Intelligent characterization and analysis system and method for chemical experiment products based on deep learning

Abstract

The application discloses a chemical experiment product intelligent characterization analysis system and method based on deep learning, and belongs to the technical field of chemical informatics and artificial intelligence. The system comprises: a data acquisition and standardization preprocessing module, which is used for automatically acquiring original spectrum data of an analytical instrument and generating standardized spectrum data; a deep learning analysis engine, which is used for analyzing the standardized spectrum data to obtain an analysis result comprising at least a product structure; and a reaction path automatic tracking and knowledge graph construction module, which is used for automatically associating the analysis result with synthesis records in an upstream experiment execution system, inferring a conversion path and storing the conversion path in a reaction knowledge graph. The application realizes the automation and standardization of spectrum analysis, can automatically associate the analysis result with synthesis conditions and conversion paths to construct a knowledge graph, supports mixture analysis and complex compound identification, and realizes model iteration optimization through human-computer interaction feedback, thereby significantly improving the efficiency and accuracy of chemical experiment product characterization analysis.

Description

Technical Field

[0001] This invention relates to the fields of cheminformatics, analytical chemistry and artificial intelligence, and more specifically, to a deep learning-based intelligent characterization and analysis system and method for chemical experimental products. Background Technology

[0002] With the rapid development of automated chemical synthesis technology, experimental execution has become highly automated, but product characterization and analysis still heavily rely on manual labor. In the traditional process, after the automated platform collects samples, the analytical instruments obtain raw spectral data, which are then manually interpreted by chemists to determine the success of the reaction, the purity of the product, and to identify the compound structure. This model has significant drawbacks: First, manual analysis is time-consuming, becoming a bottleneck in the automated process and forcing automated equipment to wait, thus failing to fully utilize its high-throughput advantages; second, the analytical results are highly subjective, with different chemists potentially interpreting the same spectrum differently, affecting the reliability and consistency of experimental results; third, manual analysis results are mostly recorded in unstructured text, making it difficult to accurately correlate with upstream synthesis parameters, resulting in an extremely slow process of discovering structure-activity relationships and optimizing reaction conditions from massive amounts of data.

[0003] In existing technologies, some patents attempt to use machine learning methods for reaction prediction or mass spectrometry data analysis, but most are limited to specific reaction types or single spectral analysis scenarios, lacking the ability to fuse and process multi-source spectral data, and struggling to automate the entire process from spectral analysis to synthesis record association and reaction pathway reasoning. Furthermore, traditional spectral library retrieval methods can only match known compounds, failing to provide effective structural identification assistance for newly synthesized or unlisted molecules, and also having significant limitations when handling complex mixture samples. Therefore, there is an urgent need to develop an intelligent characterization and analysis system that can automate and standardize spectral analysis and seamlessly link with upstream experimental data, thus bridging the final stage of automated chemical research and development. Summary of the Invention

[0004] In view of the above-mentioned technical problems in related technologies, the present invention proposes a deep learning-based intelligent characterization and analysis system for chemical experimental products, which can overcome the above-mentioned shortcomings of the prior art.

[0005] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows: A deep learning-based intelligent characterization and analysis system for chemical experimental products; This deep learning-based intelligent characterization and analysis system for chemical experimental products includes: The data acquisition and standardization preprocessing module is used to automatically acquire the raw spectral data of the analytical instrument and preprocess the raw spectral data to generate standardized spectral data. A deep learning parsing engine is used to receive the standardized spectral data and parse the standardized spectral data based on a deep learning model to obtain parsing results that include at least the product structure. The automatic reaction path tracking and knowledge graph construction module is used to automatically associate the analysis results with the synthesis records in the upstream experimental execution system, infer the transformation path from reactants to products, and store the synthesis records, the analysis results, and the transformation path as knowledge units in the reaction knowledge graph.

[0006] Furthermore, the deep learning parsing engine includes: The intelligent chromatographic analysis subnetwork is used to automatically detect peaks, calculate integrals, and analyze peak purity in the chromatographic data of the standardized spectral data; A smart spectral analysis subnetwork is used to identify compound structures from the spectral data in the standardized spectral data.

