsistent data, disconnect between algorithms and models, low R&D efficiency due to reliance on experience in design, and insufficient accuracy of solutions in traditional chemical R&D.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a high-dimensional parameter global optimization intelligent chemical design system, comprising: The Chemistry and Materials Database Construction Module is used to extract structured data from chemical research and development data and expand the database based on the structured data to form a Chemistry and Materials Database. The chemical descriptor extraction module is used to generate pre-trained molecular descriptors based on spectroscopic pre-training by fusing quantum chemical Raman spectroscopy theory and experimental datasets using a spectroscopic inversion intelligent algorithm. The high-dimensional structure-activity relationship prediction model training module is used to construct a spectroscopic structure-activity relationship prediction model by training iteratively using machine learning techniques, based on the global optimization algorithm of high-dimensional related systems, combined with molecular descriptors and structured data. The intelligent design and reverse engineering module is used to retrieve the chemical and materials database by calling the spectroscopic structure-activity relationship prediction model based on the performance requirements of the input target chemical, and output candidate formulations, process conditions and synthesis route schemes.

[0006] The present invention has the following beneficial effects: This intelligent chemical design system with high-dimensional parameter global optimization clarifies its core architecture. By integrating four core modules—chemical and materials database construction, chemical descriptor extraction, high-dimensional structure-activity relationship prediction model training, and intelligent design and reverse engineering—it forms a closed-loop technical solution encompassing data, features, models, and design. Its core advantage lies in overcoming the limitations of traditional chemical R&D, such as fragmented data, disjointed algorithms, and reliance on experience. For the first time, it achieves integrated design from automated data acquisition to intelligent solution output, providing a systematic solution for high-dimensional parameter global optimization. This not only lays a solid foundation for the entire technical solution but also efficiently coordinates the technical collaboration across various stages, significantly improving the intelligence level and efficiency of chemical R&D.

[0007] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0008] Figure 1 This is a flowchart of the chemical intelligent design system method for high-dimensional parameter global optimization according to the present invention. Detailed Implementation

[0009] Please see Figure 1 This invention provides a technical solution: a high-dimensional parameter global optimization intelligent chemical design system, comprising: The chemical and materials database construction module is used to extract structured data from chemical research and development data (related literature and patent texts). The structured data includes entities, functions, properties, actions and dosages, and process parameters in chemical experiments. Based on the structured data, a chemical and materials database covering energy catalysts, battery electrolytes, battery positive and negative electrodes, and optical thin films is formed. The database contains hundreds of millions of related data and no less than 11 million reaction path data.

[0010] The process of extracting structured data from chemical research and development data is as follows: Literature and patent texts related to chemical research and development are used as data for chemical research and development.

[0011] The chemical research and development data is analyzed using natural language processing technology to parse the sentence structure, break down the core semantic units of subject, verb, and object, locate the target paragraphs describing the chemical experiments, and output the parsed set of semantic units and the corresponding target paragraphs of the chemical experiments.

[0012] For the set of semantic units and the target paragraph of the chemical experiment, information categories are labeled, and the categories of information to be extracted are defined as entities, functions, properties, actions, dosages and process parameters. The set of semantic units labeled with information categories is then output.

[0013] Based on the set of semantic units, target information is extracted according to corresponding categories: entities in chemical experiments (including compounds, crystals, catalysts, solvents, and reagents), functions in chemical experiments (including catalytic activity, electrochemical performance, optical transmittance, and stability), properties in chemical experiments (including physicochemical properties, molecular spectral characteristics, and thermodynamic parameters), actions in chemical experiments (including synthesis, stirring, centrifugation, heating, cooling, characterization, and testing), quantitative dosage data containing mass units (mg / g / kg), volume units (mL), and molar units (mol), and process parameters in chemical experiments (including temperature, pressure, reaction time, stirring speed, and reaction pH), and a preliminary set of extracted multi-category information is output.

[0014] The multi-category information set is standardized and transformed to unify the data recording format and measurement standards, and output a standardized information set with a unified format.

[0015] Based on the standardized information set, information correlation is verified, and duplicate and contradictory data are eliminated through logical consistency judgment to form structured data.

[0016] By introducing natural language processing technology to automate the transformation from literature and patent texts into standardized structured data, the core benefit lies in solving the technical pain points of traditional chemical R&D data extraction, which relies on manual labor, is inefficient, suffers from inconsistent data formats, and is riddled with redundancy and contradictions. This process, through step-by-step operations such as sentence parsing, category labeling, standardization transformation, and correlation verification, ensures the accuracy, completeness, and standardization of the extracted data. This provides high-quality, standardized foundational data for subsequent database construction, avoids model training bias caused by data quality issues, and guarantees the stable operation of the entire system.

