A deep learning material diffraction pattern generation method and device
By generating XRD diffraction patterns directly from the chemical formula of materials using a deep learning model, the problem of relying on crystal structure information in existing technologies is solved, enabling rapid and accurate prediction of material characteristics and virtual experiments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI INST OF TECH
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies cannot directly predict XRD diffraction patterns from material chemical formulas and rely on crystal structure information, resulting in high costs for rapid material screening and virtual experiments.
A deep learning model is used to directly generate XRD diffraction patterns from material chemical formulas through material element property feature extraction and multi-scale Transformer module. The prediction accuracy is improved by adaptive residual connection module, and the prediction uncertainty is calculated by combining random deactivation forward propagation.
It enables the rapid generation of accurate XRD diffraction patterns without crystal structure information, improving prediction accuracy and reliability, reducing data acquisition difficulty, and supporting material screening and virtual experiments.
Smart Images
Figure CN122157906A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of materials science and artificial intelligence, and in particular to a deep learning method and apparatus for directly predicting XRD diffraction patterns based on chemical formulas. Background Technology
[0002] X-ray diffraction (XRD) is an important structural characterization technique in materials science, used to analyze information such as crystal structure, phase composition, and lattice parameters. Traditionally, obtaining XRD diffraction patterns requires experimental methods, i.e., testing material samples using an X-ray diffractometer. This process typically involves multiple steps, including material synthesis, sample preparation, and instrument testing, resulting in a lengthy and costly process. In recent years, deep learning methods have been increasingly applied to materials science, with some studies attempting to predict the crystal structure or physical properties of materials using machine learning models. However, most existing methods rely on the material's crystal structure information (e.g., CIF files) as input, which typically requires experimental or complex computational methods, leading to high acquisition costs. Existing methods rarely predict complete XRD diffraction patterns directly from the material's chemical formula, thus limiting their application in rapid material screening and virtual experiments. Summary of the Invention
[0003] This invention provides a method for generating diffraction patterns of deep learning materials, the method comprising the following steps:
[0004] Step 1: Obtain the chemical formula of the material;
[0005] Step 2: Based on the chemical formula of the material, obtain the material element attribute characteristics through material element attribute extraction methods, and enhance the characteristics based on the element composition and physicochemical properties. Perform weighted statistics on the physicochemical properties of the elements according to the mole fraction of the elements in the chemical formula, calculate the average value, maximum value, minimum value and variance, and obtain the overall characteristics of the material.
[0006] Step 3: Input the enhanced material features into the pre-trained deep learning model to generate the corresponding XRD diffraction pattern data;
[0007] Step 4: Calculate the prediction uncertainty by performing multiple forward propagation calculations with random deactivation during the inference phase, and output the XRD diffraction pattern data and the corresponding prediction uncertainty.
[0008] Step 5: Post-process the generated XRD diffraction pattern data, including diffraction peak identification, peak position calibration and intensity normalization, to improve the accuracy and usability of the diffraction pattern.
[0009] The deep learning model is used to directly predict XRD diffraction patterns based on chemical formulas, without requiring crystal structure information.
[0010] The deep learning model used in this invention includes a feature extraction module, an adaptive residual connection module, a multi-scale Transformer module, and an output module, which are implemented as follows:
[0011] - Feature extraction module: processes the input material features and performs multi-attribute feature fusion based on elemental composition and physicochemical properties;
[0012] - Multi-scale Transformer module: Models the interaction between elements, models different combinations of elements through different attention heads, improves the accuracy of diffraction peak position prediction, contains 4 encoder layers, each layer uses 8 attention heads;
[0013] - Adaptive Residual Connection Module: Captures a deep representation of material features. It uses learnable weights to weight and fuse the outputs of the main branch and the residual branch, improving the ability to model the relationship between elements. Each residual block contains two one-dimensional convolutional layers with 256 and 512 kernels, respectively.