[0007] Furthermore, the intelligent spectral analysis subnetwork adopts a multi-task learning framework, simultaneously training three tasks: compound structure identification, functional group recognition, and spectrum-structure consistency verification.

[0008] Furthermore, the automatic reaction path tracking and knowledge graph construction module includes: The data association submodule is used to associate the product structure, purity, and by-product information in the analysis results with the synthesis record; The transformation path reasoning submodule is used to reason about the transformation path from reactants to products based on the reactant and product structures. The knowledge graph construction submodule is used to store the synthesis record, the parsing result, and the transformation path as knowledge units in the reaction knowledge graph.

[0009] Furthermore, it also includes dedicated analytical networks for peptides or natural product compounds, which employ targeted molecular feature extraction and network structure design to improve the analytical accuracy of the corresponding categories of compounds.

[0010] In addition, this invention also discloses a method for intelligent characterization and analysis of chemical experimental products based on deep learning; This deep learning-based intelligent characterization and analysis method for chemical experimental products includes the following steps: Step 1, Multi-source spectral data acquisition and standardized preprocessing: Automatically connect to the data interface of the analysis instrument, acquire raw spectral data in real time, and preprocess the acquired data to generate standardized spectral data tensors; Step 2, Spectral decoupling and structure identification based on deep learning: Input the standardized spectral data into the deep learning parsing engine to predict the corresponding compound structure and output the candidate structural formula and its confidence level; Step 3, Automatic Reaction Path Association and Knowledge Graph Construction: The product structure information parsed in Step 2 is automatically associated with the synthesis records in the upstream experimental execution system. Based on the reactant and product structures, a reasonable transformation path from reactant to product is deduced, and "synthesis conditions - characterization results - transformation path" is stored as a knowledge unit in the reaction knowledge graph database.

[0011] Furthermore, it also includes: Before or during step 2, when a chromatographic peak is determined to be a mixture or a mixed spectrum is directly input, the spectrum deconvolution and mixture analysis steps are executed: the spectrum deconvolution module is called to decouple the mixed spectrum, separate the contribution spectra of each component, and then input them into the spectral analysis subnetwork for structure identification.

[0012] Furthermore, it also includes: Step 4: Visualization of Results and Human-Computer Interaction Confirmation and Feedback Optimization: Visualize the analysis results on the graphical interface and receive verification and correction operations from chemists. Use the corrected data as new training data and feed it back to the deep learning engine for iterative optimization.

[0013] Furthermore, in step 2, the deep learning parsing engine adopts an ensemble learning strategy, which processes the data through multiple different base learners and combines the outputs of each base learner to obtain the final structure prediction result. The multiple different base learners include a one-dimensional convolutional neural network for processing the original spectral sequence, a gradient boosting tree for processing artificially designed chemical feature vectors, and a graph neural network for converting the spectral data into a molecular fingerprint to mine molecular topological structure features.

[0014] Furthermore, in step 2, the deep learning parsing engine uses a large language model for parsing, specifically including: The standardized spectral data is converted into spectral embedding vectors using a spectral encoder. The spectral embedding vector is used as a soft cue input into a fine-tuned chemical domain large language model, which then generates structural analysis results. Furthermore, before generating results, the large language model automatically retrieves historical experimental records in the reaction knowledge graph that are similar to the current graph, and adds the retrieved similar cases as contextual examples to the prompt words.

[0015] The beneficial effects of this invention are as follows: By setting up a data acquisition and standardized preprocessing module to automatically acquire and process multi-source spectral data, and combining it with a deep learning analysis engine to jointly analyze chromatographic and spectroscopic data, the spectral analysis process is automated and standardized, significantly improving analysis efficiency and consistency, and avoiding the subjectivity and time-consuming problems of manual operation; by using an automatic reaction path tracking and knowledge graph construction module to automatically associate the analysis results with the synthesis record and construct a searchable knowledge graph, the entire chain of knowledge from experimental conditions to product structure to transformation path can be accumulated and reused, thus providing a data foundation for structure-activity relationship mining and reaction condition optimization; by setting up a dedicated analysis network and spectral deconvolution module, the accuracy of analysis for complex compounds and mixtures is effectively improved, and by combining human-computer interaction confirmation and feedback optimization mechanisms, continuous iteration and knowledge updates of the model are achieved. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0017] The intelligent characterization and analysis system for chemical experimental products based on deep learning according to an embodiment of the present invention includes: The data acquisition and standardization preprocessing module is used to automatically acquire the raw spectral data of the analytical instrument and preprocess the raw spectral data to generate standardized spectral data. A deep learning parsing engine is used to receive the standardized spectral data and parse the standardized spectral data based on a deep learning model to obtain parsing results that include at least the product structure. The automatic reaction path tracking and knowledge graph construction module is used to automatically associate the analysis results with the synthesis records in the upstream experimental execution system, infer the transformation path from reactants to products, and store the synthesis records, the analysis results, and the transformation path as knowledge units in the reaction knowledge graph.