[0017] The process of expanding a chemistry and materials database based on structured data is as follows: The database is pre-defined to cover the application areas of energy catalysts, battery electrolytes, battery positive and negative electrodes, and optical thin films.

[0018] Collect specialized data (specialized experimental data, supplementary data from published literature, and patent technology data) within the application field, perform format adaptation processing on the collected specialized data to ensure that the data recording format and measurement standards of the specialized data are consistent with those of the structured data, and output a comprehensive data set with a unified format.

[0019] The comprehensive dataset is filtered to remove invalid data that does not meet the domain relevance requirements. For missing data items, a data interpolation strategy based on the same category is used to complete the data. The filtered and completed preprocessed dataset is then output.

[0020] Based on the preprocessed data set, a reaction path dataset is extracted, including reactants, intermediate products, products and corresponding reaction conditions. The extracted reaction path information is logically verified and integrated to form a standardized reaction path dataset, which contains no less than 11 million valid reaction path data.

[0021] The preprocessed data set is linked and integrated with the standardized reaction path dataset to establish a multi-dimensional data association mapping relationship between entity, function, property, process parameter, and reaction path, forming an integrated data set containing associated data.

[0022] A database storage framework is constructed based on the integrated data set. Data classification indexes and retrieval interfaces are set up. The integrated data set is imported into the storage framework according to the association mapping relationship, the initial database construction is completed, and the initial chemistry and materials database is output.

[0023] The initial chemistry and materials database is subjected to data integrity and accuracy checks. The consistency of data association logic is verified by sampling. Based on the verification results, the data index and storage structure are optimized to form a chemistry and materials database.

[0024] By focusing on specific application areas, integrating specialized data, and establishing data association mappings, its core benefit lies in solving the problems of traditional databases, such as narrow data coverage, weak domain specificity, low data correlation, and poor retrieval efficiency. This process, through format adaptation, data filtering, association integration, and validation optimization, constructs a specialized database that is domain-focused, standardized, closely linked, and easy to retrieve. This not only provides a rich and accurate data source for model training but also meets the high-efficiency requirements of subsequent intelligent retrieval, ensuring that highly relevant data can be quickly matched when searching for candidate solutions, thereby improving the relevance and reliability of the design solutions.

[0025] The chemical descriptor extraction module is used to communicate with the high-quality chemical and materials database construction unit, configure the spectroscopic inversion intelligent algorithm, and generate spectroscopically pre-trained molecular descriptors by fusing quantum chemical Raman spectroscopy theory and experimental datasets according to the spectroscopic inversion intelligent algorithm. The chemical descriptor extraction unit has a built-in automated chemical descriptor extraction program. The molecular descriptors specifically include bond length, bond angle, d orbital electrons, Coulomb matrix, point charge, work function, electric dipole moment and its included angle, and charge transfer characteristic parameters.

[0026] The process is as follows: To obtain the theoretical system of quantum chemical Raman spectroscopy and the Raman spectroscopy experimental dataset related to chemical research and development, clarify the correlation principle between molecular structure and spectral response included in the theoretical system, verify the source validity of the experimental dataset, and output a complete theoretical framework of quantum chemical Raman spectroscopy and the verified Raman spectroscopy experimental dataset.

[0027] The verified Raman spectroscopy experimental dataset is preprocessed by performing data denoising, baseline correction, spectral peak calibration, and data standardization operations in sequence. This ensures that the processed data format matches the computational parameter dimensions of the quantum chemical Raman spectroscopy theoretical framework, resulting in a standardized Raman spectroscopy experimental dataset.

[0028] Based on the aforementioned quantum chemical Raman spectroscopy theoretical framework, we explore the intrinsic correlation between molecular microstructure (bond length, bond angle, d orbital electron distribution, Coulomb matrix) and Raman spectral characteristics (peak position, peak intensity, half-maximum width), establish a structure-spectrum correspondence mapping table, and output a molecular structure-Raman spectrum correlation mapping table.

[0029] The standardized Raman spectroscopy experimental dataset and the molecular structure-Raman spectroscopy correlation mapping table are used as inputs. A spectroscopic-structure inversion model is constructed through neural network training. The model learns the nonlinear mapping relationship from spectral features to molecular microstructure and outputs a spectroscopic-structure inversion model that has been trained and converged.

[0030] A set of typical molecular samples from the target fields of chemical research and development (energy catalysts, battery electrolytes, battery positive and negative electrodes, and optical thin films) is selected. The Raman spectral data of the typical molecular sample set is input into the trained and converged spectroscopic-structure inversion model. The model inference outputs a set of molecular microstructure characteristic parameters.