[0014] - Output module: Generates XRD diffraction pattern data and predicts uncertainty. The output dimension is 1×5250, corresponding to the intensity values of 5250 diffraction angles.
[0015] The material elemental properties are selected from at least one of the following physicochemical properties: elemental composition, atomic number, atomic radius, electronegativity, melting point, number of valence electrons, electron affinity, and first ionization energy.
[0016] The specific implementation of the feature enhancement is as follows: by obtaining the material element embedding vector and element attribute statistical features, and then concatenating the two and performing normalization processing, the final material feature vector is obtained.
[0017] The multi-scale Transformer module is specifically implemented as follows: it includes multiple encoder layers, each employing multiple attention heads to focus on the feature interactions of different element combinations, thereby achieving multi-scale modeling of interactions between elements and improving the predictive ability for complex material systems. Each encoder layer contains multiple self-attention heads, which learn the interaction relationships between elements through attention weights of different dimensions, thus achieving multi-scale feature modeling.
[0018] The adaptive residual connection module is specifically implemented as follows: each residual block contains two convolutional layers and an adaptive skip connection. The residual connection weights are dynamically calculated through an auxiliary neural network to adaptively adjust the residual connection ratio according to the input features, thereby enhancing the model's adaptability to different material types.
[0019] The specific calculation method for the prediction uncertainty is as follows: 20 forward propagations with random deactivation are performed during the inference phase, and the average and standard deviation of the output are calculated. The standard deviation is used as a measure of prediction uncertainty.
[0020] The training process of the pre-trained deep learning model includes: training the deep learning model with training data containing 110,000 material samples and their corresponding XRD diffraction patterns; optimizing the model parameters using a hybrid loss function, which includes mean squared error loss, uncertainty loss and feature consistency loss; using a gradient descent-based optimizer with a learning rate of 0.001 and a batch size of 512.
[0021] In another aspect, the present invention provides a deep learning material diffraction pattern generation apparatus, the apparatus comprising:
[0022] - Input module, used to receive material chemical formulas, with a text input interface;
[0023] - Feature extraction module, used to obtain material element attribute features by means of material element attribute extraction according to the chemical formula of the material, and to enhance the features based on element composition and physicochemical properties. The physicochemical properties of the elements are weighted and statistically analyzed according to the mole fraction of the elements in the chemical formula, and the average value, maximum value, minimum value and variance are calculated to obtain the overall characteristics of the material;
[0024] - The model inference module is used to input the enhanced material features into the pre-trained deep learning model, generate the corresponding XRD diffraction pattern data, and predict the uncertainty by performing multiple forward propagation calculations with random deactivation during the inference stage.
[0025] - The post-processing module is used to perform diffraction peak identification, peak position calibration and intensity normalization on the generated XRD diffraction pattern data;
[0026] - Output module, used to output the processed XRD diffraction pattern data and the corresponding prediction uncertainty, in JSON or CSV format;
[0027] The present invention also provides a computer device, the computer device including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the deep learning material diffraction pattern generation method described above. The program stored in the memory includes a feature extraction algorithm, a model inference algorithm, an uncertainty calculation algorithm, and a post-processing algorithm.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] 1. This invention only uses the chemical formula of the material as input, without requiring crystal structure information, thus reducing the difficulty of data acquisition;
[0030] 2. This invention improves the ability to express material features by fusing element embedding features with element attribute statistical features;
[0031] 3. This invention uses a multi-scale Transformer structure to model the interactions between elements, which improves the accuracy of XRD diffraction peak position prediction;
[0032] 4. This invention improves the reliability of model prediction results by calculating prediction uncertainty through random deactivation inference;
[0033] 5. This invention can quickly generate material XRD patterns, which can be used for material screening and virtual experiments. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in 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.
