A molecular structure recognition method and device based on a convolutional neural network

Abstract

The present application relates to the field of spectrum analysis, in particular to a molecular structure identification method and device based on convolutional neural network, by receiving a spectrum to be processed; the spectrum to be processed includes a small molecule spectrum to be processed and a polymer spectrum to be processed; the spectrum to be processed is input into a pre-trained molecular structure identification convolutional neural network model to obtain functional group information; the molecular structure identification convolutional neural network model is a convolutional neural network model trained by experimental spectrum and calculated spectrum; the molecular structure corresponding to the spectrum to be processed is determined according to the functional group information. The present application uses experimental spectrum reflecting real data to ensure high prediction accuracy of the model completed by training, and uses calculated spectrum to further expand the training sample, uses the characteristics of low cost and easy availability of calculated spectrum to make the model learn more features, and reduces the training cost while improving the generalization ability and accuracy of the model.

Description

Technical Field

[0001] This invention relates to the field of spectral analysis, and in particular to a method and apparatus for molecular structure recognition based on convolutional neural networks. Background Technology

[0002] Spectral analysis is a crucial technique for determining the structure of unknown substances in experimental sciences such as chemistry, biology, and materials science. It serves as an essential bridge between instrument-measured data signals and the chemical structure of a substance, and is widely used in molecular characterization of new drugs, biomarker discovery in metabolomics, and the identification of toxic pollutants in environmental samples. Its accuracy directly impacts the efficiency of scientific research. Existing spectral analysis methods can be mainly categorized into three types: rule-based manual analysis, traditional machine learning methods, and deep learning methods based on convolutional neural network models.

[0003] Of the three methods mentioned above, rule-based manual analysis is time-consuming, labor-intensive, highly repetitive, and subjective. For complex molecules (such as natural products and macromolecular polymers), the spectral information overlaps and interferes severely, making manual analysis extremely difficult and resulting in low accuracy. On the other hand, traditional machine learning methods rely heavily on feature engineering, which results in significant information loss and fails to fully utilize the complete information of the original spectrum. Furthermore, when the number of features increases but the number of samples is insufficient, the model performance drops sharply. Finally, deep learning methods based on convolutional neural network models heavily rely on a large amount of high-quality labeled data, requiring massive amounts of high-quality and accurately labeled spectrum-structure pairing data. This is extremely costly to obtain in specialized fields and its application is limited in areas where data is scarce or where a clear mechanism is required, resulting in poor versatility.

[0004] Therefore, how to reduce the cost of model training and improve the versatility of molecular structure recognition while ensuring high accuracy of model analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a molecular structure recognition method and device based on convolutional neural networks, so as to solve the problem that high accuracy, low training cost and high versatility of molecular structure recognition models in the prior art cannot be achieved simultaneously.

[0006] To address the aforementioned technical problems, this invention provides a molecular structure recognition method based on convolutional neural networks, comprising:

[0007] Receive spectra to be processed; the spectra to be processed include spectra of small molecules to be processed and spectra of polymers to be processed.

[0008] The spectrum to be processed is input into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained by experimental and calculated spectra.

[0009] The molecular structure corresponding to the spectrum to be processed is determined based on the functional group information.

[0010] Optionally, in the aforementioned molecular structure recognition method based on convolutional neural networks, accepting the spectrum to be processed includes:

[0011] Receive at least one of the infrared spectrum, Raman spectrum, and NMR spectrum to be processed.

[0012] Optionally, in the aforementioned molecular structure recognition method based on convolutional neural networks, the training method for the molecular structure recognition convolutional neural network model includes:

[0013] Receive raw spectral data; the raw spectral data includes experimental spectra and calculated spectra, and each spectrum in the raw spectral data has corresponding known molecular structure information;

[0014] The original spectral data is preprocessed to obtain standard spectral data;

[0015] Based on the known molecular structure information, each spectrum in the standard spectral data is labeled to obtain labeled spectral data;

[0016] The spectra in the labeled spectral data are assigned to the training set, validation set and test set according to a preset first ratio, and the distribution ratio of each functional group category is consistent in the three subsets.

[0017] The training set and the validation set are used to train the preset initial convolutional neural network model until the difference between the simulation results of the spectra in the test set and the known molecular structure information of the spectra in the test set is less than a preset allowable threshold. The latest convolutional neural network model is then output as the molecular structure recognition convolutional neural network model.

[0018] Optionally, in the aforementioned molecular structure recognition method based on convolutional neural networks, the first proportion is that the proportion of spectra in the training set is not less than 70%.

[0019] The ratio of the number of spectra in the validation set to the number of spectra in the test set ranges from 0.5 to 1.5, including the endpoint values.

[0020] Optionally, in the aforementioned molecular structure recognition method based on convolutional neural networks, the original spectral data is preprocessed to obtain standard spectral data, including:

[0021] The original spectral data is subjected to at least one preprocessing step, namely baseline correction, noise reduction, and normalization, to obtain standard spectral data.

[0022] Optionally, in the aforementioned molecular structure recognition method based on convolutional neural networks, the convolutional neural network model for molecular structure recognition is a convolutional neural network model employing a progressive channel growth strategy.

[0023] Optionally, in the aforementioned molecular structure recognition method based on convolutional neural networks, the convolution kernel of the molecular structure recognition convolutional neural network model is a one-dimensional convolution kernel;

[0024] The one-dimensional convolution kernel is used to perform sliding convolution operations along a single dimension of the molecular structure sequence to extract local features.

[0025] Optionally, in the aforementioned molecular structure recognition method based on convolutional neural networks, the fully connected layer of the molecular structure recognition convolutional neural network model is a fully connected layer employing progressive dimensionality reduction.

[0026] Optionally, in the aforementioned molecular structure recognition method based on convolutional neural networks, the optimizer of the molecular structure recognition convolutional neural network model is an adaptive learning rate optimizer with momentum.

[0027] The adaptive learning rate optimizer includes a weight decay term.

[0028] A molecular structure recognition device based on a convolutional neural network, comprising:

[0029] A receiving module is used to receive spectra to be processed; the spectra to be processed include spectra of small molecules to be processed and spectra of polymers to be processed.

