A method of crystal property prediction
By combining Fourier analysis networks and Transformer encoders, a pseudo-periodic prediction model was established, which solved the problem of fuzzy extraction of periodic features in crystal property prediction, and achieved efficient and accurate crystal property prediction, improving the accuracy and efficiency of the prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUNAN UNIV
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies for predicting crystal properties suffer from high costs, long cycles, and difficulty in high-throughput screening. Furthermore, deep learning-based models fail to effectively encode lattice periodic arrangements, making them black boxes with poor interpretability and large prediction errors, which reduces the reliability and efficiency of the prediction results.
By explicitly encoding the periodicity of the crystal using a Fourier analysis network (FAN layer), and combining it with Transformer encoder pre-training and loss feature adjustment, a pseudo-periodic prediction model is established to replace the traditional attention mechanism, capture long-range interactions of the crystal, and achieve accurate correction of the prediction results through multi-index verification and error correlation compensation.
This method solves the problem of fuzzy extraction of periodic features in traditional methods, improves the accuracy and efficiency of prediction results, simplifies data processing steps, and reduces prediction errors.
Smart Images
Figure CN122494019A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of crystal property prediction technology, and in particular to a method for predicting crystal properties. Background Technology
[0002] The mechanical, optical, electrical, and thermodynamic properties of crystalline materials are highly correlated with their atomic arrangement, lattice structure, and periodicity. Traditional methods for obtaining crystal properties mainly rely on first-principles calculations and experimental characterization, which suffer from high costs, long cycles, and difficulty in high-throughput screening. With the development of artificial intelligence, deep learning-based crystal property prediction has gradually become a research hotspot. However, existing technologies have significant shortcomings: they do not effectively encode the periodic arrangement of the lattice; the models are black boxes with poor interpretability and cannot correlate structure and properties; and the index parameters in the prediction results contain errors that have not been further verified, reducing the reliability of the prediction results and the efficiency of the entire prediction process. Summary of the Invention
[0003] To overcome the shortcomings of the prior art, this application provides a method for predicting crystal properties.
[0004] This application provides a method for predicting crystal properties, the method comprising:
[0005] Step S1: Obtain the original crystal structure data, normalize the original crystal structure data to obtain the crystal feature vector, obtain the encoded atomic-level input embedding vector based on the Fourier analysis network FAN layer and the crystal feature vector, train the Transformer encoder according to historical crystal structure data, preset the simulated atomic-level input embedding vector, input the simulated atomic-level input embedding vector into the trained Transformer encoder for feature learning, and obtain the preprocessing loss feature value of the atomic coordinates, atomic classification and lattice parameters. According to the preprocessing loss feature value and the atomic-level input embedding vector to be tested, obtain the pre-trained periodic embedding.
[0006] Step S2: Based on the physical laws of crystal periodicity in the original crystal structure data, establish a pseudo-period prediction model. Based on the pre-trained periodic embedding and pseudo-period prediction model, combined with feedforward network processing, obtain the initial prediction results of crystal properties.
[0007] Step S3: Compare the ratio between bulk modulus and shear modulus in the initial crystal property prediction results with the preset Poisson's ratio; compare the formation energy in the initial crystal property prediction results with the preset energy comparison threshold one; match the band gap value with the preset crystal structure type to be matched; compare the total energy with the preset energy comparison threshold two; and compare the electron affinity energy with the preset energy comparison threshold three. A result set is obtained by combining these results. Based on the test error between the predicted values of all indicators in the initial crystal property prediction results and the historical data reference values in the result set, error compensation is performed on the predicted values of all indicators in the initial crystal property prediction results, and the final crystal property prediction result is output.
[0008] Preferably, the original crystal structure data is obtained, and the atomic type, atomic fraction coordinates and lattice parameters in the original crystal structure data are normalized to obtain normalized feature information. The atomic features, coordinate features and lattice parameters in the normalized feature information are concatenated to obtain the crystal feature vector.