[0035] Figure 1 This is a flowchart of the steps in a method for generating diffraction patterns of deep learning materials according to Embodiment 1 of the present invention;
[0036] Figure 2 This is a schematic diagram of the structure of the deep learning model provided in Embodiment 1 of the present invention;
[0037] Figure 3 This is a structural block diagram of a deep learning material diffraction pattern generation device provided in Embodiment 2 of the present invention;
[0038] Figure 4 This is a schematic diagram of the structure of the model training module provided in Embodiment 2 of the present invention. Detailed Implementation
[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only 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.
[0040] Please see Figure 1 and Figure 2 The present invention provides a method for generating diffraction patterns of deep learning materials, comprising:
[0041] Step 101: Obtain the chemical formula of the material.
[0042] Step 102: Based on the chemical formula of the material, obtain the material element attribute characteristics through material element attribute extraction methods, and enhance the characteristics based on the element composition and physicochemical properties. Perform weighted statistics on the physicochemical properties of the elements according to the mole fraction of the elements in the chemical formula, calculate the average value, maximum value, minimum value and variance, and obtain the overall characteristics of the material.
[0043] Material elemental properties refer to various characteristics that can describe the properties of a material, selected from at least one of the following physicochemical properties: elemental composition, atomic number, atomic radius, electronegativity, melting point, number of valence electrons, electron affinity, and first ionization energy.
[0044] It should be noted that this embodiment obtains the elemental property characteristics of the material through material elemental property extraction methods, and enhances the features based on elemental composition and physicochemical properties. The physicochemical properties of the elements are weighted and statistically analyzed according to the mole fraction of the elements in the chemical formula, and the average, maximum, minimum and variance are calculated to obtain the overall characteristics of the material. These enhanced features contain rich physicochemical information of the material, providing a comprehensive foundation for the subsequent generation of XRD patterns, while avoiding the circular reasoning problem caused by using structural information.
[0045] Step 103: Input the enhanced material features into the pre-trained deep learning model to generate the corresponding XRD diffraction pattern data.
[0046] The deep learning model refers to the neural network model designed in this embodiment for generating XRD diffraction patterns.
[0047] XRD diffraction pattern data refers to numerical data representing the X-ray diffraction pattern of a material, which typically includes the diffraction angle and the corresponding diffraction intensity.
[0048] Step 104: Calculate the prediction uncertainty by performing multiple forward propagation calculations with random deactivation during the inference phase, and output the XRD diffraction pattern data and the corresponding prediction uncertainty.
[0049] Prediction uncertainty refers to the confidence estimate of the model on the generated XRD diffraction pattern. It is used to evaluate the reliability of the model's prediction by performing multiple forward propagation calculations with random deactivation during the inference phase.
[0050] It should be noted that the deep learning model in this embodiment integrates adaptive residual connections and a multi-scale Transformer architecture, which can effectively capture the deep representation of material features and inter-element interactions, thereby generating accurate XRD diffraction patterns and providing prediction uncertainty estimates without using any structural information.
[0051] Step 105: Post-process the generated XRD diffraction pattern data, including diffraction peak identification, peak position calibration, and intensity normalization, to improve the accuracy and usability of the diffraction pattern.
[0052] It should be noted that the output XRD diffraction pattern data can be used for visualization or further material analysis. Prediction uncertainty helps users assess the reliability of the prediction results; for predictions with high uncertainty, users can choose to conduct experimental verification.
[0053] In one specific embodiment of this example, the deep learning model includes:
[0054] The feature extraction module is used to process the input material features and perform multi-attribute feature fusion based on elemental composition and physicochemical properties.
[0055] The multi-scale Transformer module is used to model the interactions between elements. It models different combinations of elements using different attention heads, thereby improving the accuracy of diffraction peak position prediction.
[0056] The adaptive residual connection module is used to capture a deep representation of material features. It improves the ability to model the combination relationship of elements by weighting and fusing the output of the main branch and the output of the residual branch with learnable weights.
[0057] The output module is used to generate XRD diffraction pattern data and predict uncertainties.
[0058] It should be noted that the deep learning model structure design in this embodiment is as follows: 1. S201 Input: Receives material feature input as the initial input to the model.