[0031] The set of molecular microstructure feature parameters is subjected to feature screening and dimensional optimization. Bond length, bond angle, d orbital electrons, Coulomb matrix, point charge, work function, electric dipole moment and its included angle, and charge transfer related parameters are integrated to form a feature vector of unified dimension, and an initial set of molecular descriptors is output.

[0032] The initial set of molecular descriptors is correlated with standard spectral data of known molecular structures. The matching degree between the descriptors and the spectral responses is calculated. The training parameters of the spectroscopic inversion intelligent algorithm are adjusted according to the verification results. The feature representation ability of the molecular descriptors is iteratively optimized, and the spectroscopically pre-trained molecular descriptors are output.

[0033] This paper elucidates the process of generating molecular descriptors pre-trained by spectrometry. It integrates a spectroscopic inversion intelligent algorithm, quantum chemical Raman spectroscopy theory, and experimental datasets. Its core advantage lies in overcoming the limitations of traditional molecular descriptors, such as insufficient representational ability, weak correlation with spectral information, and difficulty in reflecting the relationship between molecular microstructure and properties. Through spectral data preprocessing, inversion model construction, descriptor optimization, and correlation verification, the generated molecular descriptors can accurately capture the intrinsic correlation between molecular microstructure and spectral response, possessing stronger feature representation capabilities. This provides high-quality feature input for subsequent high-dimensional structure-activity relationship prediction models, fundamentally improving the model's prediction accuracy and generalization ability.

[0034] The process of constructing a spectroscopic-structure inversion model through neural network training, learning the nonlinear mapping relationship from spectral features to molecular microstructure, and outputting the trained and converged spectroscopic-structure inversion model is as follows: The spectral feature vector consisting of peak position, peak intensity, and half-maximum width is extracted from the standardized Raman spectroscopy experimental dataset. The corresponding molecular microstructure parameters such as bond length, bond angle, d orbital electron distribution, and Coulomb matrix are extracted from the association mapping table as labels. The training set and validation set are divided in a 7:3 ratio, and the training dataset, validation dataset, and corresponding label set are output.

[0035] An initial architecture for a spectral-guided graph neural network (GNN) is constructed, comprising an input layer, a spectroscopic feature extraction layer, a structure mapping layer, and an output layer. The input layer receives the spectral feature vectors and converts them into a tensor format that the model can process. The spectroscopic feature extraction layer consists of 3-5 graph convolutional layers and 2-3 pooling layers. The graph convolutional layers calculate the local correlation information of the spectral features using an adjacency matrix, while the pooling layers reduce the dimensionality of the features and preserve key information, jointly extracting higher-order spectral features. The structure mapping layer consists of 2-4 fully connected layers, used to establish a nonlinear mapping channel from higher-order spectral features to molecular microstructure parameters. The output layer uses a linear activation function to output the predicted values ​​of molecular microstructure parameters, thus displaying the initial architecture of the graph neural network (GNN).

[0036] The initial architecture of the graph neural network is configured with the following training parameters: the loss function is set to mean squared error loss function, the Adam optimizer is configured, the learning rate is set to 0.001-0.01, the maximum number of iterations is set to 1000-5000, the batch size is set to 32-128, and the early stopping mechanism trigger threshold is set to no decrease in loss on the validation set for 50-100 consecutive iterations. The configured spectroscopic-structure inversion model to be trained is then output.

[0037] The training dataset and corresponding label set are input into the spectroscopic-structure inversion model to be trained, and iterative training is initiated: each batch of training data is converted into tensors by the input layer and then fed into the spectroscopic feature extraction layer. The correlation weights between spectral features are calculated through the graph convolutional layer to generate a feature matrix. After the pooling layer filters key features, the higher-order spectral features are output. The higher-order spectral features are fed into the structure mapping layer, and nonlinear transformation is performed through the fully connected layer to obtain the predicted values ​​of molecular microstructure parameters. The loss value of the predicted value and the corresponding label set is calculated. The weight parameters of each layer are updated backward along the network layers through the backpropagation algorithm to complete the single-batch training, and the intermediate model after each iteration is output.

[0038] After each set number of iterations (100-200), the validation dataset is input into the currently updated intermediate model. The prediction accuracy and root mean square error (RMSE) of the predicted values ​​and the validation label set are calculated. It is then determined whether the convergence criteria are met (when the prediction accuracy is ≥90% and RMSE ≤0.05, the model training is considered converged, and the converged spectroscopic-structural inversion model is output). If the convergence condition is not met, the learning rate is reduced by 0.8 times, and iterative training continues until the convergence criteria are met, and the converged spectroscopic-structural inversion model is output.