[0009] Perform a linear projection transformation on the crystal feature vector to obtain the projected features;
[0010] Based on the FAN layer of the Fourier analysis network, the periodic crystal structure features in the projection features are explicitly encoded, and the atomic-level input embedding vector to be tested is output.
[0011] Historical crystal structure data is acquired, and the Transformer encoder is pre-trained based on the general spatial distribution and periodic structure characteristics of the crystals in the historical crystal structure data to obtain a trained Transformer encoder.
[0012] Preferably, based on the atomic-level input embedding vector to be tested, a simulated atomic-level input embedding vector is preset, and the simulated atomic-level input embedding vector is input into the trained Transformer encoder for feature learning to obtain the predicted atom position, the pre-determined atom type, and the predicted lattice parameters;
[0013] The difference between the predicted atomic positions and the actual atomic positions in the historical crystal structure data is used to obtain the coordinate reconstruction loss feature value.
[0014] The predicted atom type is compared with the actual atom type in the historical crystal structure data, and the number of incorrect judgments is counted to obtain the atom classification loss feature value.
[0015] The difference between the predicted lattice parameters and the actual lattice parameters in the historical crystal structure data is used to obtain the lattice parameter loss characteristic value.
[0016] The coordinate reconstruction loss feature value, the atom classification loss feature value, and the lattice parameter loss feature value are combined to form the preprocessing loss feature value.
[0017] Preferably, the trained Transformer encoder is adjusted based on the preprocessed loss feature values to obtain the adjusted Transformer encoder;
[0018] Based on the adjusted Transformer encoder, which employs a combination of multi-head attention, layer normalization, and feedforward network structure, feature learning is performed on the atomic-level input embedding vector to extract the spatial arrangement, bonding relationships, and periodic general features of atoms in the crystal, and output a pre-trained periodic embedding with common laws of crystal structure.
[0019] Preferably, based on the physical laws of crystal periodicity in the original crystal structure data, a cell period offset term and a periodic mirror mapping rule are preset. Based on the cell period offset term, the periodic mirror mapping rule, and historical crystal structure data, a pseudo-finite periodic attention mechanism is trained. The traditional multi-head attention inside the pseudo-trained Transformer encoder is replaced with the pseudo-finite periodic attention mechanism to establish a pseudo-periodic prediction model.
[0020] The pre-trained period is embedded into the input pseudo-period prediction model to extract long-range crystal interaction features. The long-range crystal interaction features are normalized and processed by a feedforward network to obtain initial prediction results of crystal properties. The initial prediction results of crystal properties include bulk modulus, shear modulus, formation energy, band gap value, total energy, and electron affinity.
[0021] Preferably, the preset Poisson's ratio, energy comparison threshold one, crystal structure type to be matched, energy comparison threshold two, and energy comparison threshold three are used.
[0022] The ratio of the bulk modulus to the shear modulus is compared with Poisson's ratio to obtain judgment result one; the formation energy is compared with the energy to threshold one to obtain judgment result two; the band gap value is matched with the crystal structure type to be matched to obtain judgment result three; the total energy is compared with the energy to threshold two to obtain judgment result four; and the electron affinity energy is compared with the energy to threshold three to obtain judgment result five.
[0023] The judgment result 1, judgment result 2, judgment result 3, judgment result 4 and judgment result 5 are combined to form a result set;
[0024] Historical data reference values are obtained, and the predicted values and historical data reference values of all indicators in the initial prediction results of crystal properties are statistically analyzed to obtain the error to be judged. An error division threshold is preset. If the judgment results of all indicators in the result set are in compliance and the error to be judged is less than the error division threshold, then the first case feature is output.
[0025] Based on characteristic one of the aforementioned situations, the initial prediction result of the crystal properties is determined to be the final prediction result, and the verification result is output.
[0026] Preferably, if any result in the result set is non-compliant, or any error in the error to be judged is within the error division threshold, then output characteristic two.