[0059] 2. S202 Feature Extraction Module: The module fuses the input material features based on elemental composition and physicochemical properties to generate an enhanced feature representation without using any structural information.
[0060] The specific implementation of feature enhancement is as follows: First, obtain the material element embedding vector based on the chemical formula. For the i-th element in the material... The element embedding vector is defined as follows: .
[0061] in Represents the i-th chemical element in the material. This represents the semantic embedding mapping function of elements. Represents the element embedding vector and , where d is the embedding dimension. To learn the semantic relationships between elements, this is achieved by maximizing the co-occurrence probability of an element with its context, specifically as follows:
[0062] Where N represents the number of element samples, and C(i) represents the number of elements. The collection of context elements This represents conditional probability. The specific definition of conditional probability between elements is:
[0063] Where M represents the total number of elements. Represents element Embedded vector, This indicates the embedding of context elements. For material chemical formulas:
[0064] The method for obtaining the overall characteristics of a material is to assume that the number of moles of each element is 1. The material embedding vector is:
[0065] Where the weights are: .
[0066] To enhance the representation of material features, statistical calculations were performed on the element embeddings, yielding the following features: Mean characteristics:
[0067] Variance characteristics:
[0068] Maximum value characteristics:
[0069] Minimum value characteristics:
[0070] The final material representation is obtained by combining the above statistical characteristics:
[0071] in:
[0072] Secondly, obtain the statistical characteristics of elemental properties, assuming the material's chemical formula is:
[0073] in This represents the i-th element. This represents the number of atoms of an element in its chemical formula, and k represents the number of different types of elements contained in the material. Let the elemental property be p(e... i ), where p(⋅) represents a certain physicochemical property, such as atomic number, atomic radius, electronegativity, number of valence electrons, melting point, etc. Element weights are defined as...
[0074] in Represents element The proportion in the material. The average value of the material properties.
[0075] This formula represents the overall material property as: Overall Material Property = Element Property × Element Proportion. The maximum value of the property is:
[0076] The minimum value of the attribute is:
[0077] The attribute range is:
[0078] The standard deviation of the attribute is:
[0079] For m types of element attributes, calculate the statistical characteristics:
[0080] The final material feature vector is:
[0081] Where d is the feature dimension. Finally, the two sets of features above are concatenated into a single feature vector, and min-max normalization is performed. The specific formula is as follows:
[0082] in This represents the concatenated feature vector. This represents the minimum value of the concatenated eigenvectors. This represents the maximum value of the concatenated eigenvectors. This represents the final eigenvector after normalization.
[0083] 3. 203 Multi-scale Transformer Module: Composed of multiple encoder layers, each layer employs multiple attention heads to focus on the feature interactions of different element combinations, thereby achieving multi-scale modeling of interactions between elements and improving the prediction capability for complex material systems.
[0084] 4. 204 Adaptive Residual Connection Module: Composed of multiple residual blocks, each containing two convolutional layers and adaptive skip connections. The residual connection weights are dynamically calculated through an auxiliary neural network, enabling adaptive adjustment of the residual connection ratio based on input features, thereby enhancing the model's adaptability to different material types.
[0085] The adaptive residual connection module is implemented as follows: each residual block contains two convolutional layers and adaptive skip connections. The residual connection weights are dynamically calculated using an auxiliary neural network, enabling adaptive adjustment of the residual connection ratio based on input features, thus enhancing the model's adaptability to different material types. The specific formula is as follows: Let the input of the residual block be x, and the main branch output be... The residual connection weights are The output of the residual block is then:
[0086] in Calculated using an auxiliary neural network:
[0087] Where FC1 represents the first fully connected layer with an input dimension of 256 and an output dimension of 128; ReLU represents the activation function; FC2 represents the second fully connected layer with an input dimension of 128 and an output dimension of 1; σ represents the sigmoid function, which is used to restrict the output to the range [0,1].