[0039] Employing a spectral-guided graph neural network architecture, through scientific dataset partitioning, network parameter optimization, and iterative validation, its core advantage lies in solving the technical challenges of modeling the nonlinear mapping between spectral features and molecular microstructures, slow model training convergence, and low prediction accuracy. This training process, through rational design of network layers, configuration of training parameters, and setting of convergence criteria, ensures that the inversion model can stably learn the mapping relationship from spectral features to molecular structures. The trained model exhibits good convergence and high prediction accuracy, enabling efficient inversion from spectral information to molecular microstructures, providing core technical support for the accurate generation of molecular descriptors.

[0040] The process of verifying the correlation between the initial set of molecular descriptors and standard spectral data of known molecular structures, and calculating the matching degree between the descriptors and the spectral responses, is as follows: Obtain an initial set of molecular descriptors and a standard spectral dataset of known molecular structures. The standard spectral dataset contains the standard Raman spectrum peak positions, peak intensities, full width at half maximum (FWHM) of the corresponding molecules, as well as the corresponding molecular structure reference parameters.

[0041] The initial molecular descriptor set and the standard spectral dataset are normalized and standardized (the initial molecular descriptor set is dimensionally normalized to eliminate the dimensional differences of different feature parameters, and the standard spectral dataset is standardized to unify the numerical range and accuracy of the spectral data, so that the data format and numerical scale of the two are compatible), and the normalized molecular descriptor set and the standardized standard spectral dataset are output.

[0042] Based on the standardized standard spectral dataset and its corresponding molecular structure benchmark parameters, a descriptor-spectral response correlation evaluation system was established. Pearson correlation coefficient, root mean square error (RMSE), and peak matching rate were selected as quantification indicators for matching, and the calculation logic of each indicator was clarified: the Pearson correlation coefficient characterizes the linear correlation strength between descriptor features and spectral response; RMSE measures the deviation between the descriptor-derived spectral data and the standard spectral data; and the peak matching rate is used to statistically determine peak position deviations ≤ 0.5 cm. -1 Furthermore, the proportion of spectral peaks with peak intensity deviation ≤ 5% is used to output the correlation evaluation system and index calculation logic.

[0043] The normalized set of molecular descriptors is input into the trained and converged spectroscopic-structure inversion model to derive the corresponding predicted Raman spectrum dataset, which includes predicted peak position, predicted peak intensity, and predicted full width at half maximum (FWHM).

[0044] Based on the descriptor-spectral response correlation evaluation system and the calculation logic of the matching quantification index, the predicted Raman spectrum dataset and the standardized standard spectrum dataset are matched one-to-one according to molecular structure. The Pearson correlation coefficient, RMSE and peak matching rate of each molecule are calculated respectively. The comprehensive matching degree value is calculated by weighted averaging (with weight ratios of 30%, 40% and 30% respectively). The set of single-molecule matching degree indexes and the comprehensive matching degree value are output.

[0045] Set the matching degree qualification thresholds: Pearson correlation coefficient ≥ 0.92, RMSE ≤ 0.08, spectral peak matching rate ≥ 85%, and comprehensive matching degree value ≥ 0.88. Compare the set of single-molecule matching degree indicators and the comprehensive matching degree value with the qualification thresholds to determine whether the correlation between the initial molecular descriptor and the spectral response meets the standards, and output the correlation verification results.

[0046] The method for verifying the correlation between molecular descriptors and spectral responses, through comprehensive evaluation using multiple quantitative indicators (Pearson correlation coefficient, RMSE, and peak matching rate), offers the core advantage of addressing the lack of scientific validation and unreliable characterization of traditional molecular descriptors. This validation process, through data standardization, the establishment of a correlation evaluation system, and threshold determination, objectively quantifies the degree of matching between descriptors and spectral responses. It promptly identifies descriptors with insufficient characterization capabilities and iteratively optimizes them, ensuring that the final output molecular descriptors possess high reliability and strong characterization capabilities. This avoids subsequent model training failures or prediction deviations due to descriptor quality defects, thus guaranteeing the technical stability of the entire system.

[0047] The high-dimensional structure-activity relationship prediction model training module is used to interconnect with the chemical and materials database construction unit and the chemical descriptor extraction unit, respectively. It has a built-in high-dimensional correlation system global optimization algorithm. Based on the high-dimensional correlation system global optimization algorithm, combined with molecular descriptors and structured data, it uses machine learning technology to perform training iterations to construct a spectroscopic structure-activity relationship prediction model that integrates theoretical and experimental data. The spectroscopic structure-activity relationship prediction model is used to discover the hidden intrinsic correlations between chemical objects.