[0027] The remaining case features after removing case features one and case features two from the result set and the error to be judged are determined to be case feature three. Based on case feature three, the original crystal structure data is re-acquired and analyzed, and secondary data processing information is output.
[0028] Based on characteristic two of the situation, and according to the correlation between the indicators, the indicators in the initial prediction results of crystal properties are grouped into groups based on the correlation between their errors, thus obtaining the correlated error groups;
[0029] The ratio of the measured error to the historical data reference value is used to obtain the preprocessing error coefficient. Based on the preprocessing error coefficient and the measured error, error compensation is performed on the predicted values of all indicators in the initial prediction results of crystal properties, and the final prediction results of crystal properties are output.
[0030] Compared with the prior art, the present invention has the following characteristics and beneficial effects:
[0031] By normalizing the crystal and explicitly encoding its periodicity using the FAN layer, combined with Transformer encoder pre-training and loss feature adjustment, a pre-trained periodic embedding that combines common patterns with the characteristics of the crystal under test is obtained. This solves the problem of fuzzy periodic feature extraction in traditional methods. Based on the physical laws of crystal periodicity, a pseudo-periodic prediction model is constructed to replace the traditional attention mechanism, effectively capturing long-range crystal interactions. Through multi-index verification and error correlation compensation, accurate correction of prediction results is achieved, avoiding the limitations of single-index judgment. For cases where the prediction does not meet the threshold or the error exceeds the standard, the prediction value is dynamically corrected through index correlation grouping and error coefficient compensation, significantly reducing prediction error and improving the accuracy of the results. The normalization process and automated error compensation process simplify data processing steps and improve prediction efficiency. Attached Figure Description
[0032] Figure 1 This is a flowchart illustrating the steps of a crystal property prediction method, which is the main feature of this embodiment. Detailed Implementation
[0033] The present invention will be further described in detail below with reference to the following embodiments.
[0034] Reference Figure 1 A method for predicting crystal properties, comprising the following steps:
[0035] Step S1: Obtain the original crystal structure data, normalize the original crystal structure data to obtain the crystal feature vector, obtain the encoded atomic-level input embedding vector based on the Fourier analysis network FAN layer and the crystal feature vector, train the Transformer encoder according to historical crystal structure data, preset the simulated atomic-level input embedding vector, input the simulated atomic-level input embedding vector into the trained Transformer encoder for feature learning, and obtain the preprocessing loss feature value of the atomic coordinates, atomic classification and lattice parameters. Based on the preprocessing loss feature value and the atomic-level input embedding vector to be tested, obtain the pre-trained periodic embedding.
[0036] Step S2: Based on the physical laws of crystal periodicity in the original crystal structure data, a pseudo-period prediction model is established. Based on the pre-trained periodic embedding and pseudo-period prediction model, combined with feedforward network processing, the initial prediction results of crystal properties are obtained.
[0037] Step S3: Compare the ratio between bulk modulus and shear modulus in the initial crystal property prediction results with the preset Poisson's ratio; compare the formation energy in the initial crystal property prediction results with the preset energy comparison threshold one; match the band gap value with the preset crystal structure type to be matched; compare the total energy with the preset energy comparison threshold two; and compare the electron affinity energy with the preset energy comparison threshold three. A result set is obtained by combining these results. Based on the test error between the predicted values of all indicators in the initial crystal property prediction results and the historical data reference values in the result set and the initial crystal property prediction results, error compensation is performed on the predicted values of all indicators in the initial crystal property prediction results, and the final crystal property prediction results are output.
[0038] Specifically, by normalizing and explicitly encoding crystal periodicity using the FAN layer, combined with Transformer encoder pre-training and loss feature adjustment, a pre-trained periodic embedding that combines common patterns with the characteristics of the crystal under test is obtained. This solves the problem of fuzzy periodic feature extraction in traditional methods. Based on the physical laws of crystal periodicity, a pseudo-periodic prediction model is constructed to replace the traditional attention mechanism, effectively capturing long-range crystal interactions. Through multi-index verification and error correlation compensation, accurate correction of prediction results is achieved, avoiding the limitations of single-index judgment. For cases where the prediction does not meet the threshold or the error exceeds the standard, the prediction value is dynamically corrected through index correlation grouping and error coefficient compensation, significantly reducing prediction error and improving the accuracy of the results. Normalization and automated error compensation processes simplify data processing steps and improve prediction efficiency.