[0088] The multi-scale Transformer module is implemented by comprising multiple encoder layers, each employing multiple attention heads to focus on the feature interactions of different element combinations. This enables multi-scale modeling of inter-element interactions, improving the predictive ability for complex material systems. The specific formula is as follows: Let the input feature be X ∈ R^{L×d}, where L is the sequence length and d is the feature dimension. The calculation process of the multi-scale attention mechanism is as follows: First, calculate the query, key, and value matrix:
[0089] Then, the attention weights are calculated. For the m-th attention head, the window size is... Attention weights are:
[0090] in , Let these represent the query and key matrices of the m-th attention head, respectively. The dimension of the key is represented; then the attention output is calculated:
[0091] Then concatenate the outputs of multiple attention heads:
[0092] Where H represents the number of attention heads. This represents the output projection matrix; finally, the output of the multi-head attention layer is obtained:
[0093] LayerNorm represents the layer normalization operation.
[0094] 5.205 Output Module: Maps the processed features to the dimensions of the XRD diffraction pattern, generates the final XRD diffraction pattern data, and evaluates the reliability of the prediction results by performing multiple forward propagation calculations with random deactivation during the inference phase to calculate the uncertainty.
[0095] The specific method for calculating prediction uncertainty is as follows: During the inference phase, perform 20 forward propagations with random deactivation, and calculate the mean and standard deviation of the outputs. The standard deviation is used as a measure of prediction uncertainty. The specific formula is as follows: Let the model be f_θ, where θ are the model parameters. During the inference phase, keep random deactivation enabled. For the input feature x, perform T = 20 forward propagations to obtain the output set:
[0096] in
[0097] Calculate the predicted mean:
[0098] Calculate the forecast uncertainty (standard deviation):
[0099] Where σ represents the prediction uncertainty, the larger the value, the lower the reliability of the prediction result.
[0100] 6.206 Output: Outputs XRD diffraction pattern data and prediction uncertainty.
[0101] Embodiment 1 of the present invention also provides a deep learning material diffraction pattern generation device, the device comprising: 1. 301 Input Module, used to receive material chemical formula input, with text input as the interface type.
[0102] 2. The 302 feature extraction module, connected to the 301 input module, is used to extract element embedding features and element physicochemical property features, without using structural information. Specifically, it obtains the material element embedding vector and element property statistical features based on the input material chemical formula, concatenates the two, and then normalizes them to obtain the final material feature vector.
[0103] 3. A 303 model inference module, connected to the 302 feature extraction module, is used to perform inference using a deep learning model, inputting the enhanced material features into a pre-trained deep learning model to generate corresponding XRD diffraction pattern data and calculate prediction uncertainty.
[0104] 4. A 304 post-processing module, connected to the 303 model inference module, is used to perform diffraction peak identification, peak position calibration, and intensity normalization processing to improve the accuracy and usability of the diffraction pattern.
[0105] 5. 305 Output Module, connected to the 304 Post-processing Module, is used to output the processed XRD pattern data and prediction uncertainty, with the output format being JSON or CSV.
[0106] The device operates as follows: First, the input chemical formula of the material is received by the user through the input module 301. Then, the input chemical formula is processed by the feature extraction module 302 to extract material features. Next, the XRD diffraction pattern is generated using a deep learning model through the model inference module 303. Then, the generated pattern is optimized by the post-processing module 304. Specifically, this includes: diffraction peak identification using a local extremum detection method, which identifies possible diffraction peak positions by calculating the first and second derivatives of the pattern; to improve the physical consistency of the predicted pattern, local extrema are identified through a diffraction peak detection algorithm, and the peak positions are fine-tuned in combination with physical constraints, thereby improving the consistency between the predicted pattern and the actual experimental pattern; peak position calibration uses a standard sample-based calibration method, which establishes a calibration model to correct the peak positions by comparing the deviation between the predicted peak positions and the standard peak positions; intensity normalization uses a maximum intensity normalization method, which divides the intensity of all diffraction peaks by the intensity of the maximum peak, so that the patterns are compared under the same intensity scale; finally, the final XRD pattern data and prediction uncertainty are output through the output module 305.