[0048] The global optimization algorithm for the high-dimensional correlation system adopts an improved particle swarm optimization-gradient descent hybrid optimization strategy, setting the population size to 50-100, the inertia weight to 0.4-0.8, the learning factor to 1.5-2.0, and the gradient descent step size to 0.001-0.01, to simultaneously optimize the correlation weights and model parameters between high-dimensional features.

[0049] The specific strategies and parameters of the global optimization algorithm for high-dimensional correlation systems were clarified. An improved particle swarm optimization-gradient descent hybrid optimization strategy was adopted, the core advantage of which lies in solving the problems of traditional optimization algorithms being prone to getting trapped in local optima, slow convergence speed, and insufficient optimization accuracy in high-dimensional feature spaces. By reasonably setting parameters such as population size, inertia weight, and learning factor, the algorithm can achieve simultaneous optimization of correlation weights between high-dimensional features and model parameters, taking into account both global search capability and local optimization accuracy. This effectively improves the efficiency and accuracy of high-dimensional parameter optimization, providing key algorithmic support for the efficient training of structure-activity relationship prediction models and ensuring that the model can uncover the deep intrinsic correlations between chemical objects.

[0050] Based on the global optimization algorithm for high-dimensional correlation systems, combined with molecular descriptors and structured data, the process of constructing a spectroscopic structure-activity relationship prediction model through machine learning training and iteration is as follows: The molecular descriptors and structured data are validated for consistency, logically contradictory data are removed, and dimensional differences are eliminated through feature normalization. The training set and validation set are divided in an 8:2 ratio, and the preprocessed training dataset, validation dataset and corresponding labels (labels are the measured values ​​of chemical functions / properties) are output.

[0051] An initial spectral structure-activity relationship model architecture integrating theoretical and experimental data was constructed, and model training parameters were configured: Based on quantum chemical Raman spectroscopy theory, a feature fusion layer, a high-dimensional correlation hidden layer, and a prediction output layer were set up. The feature fusion layer is used to concatenate the preprocessed molecular descriptor features and structured data features to generate a high-dimensional fused feature vector. The high-dimensional correlation hidden layer contains 4-6 fully connected layers and 2 attention mechanism layers to explore the intrinsic correlation between high-dimensional features. The prediction output layer uses the Sigmoid activation function to map to the chemical function / property prediction space. The loss function is set as either the cross-entropy loss function (for classification tasks) or the mean squared error loss function (for regression tasks), and the AdamW optimizer is selected. The initial learning rate is 0.001-0.005, and a learning rate decay strategy of 50% every 200 iterations is set. At the same time, an early stopping mechanism is set (training stops if the loss on the validation set does not decrease after 100 consecutive iterations). The configured model to be trained is then output.

[0052] The training dataset is input into the model to be trained. A high-dimensional fused feature vector is generated through a feature fusion layer. The attention mechanism of the high-dimensional association hidden layer is used to strengthen the weights of key features. The global optimization algorithm of the high-dimensional association system is used to simultaneously optimize the weight parameters and feature association coefficients of each layer of the model. After every 50 iterations, the parameters are updated through the backpropagation algorithm. The validation dataset is input into the current model, and the validation metrics are calculated, namely the mean absolute error (MAE) and the coefficient of determination (R²) between the predicted values ​​of functions or properties and the labels.

[0053] The convergence thresholds are set as MAE ≤ 0.06 and R² ≥ 0.93. If the validation metrics of the current intermediate model meet the thresholds, the model training is considered converged, and a spectral structure-activity relationship prediction model integrating quantum chemical theory (quantum chemical Raman spectroscopy theory) and experimental data (structured data, molecular descriptors) is output. If the convergence thresholds are not met, the inertia weights and learning factors of the global optimization algorithm for the high-dimensional correlation system are adjusted, and iterative training continues until the convergence criteria are met.

[0054] By combining high-dimensional correlation system global optimization algorithms with machine learning techniques, its core advantage lies in solving the problems of insufficient fusion of theoretical and experimental data, weak high-dimensional data processing capabilities, and poor predictive stability in traditional structure-activity relationship models. This process, through data preprocessing, model architecture design, parameter configuration, and iterative training, constructs a model that deeply integrates quantum chemical theory and experimental data, efficiently processes high-dimensional feature data, and accurately uncovers the intrinsic correlations between molecular structure, process parameters, and chemical properties. The trained model exhibits high predictive accuracy and strong stability, providing a core decision-making tool for intelligent chemical design and significantly reducing the uncertainty of design relying on experience.