[0030] The neural network model module is used to input the spectrum to be processed into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained by experimental and calculated spectra.

[0031] The molecular derivation module is used to determine the molecular structure corresponding to the spectrum to be processed based on the functional group information.

[0032] The molecular structure recognition method based on convolutional neural networks provided by this invention involves receiving a spectrum to be processed, including spectra of small molecules and polymers; inputting the spectrum into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained using experimental and computational spectra; and determining the molecular structure corresponding to the spectrum based on the functional group information. This invention utilizes both experimentally obtained spectra and computationally obtained spectra to train the convolutional neural network model, resulting in a corresponding molecular structure recognition convolutional neural network model. While using experimental spectra reflecting real data to ensure high prediction accuracy of the trained model, the invention further expands the training sample using computational spectra. Leveraging the low cost and easy availability of computational spectra, the model learns more features, improving its generalization ability and accuracy while reducing training costs. This invention also provides a molecular structure recognition device based on convolutional neural networks with the aforementioned beneficial effects. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart illustrating a specific implementation of the molecular structure recognition method based on convolutional neural networks provided by the present invention;

[0035] Figure 2 A flowchart illustrating a specific implementation of the training method for the molecular structure recognition convolutional neural network model provided by the present invention;

[0036] Figure 3 A schematic diagram of a specific embodiment of the molecular structure recognition device based on a convolutional neural network provided by the present invention;

[0037] Figure 4 A schematic diagram of a convolutional neural network structure for a specific embodiment of the molecular structure recognition device based on a convolutional neural network provided by the present invention.

[0038] Figure label:

[0039] 100 - Receiver module, 200 - Neural network model module, 300 - Molecular derivation module. Detailed Implementation

[0040] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The core of this invention is to provide a molecular structure recognition method based on convolutional neural networks, and a flowchart of one specific implementation is shown below. Figure 1 As shown, this is referred to as Specific Implementation Method One, which includes:

[0042] S101: Receive the spectrum to be processed; the spectrum to be processed includes the spectrum of the small molecule to be processed and the spectrum of the polymer to be processed.

[0043] It should be noted that the small molecule spectra refer to those corresponding to organic compounds with a molecular weight ≤ 1000 Daltons (Da). The model in this invention is not only effective for the qualitative identification of functional groups in small organic molecule compounds, but also exhibits excellent applicability and reliability for the qualitative identification of functional groups in polymers with complex chain structures and higher molecular weights. Specifically, the types of functional groups include, but are not limited to, alkanes, alkenes, alkynes, aromatics, haloalkanes, alcohols, aldehydes, ketones, carboxylic acids, acid anhydrides, acyl halides, esters, ethers, amines, amides, nitriles, imides, imines, azo compounds, thiols, thioaldehydes, phenols, enols, sulfones, sulfonic acids, hydrazines, enamines, isocyanates, isothiocyanates, phosphine, sulfonamides, sulfonates, sulfoxides, thioamides, hydrazones, carbamates, thioethers, etc., which can be selected according to actual needs.

[0044] In one specific implementation, the process of accepting the spectrum to be processed in this step includes:

[0045] Receive at least one of the infrared spectrum, Raman spectrum, and NMR spectrum to be processed.

[0046] In other words, the spectrum to be processed in this specific embodiment includes at least one of the infrared spectrum to be processed, the Raman spectrum to be processed, and the NMR spectrum to be processed. This greatly expands the application scenarios and enhances the versatility of the present invention. Of course, the spectrum to be processed can also be other types of spectra, and the present invention does not limit them here.

[0047] S102: Input the spectrum to be processed into a pre-trained molecular structure recognition convolutional neural network model to obtain functional group information; the molecular structure recognition convolutional neural network model is a convolutional neural network model trained by experimental and calculated spectra.

[0048] The molecular structure recognition convolutional neural network model mainly relies on experimental spectra, supplemented by some computational spectra, for training to ensure the accuracy of model prediction. For example, in one specific embodiment, the training of the molecular structure recognition convolutional neural network model utilizes approximately 160,000 experimental spectra, while only using approximately 10,000 computational spectra. Of course, the specific proportion and number of spectra can be changed according to actual needs, and this invention does not limit them.

[0049] As a preferred embodiment, the molecular structure recognition convolutional neural network model is a convolutional neural network model employing a progressive channel growth strategy.

[0050] For example, the channel growth of the convolutional neural network model is as follows: 1 channel → 32 channels → 64 channels → 128 channels → 256 channels. This allows basic features to be extracted with a small number of channels in shallow layers, while the number of channels is increased appropriately in deeper layers, avoiding parameter inflation caused by too many channels in the early stages.

[0051] On the other hand, the convolution kernel of the molecular structure recognition convolutional neural network model is a one-dimensional convolution kernel;

[0052] The one-dimensional convolution kernel is used to perform sliding convolution operations along a single dimension of the molecular structure sequence to extract local features.

[0053] In this specific embodiment, the convolutional kernel is defined as a one-dimensional convolutional kernel. The receptive field can be expanded by stacking multiple layers of the one-dimensional convolutional kernel, which saves more parameters and improves model efficiency than directly using a large convolutional kernel. Specifically, a 3*1 convolutional kernel can be used. Furthermore, a pooling layer can be followed immediately after each convolutional layer to gradually reduce the spatial dimension of the feature map, thereby effectively controlling the input size of subsequent fully connected layers.

[0054] Furthermore, the fully connected layer of the molecular structure recognition convolutional neural network model is a fully connected layer that employs progressive dimensionality reduction.

[0055] Continuing with the example, the dimensionality reduction can be achieved by going from 15872 to 4800 to 3200 to 1600 to n_classes (the specific value of n_classes can be adjusted to any value based on the number of functional group types to be identified), thus avoiding information loss and parameter redundancy caused by a sudden drop in dimensionality, and improving the accuracy and reliability of the simulation. Furthermore, adding a batch normalization layer can accelerate convergence and provide reasonable parameter initialization, enabling the model to maintain a fast inference speed per sample while achieving a good balance between storage space and computational efficiency.

[0056] Combining the above preferred embodiments, it is possible to achieve lightweight modeling, enable local deployment of convolutional neural network models, and improve the versatility of the present invention.