[0018] According to an embodiment of the intelligent characterization and analysis system for chemical experimental products based on deep learning, in a specific implementation, the deep learning analysis engine includes: The intelligent chromatographic analysis subnetwork is used to automatically detect peaks, calculate integrals, and analyze peak purity in the chromatographic data of the standardized spectral data; A smart spectral analysis subnetwork is used to identify compound structures from the spectral data in the standardized spectral data.

[0019] According to an embodiment of the present invention, the intelligent characterization and analysis system for chemical experimental products based on deep learning, in a specific implementation, the intelligent spectral analysis sub-network adopts a multi-task learning framework to simultaneously train three tasks: compound structure identification, functional group recognition, and spectrum-structure consistency verification.

[0020] In a specific embodiment of the intelligent characterization and analysis system for chemical experimental products based on deep learning according to an embodiment of the present invention, the automatic reaction path tracking and knowledge graph construction module includes: The data association submodule is used to associate the product structure, purity, and by-product information in the analysis results with the synthesis record; The transformation path reasoning submodule is used to reason about the transformation path from reactants to products based on the reactant and product structures. The knowledge graph construction submodule is used to store the synthesis record, the parsing result, and the transformation path as knowledge units in the reaction knowledge graph.

[0021] The intelligent characterization and analysis system for chemical experimental products based on deep learning according to an embodiment of the present invention further includes, in a specific embodiment, a dedicated analysis network for peptides or natural product compounds. The dedicated analysis network employs targeted molecular feature extraction and network structure design to improve the analysis accuracy of the corresponding category of compounds.

[0022] Secondly, the intelligent characterization and analysis method for chemical experimental products based on deep learning according to embodiments of the present invention includes the following steps: Step 1, Multi-source spectral data acquisition and standardized preprocessing: Automatically connect to the data interface of the analysis instrument, acquire raw spectral data in real time, and preprocess the acquired data to generate standardized spectral data tensors; Step 2, Spectral decoupling and structure identification based on deep learning: Input the standardized spectral data into the deep learning parsing engine to predict the corresponding compound structure and output the candidate structural formula and its confidence level; Step 3, Automatic Reaction Path Association and Knowledge Graph Construction: The product structure information parsed in Step 2 is automatically associated with the synthesis records in the upstream experimental execution system. Based on the reactant and product structures, a reasonable transformation path from reactant to product is deduced, and "synthesis conditions - characterization results - transformation path" is stored as a knowledge unit in the reaction knowledge graph database.

[0023] The intelligent characterization and analysis method for chemical experimental products based on deep learning according to an embodiment of the present invention, in a specific embodiment, further includes: Before or during step 2, when a chromatographic peak is determined to be a mixture or a mixed spectrum is directly input, the spectrum deconvolution and mixture analysis steps are executed: the spectrum deconvolution module is called to decouple the mixed spectrum, separate the contribution spectra of each component, and then input them into the spectral analysis subnetwork for structure identification.

[0024] The intelligent characterization and analysis method for chemical experimental products based on deep learning according to an embodiment of the present invention, in a specific embodiment, further includes: Step 4: Visualization of Results and Human-Computer Interaction Confirmation and Feedback Optimization: Visualize the analysis results on the graphical interface and receive verification and correction operations from chemists. Use the corrected data as new training data and feed it back to the deep learning engine for iterative optimization.