[0055] The intelligent design and reverse inference module is used to integrate the spectroscopic structure-activity relationship prediction model output by the high-dimensional structure-activity relationship prediction model training module, and is configured with a high-value chemical reverse inference and precise construction sub-unit. Based on the performance requirements of the input target chemical, it calls the spectroscopic structure-activity relationship prediction model to retrieve the chemical and materials database and outputs candidate formulations, process conditions and synthesis route schemes.

[0056] The process is as follows: The performance requirements include at least one of catalytic activity, electrochemical performance, optical transmittance, and stability. These performance requirements are converted into quantitative indicators (including indicator threshold ranges) to form a standardized performance requirement parameter set, which is then output.

[0057] The standardized performance requirement parameter set is input into the structure-property relationship prediction model. Through the built-in correlation mapping logic of the model, the performance quantification index is transformed into corresponding molecular structure feature constraints and process parameter screening thresholds, clarifying the model retrieval guidance and outputting structure-performance correlation constraints and process parameter screening thresholds.

[0058] Based on the structure-performance correlation constraints and process parameter screening thresholds, the retrieval interface of the chemistry and materials database is called to retrieve structured data (including entities, functions, properties, and process parameters) and reaction path data that match the constraints. Correlation data with a matching degree of ≥80% is extracted, and the initial matching dataset is output.

[0059] The initial matching dataset is mined using a spectroscopic structure-activity relationship prediction model. Based on the intrinsic correlation patterns of chemical objects learned by the model, entities (compounds, catalysts, solvents, reagents) that meet the structure-performance constraints are selected, along with application amount data (including mg / g / kg, mL, mol quantification units) and associated process parameters (temperature, pressure, reaction time, stirring speed, etc.), forming a candidate basic data set.

[0060] Based on the reaction pathway data in the aforementioned candidate basic data set, reactants, intermediates, products, and corresponding reaction conditions are integrated to generate multiple complete synthetic pathways. Combining the entities, dosages, and process parameters in the candidate basic data set, an initial candidate scheme set containing multiple sets of candidate formulations, process conditions, and synthetic pathways is formed.

[0061] The feasibility of the initial candidate scheme set is verified. The performance prediction value corresponding to each candidate scheme is calculated by the structure-property relationship prediction model. The predicted value is verified to be within the threshold range of the target performance quantification index. Schemes with performance prediction values ​​exceeding the threshold range or with logical contradictions (such as mismatch between process parameters and reaction path) are eliminated.

[0062] The comprehensive score is calculated based on the degree of fit between the performance prediction value and the target value (weight 60%) and the complexity of the synthesis path (weight 40%). The comprehensive scores are sorted in descending order, and the top 10-20 schemes are selected to form the final candidate scheme set.

[0063] The core advantage of this process, from quantifying performance requirements to generating candidate solutions, lies in solving the problems of traditional chemical design, such as the difficulty in translating performance requirements into technical constraints, weak targeting of candidate solutions, and long R&D cycles. This process automates and precisely generates candidate formulations, process conditions, and synthetic routes by transforming performance requirements into structured constraints and combining them with a spectroscopic structure-activity relationship prediction model to search the database. This avoids the waste of resources caused by blind experiments and significantly shortens the R&D cycle. Furthermore, the generated candidate solutions, based on rich database resources and accurate model predictions, possess high feasibility and relevance, providing a high-quality foundation for subsequent experimental verification.

[0064] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0066] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0069] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A chemical intelligent design system with high-dimensional parameter global optimization, characterized in that, include: The Chemistry and Materials Database Construction Module is used to extract structured data from chemical research and development data and expand the database based on the structured data to form a Chemistry and Materials Database. The chemical descriptor extraction module is used to generate pre-trained molecular descriptors based on the fusion of quantum chemical Raman spectroscopy theory and experimental datasets using a spectroscopic inversion intelligent algorithm. The high-dimensional structure-activity relationship prediction model training module is used to construct a spectroscopic structure-activity relationship prediction model by training iteratively using machine learning techniques, based on the global optimization algorithm of high-dimensional related systems, combined with molecular descriptors and structured data. The intelligent design and reverse engineering module is used to retrieve the chemical and materials database by calling the spectroscopic structure-activity relationship prediction model based on the performance requirements of the input target chemical, and output candidate formulations, process conditions and synthesis route schemes.