[0057] S103: Determine the molecular structure corresponding to the spectrum to be processed based on the functional group information.

[0058] In addition, the optimizer of the molecular structure recognition convolutional neural network model is an adaptive learning rate optimizer with momentum.

[0059] The adaptive learning rate optimizer includes a weight decay term.

[0060] This preferred embodiment further defines the optimizer. The optimizer defined by this preferred embodiment achieves a good balance between convergence speed and generalization ability. It retains the ultra-fast convergence brought by adaptation and momentum, and achieves very high generalization ability through weight decay. It completely solves the defects of traditional adaptive optimizers that are "fast to train but poor in effect". At the same time, it has strong robustness and reduces training cost.

[0061] Specifically, the optimizer can be the AdamW optimizer; the AdamW optimizer is a variant of the Adam algorithm that achieves more effective regularization by decoupling the weight decay and gradient update steps, with a base learning rate set to 0.00025. This optimizer configuration combines the advantages of momentum and adaptive learning rate adjustment, enabling efficient and stable searching for optimal model parameters in complex non-convex optimization spaces.

[0062] The molecular structure recognition method based on convolutional neural networks provided by this invention involves receiving a spectrum to be processed, including spectra of small molecules and spectra of polymers; inputting the spectrum to be processed into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained using experimental and computational spectra; and determining the molecular structure corresponding to the spectrum to be processed based on the functional group information. This invention utilizes both experimental spectra obtained through actual experiments and computational spectra obtained through theoretical calculations to train the convolutional neural network model to obtain the corresponding molecular structure recognition convolutional neural network model. While using experimental spectra reflecting real data to ensure high prediction accuracy of the trained model, the use of computational spectra further expands the training sample. Leveraging the low cost and easy availability of computational spectra, the model learns more features, improving the model's generalization ability and accuracy while reducing training costs.

[0063] As one specific implementation method, the flowchart of the training method for the pre-trained molecular structure recognition convolutional neural network model mentioned above is as follows: Figure 2 As shown, it includes:

[0064] A1: Receive raw spectral data; the raw spectral data includes experimental spectra and calculated spectra, and each spectrum in the raw spectral data has corresponding known molecular structure information.

[0065] The original spectral data is a collection of multiple spectra. Similarly, the standard spectral data below is also a collection of preprocessed spectra, which will not be described further hereafter.

[0066] A2: Preprocess the original spectral data to obtain standard spectral data.

[0067] As one specific implementation method, this step includes:

[0068] The original spectral data is subjected to at least one preprocessing step, namely baseline correction, noise reduction, and normalization, to obtain standard spectral data.

[0069] By performing at least one of the preprocessing steps, namely baseline correction, noise reduction, and normalization, the signal-to-noise ratio of the original spectral data can be greatly improved, the representativeness of the data in the spectral data can be enhanced, and the training effect of the subsequent model can be improved.

[0070] Furthermore, after obtaining the standard spectrum data through preprocessing, the spectrum data can be aligned according to the peak positions of the standard spectrum data, and dimensionality reduction can be performed on the data. Without reducing the representativeness of the data, the amount of data to be processed is reduced, the computational pressure is alleviated, and the processing efficiency is improved. In other words, normalization and data alignment are mathematical processing methods. Data dimensionality reduction maps high-dimensional data to a low-dimensional space while preserving as much key information of the original data as possible, thus addressing the curse of dimensionality and improving model efficiency and generalization ability (each spectrum data used for training has approximately 1300-1400 data points; reducing them to the same dimension can effectively reduce computational complexity).

[0071] A3: Based on the known molecular structure information, each spectrum in the standard spectral data is labeled to obtain labeled spectral data.

[0072] In this step, classification labels are added to the preprocessed standard spectral data based on the known molecular structure corresponding to each spectrum. These labels are used to indicate whether the molecular structure represented by the corresponding spectrum contains one or more target functional groups or substructures, thereby constructing a labeled dataset with a structure-spectral correspondence.

[0073] A4: Assign the spectra in the labeled spectral data to the training set, validation set, and test set according to a preset first ratio, and make the distribution ratio of each functional group category consistent in the three subsets.

[0074] The labeled spectral data is divided proportionally using a stratified sampling method based on the functional group categories to be identified. Specifically, the first ratio is 8:1:1, meaning the labeled spectral data is divided into the training set, the validation set, and the test set in an 8:1:1 ratio, ensuring that the distribution ratio of each functional group category remains consistent across the three subsets (i.e., the training set, the validation set, and the test set), thereby guaranteeing the balance between model training and evaluation.

[0075] A5: Train the preset initial convolutional neural network model using the training set and the validation set until the difference between the simulation results of the spectra in the test set and the known molecular structure information of the spectra in the test set is less than a preset allowable threshold. Then output the latest convolutional neural network model as the molecular structure recognition convolutional neural network model.

[0076] Of course, before performing this step, an initial convolutional neural network model needs to be built. Specifically, a deep learning model based on a convolutional neural network model needs to be built to automatically extract deep features from the spectral map and achieve efficient, accurate, and batch recognition of functional group categories. The model includes a network architecture definition module and a training strategy definition module. The network architecture definition module is used to configure the structure and parameters of convolutional layers, pooling layers, and fully connected layers, while the training strategy definition module is used to set the optimizer, learning rate scheduling, and regularization strategy. The two modules work together to achieve modular and parameterized control from model structure to training process, enhancing the reproducibility of the method and the scalability of the system.

[0077] Specifically, in the initial training of the convolutional neural network model in this step, the model is trained using the training set and the validation set. During training, the difference between the model's predictions and the true labels is calculated, and the network parameters are updated using the backpropagation algorithm. The learning rate and hyperparameters are fine-tuned based on the performance on the validation set to improve the model's recognition accuracy and generalization ability. Then, the test set is input into the trained model for functional group recognition testing. Based on the extracted spectral features, the model outputs the probability of the presence of the corresponding functional group for each test sample, further generating a recognition result chart to visually display the performance indicators such as recognition accuracy and recall for each functional group.

[0078] Through the above steps, this invention achieves rapid and automated analysis of complex spectra, and is particularly suitable for functional group identification tasks of large batches of spectral data. While maintaining the original spectral form, it effectively improves analysis efficiency and accuracy.