[0025] According to an embodiment of the present invention, in a specific implementation of the intelligent characterization and analysis method for chemical experimental products based on deep learning, in step 2, the deep learning analysis engine adopts an ensemble learning strategy, processes the data through multiple different base learners, and combines the outputs of each base learner to obtain the final structure prediction result; the multiple different base learners include a one-dimensional convolutional neural network for processing the original spectral sequence, a gradient boosting tree for processing artificially designed chemical feature vectors, and a graph neural network for converting the spectrum into a molecular fingerprint to mine molecular topological structure features.

[0026] According to an embodiment of the present invention, the intelligent characterization and analysis method for chemical experimental products based on deep learning, in a specific implementation, step 2, in which the deep learning analysis engine uses a large language model for analysis, specifically includes: The standardized spectral data is converted into spectral embedding vectors using a spectral encoder. The spectral embedding vector is used as a soft cue input into a fine-tuned chemical domain large language model, which then generates structural analysis results. Furthermore, before generating results, the large language model automatically retrieves historical experimental records in the reaction knowledge graph that are similar to the current graph, and adds the retrieved similar cases as contextual examples to the prompt words.

[0027] To facilitate understanding of the above technical solutions of the present invention, the following embodiments will be used to describe the above technical solutions of the present invention in detail.

[0028] Example 1 This embodiment provides a deep learning-based intelligent characterization and analysis system for chemical experimental products. The system includes a data acquisition and standardization preprocessing module, a deep learning analysis engine, an automatic reaction path tracking and knowledge graph construction module, and a result visualization and human-computer interaction confirmation module.

[0029] The data acquisition and standardization preprocessing module comprises an interface layer and a standardization processing layer. The interface layer automatically acquires raw chromatographic and spectroscopic data through the data interfaces of analytical instruments such as HPLC, NMR, and LC-MS. The standardization processing layer performs noise reduction, baseline correction, chemical shift calibration, and peak alignment on the acquired raw data to generate standardized spectral data. Specifically, for NMR data, standardization methods are used to process the 1H and 13C spectra; for HPLC data, standardization methods are used to process the UV absorbance data; and for LC-MS data, standardization methods are used to process the ion current data in full scan mode.

[0030] The deep learning-based analysis engine comprises an intelligent chromatographic analysis subnetwork and an intelligent spectroscopic analysis subnetwork. The intelligent chromatographic analysis subnetwork is constructed using a one-dimensional convolutional neural network (1D-CNN), with standardized chromatographic data as input. This network contains five convolutional layers and two pooling layers, with a kernel size of 3×3 for the convolutional layers and a pool size of 2×2 for the pooling layers. After training, the network can automatically identify chromatographic peaks and perform integral calculations, while simultaneously determining whether a peak represents a single compound.

[0031] The intelligent spectral analysis subnetwork is constructed using a graph neural network (GNN) combined with a multi-head attention mechanism. The input consists of standardized multidimensional NMR spectra and high-resolution mass spectrometry data. Specifically, 1D¹H and ¹³C spectra utilize a GNN sub-graph network, 2DCOSY spectra employ an edge convolutional network, and HSQC spectra use a self-attention network. The network employs a multi-task learning framework, simultaneously training for compound structure identification, functional group recognition, and spectrum-structure consistency verification tasks. After training, the network can be used to directly predict the molecular structure of compounds.

[0032] The automatic reaction pathway tracking and knowledge graph construction module comprises three sub-modules: data association, transformation path reasoning, and knowledge graph construction. The data association module automatically associates the parsed product structure, purity, and byproduct information with the synthesis records in the upstream intelligent scheduling system. The transformation path reasoning, based on reactant and product structures, utilizes a built-in chemical reaction rule base to automatically reason and verify reasonable transformation paths from reactants to products. The knowledge graph construction automatically stores the "synthesis conditions-characterization results-transformation path" of each experiment as a knowledge unit within the reaction knowledge graph.

[0033] The results visualization and human-computer interaction confirmation module provides a graphical display and a human-computer collaboration interface. Identified peaks are highlighted on the original spectrum, and the predicted compound structures and their confidence levels are displayed. Chemists can verify and correct the model's identification results through the interface. The corrected data will be used as new high-quality training data and fed back to the deep learning engine for iterative optimization.