[0107] Embodiment 2 of the present invention provides a training method for a deep learning material diffraction pattern generation model. The method first creates a dataset. Specifically, the 401 data collection unit obtains the CIF files of materials from a material database. The material data comes from public material databases, including the Materials Project database, etc. Then, the 402 Bragg diffraction peak calculation unit calculates the XRD pattern based on the CIF crystal structure. The XRD pattern is obtained by calculating the Bragg diffraction peaks of the crystal structure and constructing a continuous diffraction pattern.
[0108] Specifically, the Bragg diffraction peak positions are first calculated using the 402 Bragg diffraction peak calculation unit. For the crystal plane (hkl), the diffraction peak positions satisfy Bragg's law:
[0109] Where λ is the X-ray wavelength (Cu Kα, approximately 1.5406 Å). Let be the interplanar spacing, θ be the diffraction angle, and n be the diffraction order. The diffraction peaks are calculated using the crystal structure. ,in Let be the position of the i-th diffraction peak. This corresponds to the diffraction intensity.
[0110] Continuous XRD patterns were constructed using Gaussian peak functions through the 403 continuous pattern building unit, with each Bragg peak represented by a Gaussian function.
[0111] in This is the peak width parameter. The entire XRD pattern is obtained by superimposing all diffraction peaks:
[0112] The peak width is calculated using a 404 peak width calculation unit. The peak width is determined by the grain size, according to the Scherrer equation:
[0113] Where β is the full width at half maximum (FWHM), K is the shape factor (usually taken as 0.9), L is the grain size, and λ is the X-ray wavelength. The Gaussian distribution width is:
[0114] Finally, the generated continuous spectrum is uniformly sampled in the range 2θ∈[10°, 90°] using a 405 XRD spectrum sampling unit to obtain...
[0115] Where n=5250, the one-dimensional XRD vector required for model training is obtained.
[0116] After completing the dataset creation, model training is performed. For each material sample, the 406 feature extraction unit extracts elemental property features based on its chemical formula and performs feature enhancement as described in Example 1 to obtain the material feature vector. Then, the 407 model training unit optimizes the model parameters using a hybrid loss function, which includes mean squared error loss, uncertainty loss, and feature consistency loss. The specific formula is as follows: Hybrid Loss Function:
[0117] 1. Mean squared error loss:
[0118] Where yi is the true XRD pattern intensity, ŷi is the predicted XRD pattern intensity, and N is the number of samples.
[0119] 2. Losses due to uncertainty:
[0120] in To predict uncertainty.
[0121] 3. Feature consistency loss:
[0122] in and This represents the features of different layers.
[0123] The optimizer uses a gradient descent-based optimizer with a learning rate of 0.001 and a batch size of 512. Then, the model performance is evaluated using multiple metrics through the 408 model evaluation unit, including mean squared error, peak accuracy, and uncertainty calibration. Finally, the trained model is saved in the 409 unit for subsequent inference and prediction.
[0124] It is important to note that the quality and quantity of training data are crucial to model performance; therefore, it is necessary to collect a sufficient number of material samples and their corresponding XRD diffraction patterns. Furthermore, an early stopping strategy can be employed during model training, stopping training when the validation set performance no longer improves to prevent overfitting.
Claims
1. A method for generating diffraction patterns of deep learning materials, characterized in that, include: Obtain the chemical formula of the material; Based on the chemical formula of the material, the material element attribute features are obtained through material element attribute extraction methods, and feature enhancement is performed based on element composition and physicochemical properties. The physicochemical properties of the elements are weighted and statistically analyzed according to the element attributes in the chemical formula, and the average, maximum, minimum and variance are calculated. In addition, element embedding vector information is added to obtain the overall characteristics of the material. The enhanced material features are input into a pre-trained deep learning model to generate corresponding XRD diffraction pattern data; By performing multiple forward propagation calculations with random deactivation during the inference phase, the prediction uncertainty is calculated, and the XRD diffraction pattern data and the corresponding prediction uncertainty are output. The deep learning model is used to directly predict XRD diffraction patterns based on chemical formulas, without requiring crystal structure information, thus enabling the prediction of XRD diffraction patterns without crystal structure information.