[0079] After clarifying the molecular structure recognition method based on convolutional neural networks and the training method of the convolutional neural network model for molecular structure recognition described above, this invention further provides a specific embodiment (i.e., Embodiment 1), which is an intelligent qualitative method for infrared spectroscopy based on convolutional neural networks. This can be understood in conjunction with the preceding text. This embodiment targets experimental infrared spectral data of 140,000 small organic molecules. Based on a convolutional neural network architecture, feature extraction and classification of the spectra are performed to obtain a method for functional group recognition of small organic molecules based on infrared spectroscopy technology. The method includes the following steps:

[0080] Step 1: Data preprocessing aims to improve the quality and consistency of input data. First, the system reads 145,526 experimental infrared spectra of small organic molecules and standardizes each raw spectrum, including normalization, data alignment, and dimensionality reduction. Each raw spectrum is first normalized because the numerical ranges of different spectral bands vary significantly. Directly inputting these into the model could lead to training instability and affect convergence speed. Normalization scales all bands to a uniform range (0-1), preventing specific bands from excessively influencing the model and thus improving training stability and model accuracy. Second, data alignment is performed, aligning and unifying the wavelength shift ranges of all infrared spectra to ensure that all input spectral data are compared and modeled on the same feature scale (wavenumber axis), thereby eliminating peak position shifts caused by instrument drift, sampling interval differences, or different data sources. Finally, dimensionality reduction is used, interpolating the spectral data to a uniform dimensionality of 1024 dimensions to compress information redundancy, reduce computational complexity, and further accelerate model convergence.

[0081] Through the above processing, this invention, while maintaining the authenticity of the data, further ensures the quality of the input data and the robustness of the model training.

[0082] Step 2: Data annotation aims to establish a precise correlation between the preprocessed spectral data and its corresponding molecular structure. This example mainly focuses on the systematic annotation of 10 common functional groups in organic molecules (including alkanes, alkenes, alkynes, aromatics, haloalkanes, alcohols, aldehydes, ketones, esters, and ethers).

[0083] First, this embodiment uses the cheminformatics toolkit RDKit to process all known molecular structures involved, utilizing its built-in functional group identification and substructure matching functions to systematically scan and determine whether the molecules contain preset target functional groups or substructures. Based on the identification results, a corresponding multi-label classification vector is generated for each molecule. Each dimension of the vector corresponds to a functional group; if the molecule contains that functional group or substructure, it is labeled as 1, otherwise as 0. Finally, this binary label vector is precisely associated with the corresponding preprocessed spectral data, thereby constructing a multi-labeled dataset with a reliable "structure-spectrum" correspondence, providing a training foundation for subsequent supervised learning models.

[0084] Step 3: Data Partitioning. To ensure the effectiveness of model training and the reliability of evaluation, this embodiment employs stratified sampling to partition the labeled spectral dataset. Specifically, based on the 10 target functional group categories to be identified, the entire dataset is divided into training, validation, and test sets in an 8:1:1 ratio. This method ensures that the distribution ratio of each functional group sample in the three subsets is highly consistent with the overall distribution of the original dataset, effectively avoiding biases that may arise during model training due to uneven distribution of functional group samples. By maintaining the representativeness of the category distribution in each data subset, this partitioning method aims to improve the model's sensitivity to identifying a few functional group categories and ultimately enhance the model's overall generalization ability and robustness on unknown spectral data.

[0085] Step 4: Model Construction and Feature Extraction. A convolutional neural network is constructed to extract deeper spectral features from Step 3, enabling rapid, accurate, batch, and lossless identification of the types of functional groups contained in molecules.

[0086] The convolutional feature extraction module consists of four convolutional layers, which can be referenced. Figure 4 , Figure 4 This is a schematic diagram of a specific implementation of a convolutional neural network, used to abstract and compress the input signal layer by layer. Each convolutional processing unit includes a one-dimensional convolutional layer (conv1D), a batch normalization layer (BatchNorm), a rectified linear activation function layer (ReLU), and a max pooling layer, forming a standard processing flow. Figure 4The corresponding convolutional neural network consists of four convolutional processing units. After processing by these units, the data is flattened before being input into a fully connected layer, and finally output. Local feature extraction is achieved by increasing the number of output channels unit by unit and reducing the sequence length.

[0087] Specifically, the first convolutional processing unit is configured to receive a single-channel one-dimensional input signal and map it to 32 feature channels using a first convolutional kernel group. The first convolutional kernel group has a spatial dimension of 3, a stride of 1, and employs zero-padding to maintain the length of the output sequence. The second convolutional processing unit expands the number of channels to 64 while maintaining the same convolution and padding strategy. The third and fourth convolutional processing units increase the number of feature channels to 128 and 256, respectively. The key difference between these units and the first and second convolutional processing units is that the corresponding convolutional operations do not use zero-padding, causing the length of the feature map sequence to gradually decrease during the convolution stage. This design deepens the network while gradually reducing the size of the feature maps, allowing the model to focus on more global and essential feature patterns. Then, a batch normalization layer is used to stabilize the training process: the data distribution of the neural network layer outputs is unified and adjusted to a standard normal distribution with a mean of 0 and a variance of 1, ensuring the output is located in the non-saturated region of the activation layer, thus accelerating convergence. The activation function of the convolutional layer is a non-linear activation function, and the pooling operation adopts the max pooling method. The kernel size and sliding stride are both set to 2 to achieve binary subsampling of the feature map, compress the data and retain significant features.

[0088] After layer-by-layer processing through four convolutional and pooling blocks, the high-dimensional convolutional feature maps need to be transformed for use by the classifier. The flattening operation flattens the 256-channel, significantly shortened feature map of each sample into a one-dimensional long vector, which is then input into the fully connected classification module.

[0089] The fully connected classification module receives the multidimensional feature tensor output by the convolutional feature extraction module, flattens it into a one-dimensional feature vector, and then performs nonlinear transformation and dimensionality compression. This module consists of four sequentially connected linear transformation layers. The first three layers reduce the feature vector dimension from 15872 to 4800, 3200, and 1600 respectively. Each layer is followed by a ReLU activation function to maintain the network's nonlinear fitting capability. The final linear transformation layer maps the 1600-dimensional features to 10 output dimensions corresponding to the number of target categories, forming the classification logical value.