2. The high-dimensional parameter global optimization intelligent chemical design system according to claim 1, characterized in that, The process of extracting structured data from chemical research and development data is as follows: Use literature and patent texts related to chemical research and development as chemical research and development data; The chemical research and development data are analyzed using natural language processing technology to parse the sentence structure, and the parsed semantic unit set and the corresponding chemical experiment target paragraph are output. For the set of semantic units and the target paragraph of the chemical experiment, information category labeling is performed, and the set of semantic units labeled with information categories is output. Based on the set of semantic units, target information is extracted according to the corresponding category, and a preliminary set of extracted multi-category information is output. The multi-category information set is standardized and transformed to output a standardized information set with a unified format; Based on the standardized information set, information correlation is verified to form structured data.

3. The high-dimensional parameter global optimization intelligent chemical design system according to claim 1, characterized in that, The process of forming a chemical and materials database based on structured data expansion is as follows: Clearly define the application areas that the database is intended to cover; Collect specialized data within the application field, perform format adaptation processing on the collected specialized data to ensure that the data recording format and measurement standards of the specialized data are consistent with those of the structured data, and output a comprehensive data set with a unified format. The comprehensive dataset is filtered to output a preprocessed dataset. Based on the preprocessed data set, extract the reaction path dataset; The preprocessed dataset is then linked and integrated with the standardized reaction path dataset to form an integrated dataset. A database storage framework is constructed based on the integrated data set. Data classification indexes and retrieval interfaces are set up. The integrated data set is imported into the storage framework according to the association mapping relationship to complete the initial construction of the database and output the initial chemistry and materials database. The initial chemistry and materials database is subjected to data integrity and accuracy checks. The consistency of data association logic is verified by sampling. Based on the verification results, the data index and storage structure are optimized to form a chemistry and materials database.

4. The high-dimensional parameter global optimization intelligent chemical design system according to claim 1, characterized in that, The process of generating spectroscopically pre-trained molecular descriptors by fusing quantum chemical Raman spectroscopy theory and experimental datasets using a spectroscopic inversion intelligent algorithm is as follows: The verified Raman spectroscopy experimental dataset is preprocessed to output a standardized Raman spectroscopy experimental dataset. Based on the aforementioned quantum chemical Raman spectroscopy theoretical framework, a structure-spectral correspondence mapping table is established, and a molecular structure-Raman spectrum correlation mapping table is output. The standardized Raman spectroscopy experimental dataset and the molecular structure-Raman spectroscopy correlation mapping table are used as inputs. A spectroscopic-structure inversion model is constructed through neural network training. The nonlinear mapping relationship from spectral features to molecular microstructure is learned, and the trained and converged spectroscopic-structure inversion model is output. A typical molecular sample set is selected from the target field of chemical research and development. The Raman spectral data of the typical molecular sample set is input into the trained convergent spectroscopic-structure inversion model. The model inference outputs a set of molecular microstructure characteristic parameters. The set of molecular microstructure feature parameters is subjected to feature screening and dimensional optimization. The parameters are integrated to form a feature vector of a unified dimension, and an initial set of molecular descriptors is output. The initial set of molecular descriptors is correlated with standard spectral data of known molecular structures to verify their correlation. The matching degree between the descriptors and the spectral responses is calculated. The feature representation ability of the molecular descriptors is iteratively optimized, and molecular descriptors pre-trained by spectrometry are output.

5. The high-dimensional parameter global optimization intelligent chemical design system according to claim 4, characterized in that, The process of constructing a spectroscopic-structure inversion model through neural network training, learning the nonlinear mapping relationship from spectral features to molecular microstructure, and outputting a convergent spectroscopic-structure inversion model is as follows: Spectral feature vectors are extracted from the standardized Raman spectroscopy experimental dataset, and corresponding molecular microstructure parameters are extracted from the association mapping table as labels. The training set and the validation set are divided, and the training dataset, the validation dataset and the corresponding label set are output. Construct the initial architecture of a spectral-guided graph neural network; Configure the training parameters for the initial architecture of the graph neural network and output the configured spectroscopic-structural inversion model to be trained; Input the training dataset and corresponding label set into the spectroscopic-structure inversion model to be trained, start iterative training, and output the intermediate model after each iteration; After each set number of iterations of training, the validation dataset is input into the currently updated intermediate model. The prediction accuracy and root mean square error (RMSE) of the predicted values ​​and the validation label set are calculated. It is then determined whether the convergence criteria are met. If the convergence criteria are not met, iterative training continues until the convergence criteria are met, and the converged spectroscopic-structural inversion model is output.