2. The method for generating diffraction patterns of deep learning materials according to claim 1, characterized in that, The material elemental properties are selected from at least one of the following physicochemical properties: elemental composition, atomic number, atomic radius, electronegativity, melting point, number of valence electrons, electron affinity, and first ionization energy.
3. The method for generating diffraction patterns of deep learning materials according to claim 1, characterized in that, The deep learning model includes: The feature extraction module is used to process the input material features and perform multi-attribute feature fusion based on elemental composition and physicochemical properties; The multi-scale Transformer module is used to model the interactions between elements. It models different combinations of elements using different attention heads, thereby improving the accuracy of diffraction peak position prediction. The adaptive residual connection module is used to capture a deep representation of material features. It uses learnable weights to weight and fuse the outputs of the main branch and the residual branch, thereby improving the ability to model the relationship between elements. The output module is used to generate XRD diffraction pattern data and predict uncertainties.
4. The method for generating diffraction patterns of deep learning materials according to claim 3, characterized in that, The adaptive residual connection module includes multiple residual blocks, each containing two convolutional layers and an adaptive skip connection. The residual connection weights are dynamically calculated through an auxiliary neural network, enabling adaptive adjustment of the residual connection ratio based on input features and enhancing the model's adaptability to different material types.
5. The method for generating diffraction patterns of deep learning materials according to claim 3, characterized in that, The multi-scale Transformer module includes multiple encoder layers, each employing multiple attention heads to focus on the feature interactions of different element combinations, thereby achieving multi-scale modeling of interactions between elements and improving the prediction capability for complex material systems.
6. The method for generating diffraction patterns of deep learning materials according to claim 1, characterized in that, Also includes: The generated XRD diffraction pattern data is post-processed, including diffraction peak identification, peak position calibration, and intensity normalization, to improve the accuracy and usability of the diffraction pattern.
7. The method for generating diffraction patterns of deep learning materials according to claim 1, characterized in that, The training process of the pre-trained deep learning model includes: The deep learning model is trained using training data containing material samples and their corresponding XRD diffraction patterns. The model parameters are optimized using a hybrid loss function, which includes mean squared error loss, uncertainty loss, and feature consistency loss. The optimizer is a gradient descent-based optimizer.
8. A device for generating diffraction patterns of deep learning materials, characterized in that, include: The input module is used to receive the chemical formula of the material. The feature extraction module is used to obtain the material element attribute features by means of material element attribute extraction according to the chemical formula of the material, and to enhance the features based on the element composition and physicochemical properties. The physicochemical properties of the elements are weighted and statistically analyzed according to the mole fraction of the elements in the chemical formula, and the average value, maximum value, minimum value and variance are calculated to obtain the overall characteristics of the material. The model inference module is used to input the enhanced material features into a pre-trained deep learning model, generate corresponding XRD diffraction pattern data, and predict uncertainty by performing multiple forward propagation calculations with random deactivation during the inference stage. The post-processing module is used to perform diffraction peak identification, peak position calibration, and intensity normalization on the generated XRD diffraction pattern data. The output module is used to output the processed XRD diffraction pattern data and the corresponding prediction uncertainty; The device is used to perform the deep learning material diffraction pattern generation method according to claim 1.
9. The deep learning material diffraction pattern generation device according to claim 8, characterized in that, The deep learning model adopts the deep learning model structure as described in any one of claims 3-5.
10. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the deep learning material diffraction pattern generation method as described in any one of claims 1-7.