[0090] Specifically, at the end of the model's forward propagation path, before the output layer of the neural network, an independent Sigmoid activation function is applied to the raw output of the last fully connected layer. The Sigmoid function is applied to each output neuron independently, ensuring that the predicted probability of each class does not depend on the output of other classes. This allows a sample to have multiple high-probability labels simultaneously, satisfying the non-mutually exclusive requirement of multi-label tasks.

[0091] To optimize network training dynamics and generalization performance, the architecture integrates several auxiliary components. A batch normalization layer following each convolutional layer normalizes the output features to accelerate training convergence and stabilize gradient flow. The network further includes optional random deactivation layers to suppress overfitting by randomly masking neurons during forward propagation.

[0092] During the model parameter initialization phase, a differentiated strategy is employed to ensure training stability. Convolutional layer weights are initialized using the "Kaiming Normal" method, which aligns with the characteristics of the ReLU activation function, ensuring the signal maintains appropriate variance during forward propagation. Fully connected layer weights are initialized using a normal distribution with a mean of 0 and a standard deviation of 2%. All learnable bias parameters are initialized to zero.

[0093] This architecture, through the synergy of the convolutional feature extraction module and the fully connected classification module, can automatically learn highly discriminative feature representations from the original one-dimensional time-series data, and finally generate classification decisions through the output layer.

[0094] Step 5: Model Training. The model is trained using the data segmented in Step 3. By setting the learning rate and hyperparameters, the difference between the predicted and true labels is calculated using the binary cross-entropy loss function (BCELoss).

[0095] This embodiment provides a complete training method for a one-dimensional convolutional neural network classification model. The method is implemented through a training execution function `train_model`, which receives a predefined model instance, storage path, dataset object, and multiple training control parameters, automating the model's training, validation, and monitoring processes.

[0096] First, the method initializes the training monitoring and recording component. Specifically, a SummaryWriter instance is created, and its log directory is set to a specified storage path to record key metrics such as loss and accuracy during the training process in real time, enabling visual tracking of the training process.

[0097] Secondly, during the training phase, the training strategy definition module uses another structured dictionary to fine-tune the model training process. The batch size is set to 64, and the maximum number of training epochs is 100. During training, the module automatically saves the current best model by comparing its classification accuracy on the validation set.

[0098] Subsequently, the method uses the BCELoss function, which is suitable for multi-label classification tasks, to independently calculate the cross-entropy between the predicted probability of each class and its true binary label (0 or 1), and sums or averages the losses of all classes and all samples to obtain the difference between the predicted label and the true label.

[0099] For the optimizer configuration, the AdamW optimization algorithm was selected. This optimizer is a variant of the Adam algorithm, which achieves more effective regularization by decoupling the weight decay and gradient update steps. Its base learning rate was set to 0.00025. This optimizer configuration combines the advantages of momentum and adaptive learning rate adjustment, enabling efficient and stable searching for optimal model parameters in complex non-convex optimization spaces.

[0100] The entire training process is executed on a designated graphics processing unit (GPU). By iteratively calling the data loader, forward propagation, loss calculation, backpropagation, and parameter update steps, and combining an early stopping mechanism with a model checkpoint saving strategy, the model is ensured to be fully trained while avoiding overfitting. Ultimately, a high-performance, high-generalization-ability convolutional neural network functional group qualitative recognition model based on infrared spectral data of organic small molecules is obtained.

[0101] Step Six: Test Model Performance. Based on the learned high-dimensional spectral features, the model identifies and classifies the functional groups contained in each spectral data, and finally generates visualized recognition results and corresponding model metrics (including precision, recall, F1 score, and the micro-mean and weighted average of these metrics).

[0102] First, this embodiment provides a method for testing and evaluating multi-label binary classification models. This method uses reserved test set data to objectively evaluate the performance of the optimal model obtained from training in step five, comprehensively reflecting the model's classification efficiency through multi-dimensional indicators.

[0103] Secondly, this method continues the intelligent task recognition mechanism from the training phase. For multi-label binary classification tasks, the binary cross-entropy loss function BCELoss is selected as the evaluation metric (although backpropagation is not performed during the testing phase, this choice maintains consistency in the evaluation logic). The method initializes a testing engine, encapsulating the model, data loader, loss function, and computing device, and calls its `test_epoch` method to perform batch forward propagation. In multi-label binary classification scenarios, the engine automatically applies a Sigmoid function to the model's raw output, converting the logistic values ​​into independent probability values, and then binarizes the probabilities into the final predicted category label using a preset threshold (usually 0.5).

[0104] Finally, the method outputs a comprehensive evaluation report via the sklearn.metrics.classification_report function.

[0105] Specifically, the evaluation system calculates the following core indicators during the testing phase:

[0106] Accuracy: For each category, the proportion of true positives among the samples predicted as positive is calculated, measuring the accuracy of the model in classifying each category.

[0107] Recall: For each category, the proportion of true positives that are correctly predicted is calculated, measuring the model's coverage of each category.

[0108] F1 score: The harmonic mean of precision and recall, providing a comprehensive and balanced evaluation metric for each category.

[0109] To evaluate the model holistically, the system further calculates micro-means and weighted averages. Additionally, label-level accuracy is calculated and recorded, representing the proportion of correctly predicted labels across all samples, evaluating model performance at the micro-level. Simultaneously, the perfect match rate is calculated, a stringent metric requiring all labels for a sample to be correctly predicted, reflecting the model's overall accuracy in multi-label scenarios. All evaluation results are simultaneously output to the console and written to a log file, ensuring the traceability and reproducibility of the evaluation process. This method returns the model's raw output, predicted labels, and true labels, providing a complete data foundation for subsequent error analysis and model improvement.

[0110] Table 1. Qualitative analysis results of 10 functional groups in experimental infrared spectra of small organic molecules based on the convolutional neural network model.