[0034] Example 2 This embodiment provides a deep learning-based intelligent characterization and analysis system for chemical experimental products. The system includes a data acquisition and standardization preprocessing module, a deep learning analysis engine, an automatic reaction path tracking and knowledge graph construction module, and a result visualization and human-computer interaction confirmation module.

[0035] The data acquisition and standardization preprocessing module comprises an interface layer and a standardization processing layer. The interface layer automatically acquires raw chromatographic and spectroscopic data through the data interfaces of analytical instruments such as HPLC, NMR, and LC-MS. The standardization processing layer performs noise reduction, baseline correction, chemical shift calibration, and peak alignment on the acquired raw data to generate standardized spectral data. Specifically, for 1H NMR data, standardization processing is used to process data under 500 MHz conditions; for 13C NMR data, standardization processing is used to process data under 125 MHz conditions; for HPLC data, standardization processing is used to process absorbance data at 210 nm wavelength; and for LC-MS data, standardization processing is used to process full-scan ion current data in the range of 200–1000 m / z.

[0036] The deep learning-based analysis engine comprises an intelligent chromatography analysis subnetwork and an intelligent spectroscopic analysis subnetwork. The intelligent chromatography analysis subnetwork is built using a Transformer model, with standardized chromatographic data as input. This network consists of six encoder layers and six decoder layers; the encoder layers use convolutional networks, and the decoder layers use self-attention networks. After training, the network can automatically identify chromatographic peaks and perform integral calculations, while simultaneously determining whether a peak represents a single compound.

[0037] The intelligent spectral analysis subnetwork is constructed using a graph neural network (GNN) combined with a multi-head attention mechanism. The input consists of standardized 2D ¹H-¹³C HSQC and 2D ¹H-¹³C HMBC spectra. The network employs an edge convolutional network to process the spatial information of the spectra and a self-attention network to process the temporal information. A multi-task learning framework is used, simultaneously training the network for compound structure identification, functional group recognition, and spectrum-structure consistency verification tasks. After training, the network can be used to directly predict the molecular structure of compounds.

[0038] The automatic reaction pathway tracking and knowledge graph construction module comprises three sub-modules: data association, transformation path reasoning, and knowledge graph construction. The data association module automatically associates the parsed product structure, purity, and byproduct information with the synthesis records in the upstream intelligent scheduling system. The transformation path reasoning, based on reactant and product structures, utilizes a built-in chemical reaction rule base to automatically reason and verify reasonable transformation paths from reactants to products. The knowledge graph construction automatically stores the "synthesis conditions-characterization results-transformation path" of each experiment as a knowledge unit within the reaction knowledge graph.

[0039] The results visualization and human-computer interaction confirmation module provides a graphical display and a human-computer collaboration interface. Identified peaks are highlighted on the original spectrum, and the predicted compound structures and their confidence levels are displayed. Chemists can verify and correct the model's identification results through the interface. The corrected data will be used as new high-quality training data and fed back to the deep learning engine for iterative optimization.

[0040] Example 3 This embodiment describes an intelligent characterization and analysis system for chemical experimental products employing feature engineering and ensemble learning architecture. In this embodiment, the core of the deep learning analysis engine lies in multi-feature fusion and ensemble learning strategies. Its features include: Feature engineering processing: The data acquisition and standardized preprocessing module not only performs conventional preprocessing but also extracts over 200 artificially designed chemical features, including: chemical shift distribution characteristics of NMR spectra, statistical characteristics of coupling constants, fragment ion mode characteristics of mass spectrometry, isotope distribution characteristics, and peak shape symmetry characteristics of chromatographic peaks. These features, together with the original spectral data, constitute multimodal input.

[0041] Ensemble Learning Architecture: The spectral analysis sub-network employs an ensemble learning strategy, simultaneously training three different base learners. Base learner 1 is a one-dimensional convolutional neural network (1D-CNN), directly processing the original spectral sequence to capture local waveform features; base learner 2 is a gradient boosting tree (XGBoost), processing manually designed feature vectors of over 200 dimensions; and base learner 3 is a graph neural network (GNN), transforming the spectrum into a molecular fingerprint to mine molecular topological structure features. The ensemble strategy uses a soft voting method, calculating the final structure prediction result based on a weighted average of the confidence scores of the three base learners.