[0039] The specific step S1 includes the following sub-steps:
[0040] The original crystal structure data is obtained, and the atomic types, atomic fraction coordinates, and lattice parameters in the original crystal structure data are normalized to obtain normalized feature information. The atomic features, coordinate features, and lattice parameters in the normalized feature information are concatenated to obtain the crystal feature vector.
[0041] The projected features are obtained by performing a linear projection transformation on the crystal eigenvectors.
[0042] Based on the FAN layer of the Fourier analysis network, the periodic crystal structure features in the projection features are explicitly encoded, and the atomic-level input embedding vector to be tested is output.
[0043] Historical crystal structure data is obtained, and the Transformer encoder is pre-trained based on the general spatial distribution and periodic structure characteristics of crystals in the historical crystal structure data to obtain a trained Transformer encoder.
[0044] Based on the atomic-level input embedding vector to be tested, a simulated atomic-level input embedding vector is preset. The simulated atomic-level input embedding vector is then input into the trained Transformer encoder for feature learning to obtain the predicted atomic position, the pre-determined atomic type, and the predicted lattice parameters.
[0045] The coordinate reconstruction loss eigenvalue is obtained by subtracting the predicted atomic positions from the actual atomic positions in historical crystal structure data.
[0046] The predicted atom type is compared with the actual atom type in historical crystal structure data. The number of incorrect judgments is counted to obtain the atom classification loss feature value.
[0047] The difference between the predicted lattice parameters and the actual lattice parameters in historical crystal structure data is used to obtain the lattice parameter loss eigenvalue.
[0048] The coordinate reconstruction loss feature value, the atom classification loss feature value, and the lattice parameter loss feature value are combined to form the preprocessing loss feature value.
[0049] Based on the preprocessed loss feature values, the trained Transformer encoder is adjusted to obtain the adjusted Transformer encoder.
[0050] The adjusted Transformer encoder employs a combination of multi-head attention, layer normalization, and feedforward network structures to perform feature learning on atomic-level input embedding vectors. This extracts the spatial arrangement, bonding relationships, and periodic general features of atoms in the crystal, outputting pre-trained periodic embeddings that possess the common laws of crystal structure.
[0051] Specifically, this includes crystal feature vectors (e.g., normalizing the original crystal structure data (e.g., NaCl crystal), converting atom type (characterized by atomic number), atomic fractional coordinates (normalized to the [0,1] interval), and lattice parameters (normalized to the unit cell scale) into normalized feature information; then concatenating the atomic features (normalized atomic number values), coordinate features (normalized fractional coordinates), and lattice parameters (normalized 5.64 Å) of Na and Cl atoms to obtain a crystal feature vector with unified dimensions), and projection features (e.g., performing a linear projection transformation on the obtained crystal feature vector to map it to a feature space of a preset dimension (assuming we preset...). The feature space dimension is 64 dimensions (the preset dimension can be adjusted according to actual needs; common dimensions are 32, 64, and 128 dimensions; 64 dimensions are used as an example here) to obtain the projection features; based on the Fourier analysis network (FAN layer), the periodic structural features of NaCl crystal in the projection features (such as the periodicity of alternating Na and Cl atoms and the spatial periodicity of the cubic crystal system) are explicitly encoded, and finally the atomic-level input embedding vector of the NaCl crystal to be tested is output, which can accurately characterize the periodic distribution features of crystal atoms). The Transformer encoder is trained (obtaining the historical crystal structure of various common cubic crystal systems (such as KCl, LiCl, etc.)). The data is constructed to extract the general spatial distribution (cubic crystal system atomic arrangement rules) and periodic structure features of these crystals. Using this as training samples, a Transformer encoder is pre-trained