[0111] Functional groups Accuracy Recall rate F1 score Sample size Alkanes 0.9828 0.9783 0.9806 12036 Olefins 0.9286 0.8366 0.8802 2209 alkynes 0.9825 0.8155 0.8912 206 Aromatics 0.9767 0.982 0.9793 8318 Halogenated alkanes 0.9158 0.8442 0.8786 3531 alcohol 0.9612 0.9385 0.9497 1610 aldehyde 0.9767 0.9005 0.9371 372 ketone 0.9375 0.8920 0.9142 1648 ester 0.9493 0.9627 0.9560 2040 ether 0.9342 0.8778 0.9051 2668 Micro average 0.9628 0.9401 0.9513 34638 Weighted average 0.9621 0.9401 0.9506 34638

[0112] Table 1 shows the results of the experimental infrared spectroscopy qualitative analysis of small organic molecules based on the convolutional neural network model. The model achieved a functional group qualitative accuracy of 97.71% on the test set and a complete matching rate of 82.53% for molecular structure prediction.

[0113] Specifically, in this embodiment, experimental infrared spectral data of 10 polymers (including copolymers and homopolymers) were randomly selected, and steps one, two, and six above were repeated to test and evaluate the optimal model obtained from training in step five. The test and evaluation results are shown in Table 2.

[0114] Table 2. Qualitative analysis results of 10 functional groups in polymer experimental infrared spectra based on the convolutional neural network model.

[0115] Functional groups Accuracy Recall rate F1 score Sample size Alkanes 1.0000 0.9000 0.9474 10 Olefins 0.0000 0.0000 0.0000 1 alkynes 0.0000 0.0000 0.0000 0 Aromatics 1.0000 0.5000 0.6667 2 Halogenated alkanes 1.0000 0.5000 0.6667 4 alcohol 1.0000 0.5000 0.6667 2 aldehyde 0.0000 0.0000 0.0000 0 ketone 0.0000 0.0000 0.0000 0 ester 1.0000 1.0000 1.0000 7 ether 0.5000 1.0000 0.6667 1 Micro average 0.9545 0.7778 0.8571 27 Weighted average 0.9444 0.7778 0.8324 27

[0116] Table 2 shows the qualitative analysis results of polymer experiments using infrared spectroscopy based on the convolutional neural network model. The model achieves a qualitative accuracy of 93.00% for functional groups in polymers. This result indicates that the model in this embodiment can not only be effectively used for the qualitative identification of functional groups in small organic molecule compounds, but also has excellent applicability and reliability for the qualitative identification of functional groups in polymers with complex chain structures and higher molecular weights.

[0117] Another embodiment is provided, referred to as Embodiment 2, which is an intelligent qualitative method for Raman spectroscopy based on convolutional neural networks.

[0118] This embodiment proposes a method for spectral feature extraction and classification based on convolutional neural networks for experimental Raman spectral data of small organic molecules, enabling automatic identification of functional groups in small organic molecules based on Raman spectroscopy.

[0119] First, following the data preprocessing procedure described in Example 1, 12,528 experimental Raman spectral data points were read. For each raw spectrum, the following operations were performed sequentially: normalization was performed to eliminate intensity differences; data dimensionality reduction was implemented to improve processing efficiency and model generalization ability; functional group classification was performed on the spectrum according to molecular structure; and the training, validation, and test sets were divided in an 8:1:1 ratio. This example labels ten common functional groups found in small organic molecules, specifically including: alkanes, alkenes, alkynes, aromatics, haloalkanes, alcohols, aldehydes, ketones, esters, and ethers.

[0120] Secondly, repeat step four of Example 1 to construct the convolutional neural network model; and repeat step five to train and optimize the model using the processed Raman spectroscopy data, finally obtaining a functional group qualitative identification model with high performance and high generalization ability for Raman spectroscopy of small organic molecules.

[0121] Subsequently, step six of Example 1 is repeated to conduct systematic performance testing and comprehensive evaluation of the trained optimal model in order to verify its accuracy, robustness and reliability in the functional group identification task.

[0122] Table 3. Qualitative analysis results of 10 functional groups in experimental Raman spectra of small organic molecules based on the convolutional neural network model.

[0123] Functional groups Accuracy Recall rate F1 score Sample size Alkanes 0.9783 0.9726 0.9754 1021 Olefins 0.8889 0.8067 0.8458 119 alkynes 0.9000 0.8571 0.8780 21 Aromatics 0.9663 0.9815 0.9738 701 Halogenated alkanes 0.9059 0.8756 0.8905 418 alcohol 0.6917 0.7132 0.7023 129 aldehyde 0.8696 0.8163 0.8421 49 ketone 0.8304 0.7209 0.7718 129 ester 0.8820 0.8606 0.8712 165 ether 0.8963 0.6901 0.7798 213 Micro average 0.9304 0.9022 0.9161 2965 Weighted average 0.9292 0.9022 0.9144 2965

[0124] Table 3 shows the qualitative analysis results of Raman spectroscopy for small organic molecules based on a convolutional neural network model. The model achieved a functional group category qualitative accuracy of 96.09% on the independent test set, and a complete matching rate of 70.95% for functional group combinations at the molecular level.

[0125] Specifically, in this embodiment, experimental Raman spectral data of 10 polymers (including copolymers and homopolymers) were randomly selected, and steps one, two, and six described in Example 1 were repeated to test and evaluate the optimal model trained in this embodiment. The test and evaluation results are shown in Table 4.

[0126] Table 4. Qualitative analysis results of 10 functional groups in polymer experimental Raman spectra based on the convolutional neural network model.

[0127] Functional groups Accuracy Recall rate F1 score Sample size Alkanes 1.0000 1.0000 1.0000 10 Olefins 0.0000 0.0000 0.0000 0 alkynes 0.0000 0.0000 0.0000 0 Aromatics 1.0000 1.0000 1.0000 4 Halogenated alkanes 1.0000 1.0000 1.0000 2 alcohol 0.0000 0.0000 0.0000 0 aldehyde 0.0000 0.0000 0.0000 0 ketone 0.0000 0.0000 0.0000 0 ester 1.0000 0.4000 0.5714 5 ether 0.3333 0.3333 0.3333 3 Micro average 0.9048 0.7917 0.8444 24 Weighted average 0.9167 0.7917 0.8274 24

[0128] Table 4 shows the qualitative analysis results of polymer experimental Raman spectroscopy based on the convolutional neural network model. The model achieves a qualitative accuracy of 91.00% for functional groups in polymers. This result indicates that the model in this embodiment can not only be effectively used for the qualitative identification of functional groups in small organic molecule compounds, but also has excellent applicability and reliability for the qualitative identification of functional groups in polymers with complex chain structures and higher molecular weights.