6. The high-dimensional parameter global optimization intelligent chemical design system according to claim 4, characterized in that, The process of verifying the correlation between the initial set of molecular descriptors and standard spectral data of known molecular structures, and calculating the matching degree between the descriptors and the spectral responses, is as follows: Obtain the initial set of molecular descriptors and the standard spectral dataset of known molecular structures; The initial set of molecular descriptors and the standard spectral dataset are normalized and standardized to output the normalized set of molecular descriptors and the standardized standard spectral dataset. Based on the standardized standard spectral dataset and its corresponding molecular structure benchmark parameters, a descriptor-spectral response correlation evaluation system is established, and Pearson correlation coefficient, root mean square error, and spectral peak matching rate are selected as matching quantification indicators. The normalized set of molecular descriptors is input into the trained and converged spectroscopic-structure inversion model to derive the corresponding predicted Raman spectrum dataset, which includes predicted peak position, predicted peak intensity, and predicted full width at half maximum (FWHM). Based on the descriptor-spectral response correlation evaluation system and the calculation logic of matching quantification index, the Pearson correlation coefficient, RMSE and peak matching rate of each molecule are calculated respectively. The comprehensive matching degree value is calculated by weighted average, and the set of single-molecule matching degree indexes and the comprehensive matching degree value are output. The set of single-molecule matching degree indices and the comprehensive matching degree value are compared with the qualified threshold to determine whether the correlation between the initial molecular descriptor and the spectral response meets the standard, and the correlation verification result is output.

7. The high-dimensional parameter global optimization intelligent chemical design system according to claim 1, characterized in that, The global optimization algorithm for the high-dimensional correlation system adopts an improved particle swarm optimization-gradient descent hybrid optimization strategy, setting the population size to 50-100, the inertia weight to 0.4-0.8, the learning factor to 1.5-2.0, and the gradient descent step size to 0.001-0.01, to simultaneously optimize the correlation weights and model parameters between high-dimensional features.

8. The high-dimensional parameter global optimization intelligent chemical design system according to claim 7, characterized in that, Based on the global optimization algorithm for high-dimensional correlation systems, combined with molecular descriptors and structured data, the process of constructing a spectroscopic structure-activity relationship prediction model through machine learning training and iteration is as follows: The molecular descriptors and structured data are validated for consistency. The training set and validation set are divided proportionally, and the preprocessed training dataset, validation dataset and corresponding labels are output. Construct an initial spectral structure-activity relationship model architecture that integrates theoretical and experimental data, configure the model training parameters, and output the configured model to be trained; The training dataset is input into the model to be trained for training, and the validation dataset is input into the current model to calculate the validation metric. If the validation metrics of the current intermediate model meet the threshold, the model training is deemed to have converged, and the spectral structure-activity relationship prediction model is output; if the convergence threshold is not met, iterative training continues until the convergence criterion is reached.

9. The high-dimensional parameter global optimization intelligent chemical design system according to claim 1, characterized in that, The process of retrieving candidate formulations, process conditions, and synthetic route schemes based on the performance requirements of the input target chemical, using the structure-activity relationship prediction model to search the chemical and materials database, and outputting the structure-activity relationship prediction model is as follows: The performance requirements are converted into quantitative indicators to form a standardized performance requirement parameter set, and the standardized performance requirement parameter set is output. The standardized performance requirement parameter set is input into the structure-property relationship prediction model, and the performance quantification index is transformed into corresponding molecular structure feature constraints and process parameter screening thresholds. The structure-performance correlation constraints and process parameter screening thresholds are output. Based on the structure-performance correlation constraints and process parameter screening thresholds, the retrieval interface of the chemistry and materials database is called to retrieve structured data and reaction path data that match the constraints, and the initial matching dataset is output. The initial matching dataset is mined using a spectral structure-property relationship prediction model to filter out entities that meet structure-performance constraints, application quantity data, and associated process parameters, forming a candidate basic data set. Based on the reaction path data in the candidate basic data set, reactants, intermediates, products and corresponding reaction conditions are integrated to generate multiple complete synthetic pathways; By combining the entities, dosages, and process parameters in the candidate basic data set, an initial candidate scheme set containing multiple sets of candidate formulations, process conditions, and synthesis routes is formed. The feasibility of the initial candidate solution set is verified, and the final candidate solution set is output.

10. The high-dimensional parameter global optimization intelligent chemical design system according to claim 9, characterized in that, The process of performing feasibility verification on the initial candidate solution set and outputting the final candidate solution set is as follows: The performance prediction values ​​of each candidate scheme are calculated by the spectral structure-property relationship prediction model. The predicted values ​​are verified to be within the threshold range of the target performance quantification index. Schemes with performance prediction values ​​exceeding the threshold range and those with logical contradictions are eliminated. A comprehensive score is calculated based on the degree of fit between the performance prediction and the target value, as well as the complexity of the synthesis path. The comprehensive scores are then sorted in descending order to select the final set of candidate solutions.

Images