to obtain a well-trained Transformer encoder capable of accurately extracting the periodic features of cubic crystals. Simulated atomic-level input embedding vectors are preset (based on the atomic-level input embedding vectors of NaCl to be tested, a set of simulated atomic-level input embedding vectors is preset (simulating the atomic distribution of NaCl under different lattice constants)). For example, this simulated vector is input into the trained Transformer encoder for feature learning, outputting predicted atomic positions (e.g., predicting Na atoms). The coordinates are (0.02, 0.01, 0), the atom type is pre-determined (the predicted atom corresponding to some coordinates is Na or Cl), the lattice parameters are predicted (e.g., predicting a=b=c=5.66Å), the preprocessing loss feature value is (e.g., the difference between the predicted atom position and the actual NaCl atom position (0,0,0) in historical data is obtained to get the coordinate reconstruction loss feature value (e.g., average deviation 0.015Å); the predicted atom type is compared with the actual atom type, and the number of judgment errors is counted (assuming 1 error) to get the atom classification loss feature value; the difference between the predicted lattice parameter (5.66Å) and the actual lattice parameter (5.64Å) is obtained to get the lattice parameter loss feature value (0.02, 0.01, 0)).The combination of these three factors (0.02Å) constitutes the preprocessing loss feature value of NaCl crystals. This is used to adjust the trained Transformer encoder (adjusting the internal parameters of the Transformer encoder based on the preprocessing loss feature value, optimizing the multi-head attention weights to obtain the adjusted Transformer encoder). For example, the atom classification loss feature value is "1 incorrect judgment" (assuming the error occurs at a Cl atom coordinate, and the model misclassifies it as a Na atom), the coordinate reconstruction loss feature value is 0.015Å (small bias), and the lattice parameter loss feature value is 0.02Å (small bias), indicating that the model's accuracy in recognizing "atom type" is insufficient, but its predictions of "atom coordinates" and "lattice parameters" are relatively accurate. Based on this, the multi-head attention weights of the Transformer encoder are adjusted: assuming the encoder originally had 3 attention heads, corresponding to "atom type feature," "atom coordinate feature," and "lattice parameter feature," with initial weights of 1 / 3 (equal attention). Due to the high atom classification loss... We increased the attention head weight for "atomic type features" from 1 / 3 to 0.5, while decreasing the weight for "atomic coordinate features" and "lattice parameter features" from 1 / 3 to 0.25. After this weight adjustment, the Transformer encoder will focus on atomic type-related features (such as the difference in atomic number and electronic structure between Na and Cl atoms) during subsequent feature learning, reducing excessive attention to coordinates and lattice parameters, thereby correcting the bias of "atomic classification errors." Using this adjusted encoder, we perform feature learning on the atomic-level input embedding vector of NaCl, extracting atomic spatial arrangement (alternating Na and Cl), bonding relationships (ionic bonds), and periodic general features (cubic crystal system periodicity). The final output is a pre-trained periodic embedding with the common laws of cubic crystal structure, providing accurate feature support for subsequent crystal property prediction.
[0052] The specific step S2 includes the following sub-steps:
[0053] Based on the periodic physical laws of crystals in the original crystal structure data, a pseudo-finite periodic attention mechanism is trained by pre-setting the cell periodic offset term and the periodic mirror mapping rule, and using the cell periodic offset term, the periodic mirror mapping rule and historical crystal structure data. The traditional multi-head attention inside the pseudo-trained Transformer encoder is replaced with the pseudo-finite periodic attention mechanism to establish a pseudo-periodic prediction model.
[0054] The pre-trained period is embedded into the input pseudo-period prediction model to extract the long-range interaction features of the crystal. The long-range interaction features of the crystal are normalized and processed by a feedforward network to obtain the initial prediction results of the crystal properties, including bulk modulus, shear modulus, formation energy, band gap value, total energy and electron affinity.