[0129] Another embodiment is provided, referred to as Embodiment 3, which is an intelligent qualitative method for Raman spectroscopy based on convolutional neural networks.

[0130] This embodiment proposes a method for spectral feature extraction and classification based on convolutional neural networks for computational Raman spectroscopy data of small organic molecules, enabling automatic identification of functional groups in small organic molecules based on Raman spectroscopy technology.

[0131] First, following the data preprocessing procedure described in Example 1, calculated Raman spectra of organic small molecules from the QM9 dataset (10846 records) were read. For each raw spectrum, the following operations were performed sequentially: normalization was performed to eliminate intensity differences; data dimensionality reduction was implemented to improve processing efficiency and model generalization ability; functional group classification was performed on the spectra based on molecular structure; and the training, validation, and test sets were divided in an 8:1:1 ratio. This example labels ten common functional groups found in organic small molecules, specifically including: alkanes, alkenes, alkynes, aromatics, haloalkanes, alcohols, aldehydes, ketones, esters, and ethers.

[0132] Secondly, repeat step four of Example 1 to construct the convolutional neural network model; and repeat step five to train and optimize the model using the processed Raman spectroscopy data, finally obtaining a functional group qualitative identification model with high performance and high generalization ability for Raman spectroscopy of small organic molecules.

[0133] Subsequently, step six of Example 1 is repeated to conduct systematic performance testing and comprehensive evaluation of the trained optimal model in order to verify its accuracy, robustness and reliability in the functional group identification task.

[0134] Table 5. Qualitative analysis results of 10 functional groups in the computational Raman spectra of small organic molecules based on the convolutional neural network model.

[0135] Functional groups Accuracy Recall rate F1 score Sample size Alkanes 0.9972 0.9991 0.9981 1057 Olefins 0.9363 0.9018 0.9187 163 alkynes 0.9942 0.9659 0.9798 176 Aromatics 1.0000 0.3333 0.5000 3 Halogenated alkanes 0.6667 0.3077 0.4211 13 alcohol 0.9771 0.9660 0.9715 353 aldehyde 0.9379 0.8882 0.9124 170 ketone 0.7872 0.7351 0.7603 151 ester 0.6271 0.6981 0.6607 53 ether 0.7851 0.7610 0.7728 456 Micro average 0.9289 0.9114 0.9201 2595 Weighted average 0.9278 0.9114 0.9189 2595

[0136] Table 5 shows the qualitative analysis results of Raman spectroscopy for small organic molecules based on a convolutional neural network model. The model achieved a functional group category qualitative accuracy of 96.21% on the independent test set, and a complete matching rate of 70.32% for functional group combinations at the molecular level.

[0137] The molecular structure recognition device based on convolutional neural networks provided in the embodiments of the present invention will be described below. The molecular structure recognition device based on convolutional neural networks described below can be referred to in correspondence with the molecular structure recognition method based on convolutional neural networks described above.

[0138] Figure 3 The structural block diagram of the molecular structure recognition device based on convolutional neural network provided in the embodiments of the present invention is shown below. Figure 3 Molecular structure recognition devices based on convolutional neural networks may include:

[0139] The receiving module 100 is used to receive the spectrum to be processed; the spectrum to be processed includes the spectrum of the small molecule to be processed and the spectrum of the polymer to be processed.

[0140] The neural network model module 200 is used to input the spectrum to be processed into a pre-trained molecular structure recognition convolutional neural network model to obtain functional group information; the molecular structure recognition convolutional neural network model is a convolutional neural network model trained by experimental spectrum and calculated spectrum.

[0141] The molecular derivation module 300 is used to determine the molecular structure corresponding to the spectrum to be processed based on the functional group information.

[0142] In one specific implementation, the receiving module 100 includes:

[0143] The classification receiving unit is used to receive at least one of the infrared spectrum, Raman spectrum, and NMR spectrum to be processed.

[0144] As one specific implementation, the training device for the molecular structure recognition convolutional neural network model includes:

[0145] The training receiving module is used to receive raw spectral data; the raw spectral data includes experimental spectra and calculated spectra, and each spectrum in the raw spectral data has corresponding known molecular structure information;

[0146] The preprocessing module is used to preprocess the original spectral data to obtain standard spectral data;

[0147] The annotation module is used to annotate each spectrum in the standard spectrum data according to the known molecular structure information to obtain annotated spectrum data;

[0148] The classification module is used to classify the spectra in the labeled spectral data into the training set, validation set and test set according to a preset first ratio, and to make the distribution ratio of each functional group category consistent in the three subsets.

[0149] The training module is used to train a preset initial convolutional neural network model using the training set and the validation set until the difference between the simulation results of the spectra in the test set and the known molecular structure information of the spectra in the test set is less than a preset allowable threshold. The latest convolutional neural network model is then output as a molecular structure recognition convolutional neural network model.

[0150] In one specific implementation, the preprocessing module includes:

[0151] A multi-step preprocessing unit is used to perform at least one of the following preprocessing steps on the original spectral data: baseline correction, noise processing, and normalization, to obtain standard spectral data.

[0152] The molecular structure recognition device based on convolutional neural networks provided by this invention includes a receiving module 100 for receiving spectra to be processed; the spectra to be processed include spectra of small molecules and spectra of polymers; a neural network model module 200 for inputting the spectra to be processed into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained by experimental spectra and calculated spectra; and a molecular derivation module 300 for determining the molecular structure corresponding to the spectra to be processed based on the functional group information. This invention utilizes both experimental spectra obtained through actual experiments and calculated spectra obtained through theoretical calculations to train the convolutional neural network model to obtain the corresponding molecular structure recognition convolutional neural network model. While using experimental spectra reflecting real data to ensure high prediction accuracy of the trained model, it further expands the training sample by using calculated spectra. Taking advantage of the low cost and easy availability of calculated spectra, the model learns more features, improving the model's generalization ability and accuracy while reducing training costs.