[0042] Spectral deconvolution implementation: The nonnegative matrix factorization (NMF) algorithm is used to deconvolve the mixed spectrum to separate the contribution spectra of each component. This method is computationally efficient and suitable for online rapid analysis.

[0043] Training and Update Mechanism: The model employs offline batch training. After accumulating 1000 manually corrected data points, a complete model retraining is triggered, and the updated model is deployed during the next system maintenance.

[0044] Example 4 This embodiment describes an intelligent characterization and analysis system for chemical experimental products enhanced with a large language model and knowledge graph, which differs significantly from Embodiment 1 in its technical approach. In this embodiment, the core of the deep learning analysis engine lies in the collaborative reasoning between the large language model and an external knowledge base. Its key features are: Spectrogram-Text Multimodal Alignment: Standardized spectrogram data is converted into spectrogram embedding vectors using a pre-trained spectrogram encoder. Simultaneously, textual data corresponding to the spectrograms, such as experimental conditions and reactant information, is converted into text embedding vectors using a large language model encoder. Through contrastive learning, alignment of spectrogram embeddings and text embeddings in the multimodal semantic space is achieved.

[0045] Large Language Model (LLM)-driven analysis: The spectral embedding vector is input as soft cue into a fine-tuned chemical domain large language model, and the LLM directly generates the structural analysis results. Example LLM input template: [Spectral Embedding] Based on the above HPLC-MS and NMR spectral data, infer the molecular structure of this compound and output it in SMILES format. Highest confidence candidate structure:

[0046] Enhanced Knowledge Graph Retrieval: The system connects to the reaction knowledge graph database in real time. Before generating results, LLM automatically searches the knowledge graph for historical experimental records similar to the current graph, adding these similar cases as contextual examples to the LLM's prompts to enhance the accuracy and interpretability of the analysis.

[0047] Spectral deconvolution is implemented by employing a variational autoencoder (VAE) model to decouple the mixed spectra. By training a conditional VAE, the model learns to map the mixed spectra to a latent space, and the pure spectra of each component are separated by adjusting the latent variables. This method demonstrates stronger generalization ability for unknown mixtures.

[0048] Training and Update Mechanism: An online incremental learning approach is employed. Each manual correction immediately triggers few-sample fine-tuning of the LLM, while simultaneously updating the case library in the knowledge graph, enabling real-time iterative optimization of the model.

[0049] In summary, by utilizing the technical solutions described above, the data acquisition and standardized preprocessing module automatically acquires and processes multi-source spectral data. Combined with a deep learning analysis engine, chromatographic and spectroscopic data are jointly analyzed, thereby automating and standardizing the spectral analysis process, significantly improving analysis efficiency and consistency, and avoiding the subjectivity and time-consuming issues of manual operation. The automatic reaction path tracking and knowledge graph construction module automatically associates the analysis results with synthesis records and constructs a searchable knowledge graph, enabling the accumulation and reuse of knowledge across the entire chain, from experimental conditions to product structure to transformation pathways, thus providing a data foundation for structure-activity relationship mining and reaction condition optimization. By setting up a dedicated analysis network and spectral deconvolution module, the accuracy of analysis for complex compounds and mixtures is effectively improved. Simultaneously, combined with a human-computer interaction confirmation and feedback optimization mechanism, continuous model iteration and knowledge updates are achieved.

[0050] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent characterization and analysis system for chemical experimental products, characterized in that, include: The data acquisition and standardization preprocessing module is used to automatically acquire the raw spectral data of the analytical instrument and preprocess the raw spectral data to generate standardized spectral data. A deep learning parsing engine is used to receive the standardized spectral data and parse the standardized spectral data based on a deep learning model to obtain parsing results that include at least the product structure. The automatic reaction path tracking and knowledge graph construction module is used to automatically associate the analysis results with the synthesis records in the upstream experimental execution system, infer the transformation path from reactants to products, and store the synthesis records, the analysis results, and the transformation path as knowledge units in the reaction knowledge graph.