[0153] The molecular structure recognition device based on convolutional neural networks in this embodiment is used to implement the aforementioned molecular structure recognition method based on convolutional neural networks. Therefore, the specific implementation of the molecular structure recognition device based on convolutional neural networks can be found in the embodiment section of the molecular structure recognition method based on convolutional neural networks mentioned above. For example, the receiving module 100, the neural network model module 200, and the molecular derivation module 300 are respectively used to implement steps S101, S102, and S103 in the aforementioned molecular structure recognition method based on convolutional neural networks. Therefore, the specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0154] The present invention also provides a molecular structure recognition device based on a convolutional neural network, comprising:

[0155] Memory, used to store computer programs;

[0156] A processor is configured to execute the computer program to implement the steps of any of the above-described molecular structure recognition methods based on convolutional neural networks. The molecular structure recognition method based on convolutional neural networks provided by this invention involves receiving a spectrum to be processed; the spectrum to be processed includes a spectrum of small molecules and a spectrum of polymers to be processed; inputting the spectrum to be processed into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained using experimental and computational spectra; and determining the molecular structure corresponding to the spectrum to be processed based on the functional group information. This invention utilizes both experimental spectra obtained through actual experiments and computational spectra obtained through theoretical calculations to train the convolutional neural network model to obtain a corresponding convolutional neural network model for molecular structure recognition. While using experimental spectra reflecting real data to ensure high prediction accuracy of the trained model, computational spectra are used to further expand the training samples. The low cost and easy availability of computational spectra allow the model to learn more features, improving the model's generalization ability and accuracy while reducing training costs.

[0157] This invention also provides a computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the steps of any of the above-described molecular structure recognition methods based on convolutional neural networks. The molecular structure recognition method based on convolutional neural networks provided by this invention involves receiving a spectrum to be processed; the spectrum to be processed includes a spectrum of small molecules to be processed and a spectrum of polymers to be processed; inputting the spectrum to be processed into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained using experimental and computational spectra; and determining the molecular structure corresponding to the spectrum to be processed based on the functional group information. This invention utilizes both experimental spectra obtained through actual experiments and computational spectra obtained through theoretical calculations to train a convolutional neural network model, thereby obtaining a corresponding molecular structure recognition convolutional neural network model. While using experimental spectra that reflect real data to ensure the high prediction accuracy of the trained model, the computational spectra are used to further expand the training samples. Taking advantage of the low cost and easy availability of computational spectra, the model learns more features, thereby improving the model's generalization ability and accuracy while reducing training costs.

[0158] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0159] It should be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0160] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0161] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0162] The molecular structure recognition method and apparatus based on convolutional neural networks provided by this invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of this invention.

Claims

1. A molecular structure recognition method based on convolutional neural networks, characterized in that, include: Receive spectra to be processed; the spectra to be processed include spectra of small molecules to be processed and spectra of polymers to be processed. The spectrum to be processed is input into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained by experimental and calculated spectra. The molecular structure corresponding to the spectrum to be processed is determined based on the functional group information.

2. The molecular structure recognition method based on convolutional neural networks as described in claim 1, characterized in that, Accept spectra to be processed, including: Receive at least one of the infrared spectrum, Raman spectrum, and NMR spectrum to be processed.

3. The molecular structure recognition method based on convolutional neural networks as described in claim 1, characterized in that, The training method for the molecular structure recognition convolutional neural network model includes: Receive raw spectral data; the raw spectral data includes experimental spectra and calculated spectra, and each spectrum in the raw spectral data has corresponding known molecular structure information; The original spectral data is preprocessed to obtain standard spectral data; Based on the known molecular structure information, each spectrum in the standard spectral data is labeled to obtain labeled spectral data; The spectra in the labeled spectral data are assigned to the training set, validation set and test set according to a preset first ratio, and the distribution ratio of each functional group category is consistent in the three subsets. The training set and the validation set are used to train the preset initial convolutional neural network model until the difference between the simulation results of the spectra in the test set and the known molecular structure information of the spectra in the test set is less than a preset allowable threshold. The latest convolutional neural network model is then output as the molecular structure recognition convolutional neural network model.

4. The molecular structure recognition method based on convolutional neural networks as described in claim 3, characterized in that, In the first ratio, the proportion of spectra in the training set is not less than 70%; The ratio of the number of spectra in the validation set to the number of spectra in the test set ranges from 0.5 to 1.5, including the endpoint values.

5. The molecular structure recognition method based on convolutional neural networks as described in claim 3, characterized in that, The original spectral data is preprocessed to obtain standard spectral data, including: The original spectral data is subjected to at least one preprocessing step, namely baseline correction, noise reduction, and normalization, to obtain standard spectral data.

6. The molecular structure recognition method based on convolutional neural networks as described in claim 1, characterized in that, The molecular structure recognition convolutional neural network model is a convolutional neural network model that employs a progressive channel growth strategy.

7. The molecular structure recognition method based on convolutional neural networks as described in claim 1, characterized in that, The convolution kernel of the molecular structure recognition convolutional neural network model is a one-dimensional convolution kernel; The one-dimensional convolution kernel is used to perform sliding convolution operations along a single dimension of the molecular structure sequence to extract local features.

8. The molecular structure recognition method based on convolutional neural networks as described in claim 1, characterized in that, The fully connected layer of the molecular structure recognition convolutional neural network model is a fully connected layer that employs progressive dimensionality reduction.

9. The molecular structure recognition method based on convolutional neural networks as described in claim 1, characterized in that, The optimizer for the molecular structure recognition convolutional neural network model is an adaptive learning rate optimizer with momentum. The adaptive learning rate optimizer includes a weight decay term.

10. A molecular structure recognition device based on a convolutional neural network, characterized in that, include: A receiving module is used to receive spectra to be processed; the spectra to be processed include spectra of small molecules to be processed and spectra of polymers to be processed. The neural network model module is used to input the spectrum to be processed into a pre-trained convolutional neural network model for molecular structure recognition to obtain functional group information; the convolutional neural network model for molecular structure recognition is a convolutional neural network model trained by experimental and calculated spectra. The molecular derivation module is used to determine the molecular structure corresponding to the spectrum to be processed based on the functional group information.

Images