2. The intelligent characterization and analysis system for chemical experimental products based on deep learning according to claim 1, characterized in that, The deep learning parsing engine includes: The intelligent chromatographic analysis subnetwork is used to automatically detect peaks, calculate integrals, and analyze peak purity in the chromatographic data of the standardized spectral data; A smart spectral analysis subnetwork is used to identify compound structures from the spectral data in the standardized spectral data.

3. The intelligent characterization and analysis system for chemical experimental products based on deep learning according to claim 2, characterized in that, The intelligent spectral analysis subnetwork adopts a multi-task learning framework, simultaneously training three tasks: compound structure identification, functional group recognition, and spectrum-structure consistency verification.

4. The intelligent characterization and analysis system for chemical experimental products based on deep learning according to claim 1, characterized in that, The automatic reaction path tracking and knowledge graph construction module includes: The data association submodule is used to associate the product structure, purity, and by-product information in the analysis results with the synthesis record; The transformation path reasoning submodule is used to reason about the transformation path from reactants to products based on the reactant and product structures. The knowledge graph construction submodule is used to store the synthesis record, the parsing result, and the transformation path as knowledge units in the reaction knowledge graph.

5. The intelligent characterization and analysis system for chemical experimental products based on deep learning according to claim 1, characterized in that, It also includes dedicated analytical networks for peptides or natural product compounds, which employ targeted molecular feature extraction and network structure design to improve the analytical accuracy of the corresponding categories of compounds.

6. A deep learning-based intelligent characterization and analysis method for chemical experimental products, characterized in that, Includes the following steps: Step 1, Multi-source spectral data acquisition and standardized preprocessing: Automatically connect to the data interface of the analysis instrument, acquire raw spectral data in real time, and preprocess the acquired data to generate standardized spectral data tensors; Step 2, Spectral decoupling and structure identification based on deep learning: Input the standardized spectral data into the deep learning parsing engine to predict the corresponding compound structure and output the candidate structural formula and its confidence level; Step 3, Automatic Reaction Path Association and Knowledge Graph Construction: The product structure information parsed in Step 2 is automatically associated with the synthesis records in the upstream experimental execution system. Based on the reactant and product structures, a reasonable transformation path from reactant to product is deduced, and "synthesis conditions - characterization results - transformation path" is stored as a knowledge unit in the reaction knowledge graph database.

7. The intelligent characterization and analysis method for chemical experimental products based on deep learning according to claim 5, characterized in that, Also includes: Before or during step 2, when a chromatographic peak is determined to be a mixture or a mixed spectrum is directly input, the spectrum deconvolution and mixture analysis steps are executed: the spectrum deconvolution module is called to decouple the mixed spectrum, separate the contribution spectra of each component, and then input them into the spectral analysis subnetwork for structure identification.

8. The intelligent characterization and analysis method for chemical experimental products based on deep learning according to claim 5, characterized in that, Also includes: Step 4: Visualization of Results and Human-Computer Interaction Confirmation and Feedback Optimization: Visualize the analysis results on the graphical interface and receive verification and correction operations from chemists. Use the corrected data as new training data and feed it back to the deep learning engine for iterative optimization.

9. The intelligent characterization and analysis method for chemical experimental products based on deep learning according to claim 5, characterized in that, In step 2, the deep learning parsing engine adopts an ensemble learning strategy, which processes the data through multiple different base learners and combines the outputs of each base learner to obtain the final structure prediction result. The multiple different base learners include a one-dimensional convolutional neural network for processing the original spectral sequence, a gradient boosting tree for processing artificially designed chemical feature vectors, and a graph neural network for converting the spectral data into a molecular fingerprint to mine molecular topological structure features.

10. The intelligent characterization and analysis method for chemical experimental products based on deep learning according to claim 5, characterized in that, In step 2, the deep learning parsing engine uses a large language model for parsing, specifically including: The standardized spectral data is converted into spectral embedding vectors using a spectral encoder. The spectral embedding vector is used as a soft cue input into a fine-tuned chemical domain large language model, which then generates structural analysis results. Furthermore, before generating results, the large language model automatically retrieves historical experimental records in the reaction knowledge graph that are similar to the current graph, and adds the retrieved similar cases as contextual examples to the prompt words.