[0055] Specifically, training a pseudo-finite-periodic attention mechanism involves, for example, collecting historical structural data of cubic crystals (NaCl, KCl, LiCl, etc.) and setting predefined cell period offset terms (Δx, Δy, Δz): based on the symmetry of the cubic crystal system, three sets of offset directions are preset (corresponding to the x, y, and z axes respectively), each set of offset directions includes both positive and negative offsets. The periodic mirror mapping rule is based on the symmetry of the NaCl cubic crystal system, with a preset "symmetry mirror mapping rule": for the x-axis direction, if the atom's x-coordinate > 1 (e.g., 1.02), then the mapped x-coordinate = If the x-coordinate is -1 (i.e., 0.02), and the x-coordinate is less than 0 (e.g., -0.03), then the mapped x-coordinate is equal to x-coordinate +1 (i.e., 0.97). The mirror mapping rules for the y-axis and z-axis are completely consistent with those for the x-axis, ensuring the mirror symmetry of the unit cell along the three axes and restoring the infinitely repeating characteristics of the crystal. Combined with historical structural data, this serves as a training sample to train the "pseudo-finite-periodic attention mechanism," enabling it to "recognize" the periodicity of NaCl crystals and accurately capture the "long-range interactions between atoms in adjacent unit cells." Unlike traditional multi-head attention mechanisms (which only focus on atomic interactions within a single unit cell), the pseudo-finite-periodic attention mechanism can associate atomic features of adjacent unit cells through periodic shifts and mirror mapping, capturing long-range interactions (e.g., in NaCl crystals, Cl atoms three unit cells away from the current Na atom). The initial prediction results of crystal properties are obtained by extracting long-range interaction features of NaCl crystals through a pseudo-finite periodic attention mechanism: capturing long-range features such as "ionic bond interaction between Na atoms and Cl atoms in non-adjacent unit cells" and "changes in interatomic forces during the repetition of lattice constants" in NaCl crystals (e.g., the interaction strength and interaction distance between Na atoms in the current unit cell and Cl atoms in 1, 2, and 3 adjacent unit cells). The extracted long-range interaction features of the crystal are normalized (the feature values are normalized to the [0,1] interval to avoid excessive differences in feature values affecting the prediction), and then input into a feedforward network (to further filter and integrate the long-range interaction features of the crystal, strengthen the core features, and weaken irrelevant interference features), and finally output the initial prediction results of the properties of NaCl crystals.
[0056] The specific step S3 includes the following sub-steps:
[0057] The preset Poisson's ratio, energy comparison threshold one, crystal structure type to be matched, energy comparison threshold two, and energy comparison threshold three are used.
[0058] The ratio of bulk modulus to shear modulus is compared with Poisson's ratio to obtain judgment result one. The formation energy is compared with the energy to threshold one to obtain judgment result two. The band gap value is matched with the crystal structure type to be matched to obtain judgment result three. The total energy is compared with the energy to threshold two to obtain judgment result four. The electron affinity energy is compared with the energy to threshold three to obtain judgment result five.
[0059] The result set is composed of judgment result 1, judgment result 2, judgment result 3, judgment result 4, and judgment result 5.
[0060] Obtain historical data reference values, statistically analyze the predicted values and historical data reference values of all indicators in the initial prediction results of crystal properties to obtain the judgment error, preset the error division threshold, and output situation feature one if the judgment results of all indicators in the result set are in compliance and the judgment error is less than the error division threshold.
[0061] Based on characteristic one, the initial prediction result of the crystal properties is determined to be the final prediction result, and the verification result is output.
[0062] If any result in the result set is not compliant, or if any error in the error to be judged is within the error division threshold, then output characteristic two.
[0063] After removing case features 1 and case features 2 from the result set and the error to be judged, the remaining case features are determined to be case feature 3. Based on case feature 3, the original crystal structure data is re-acquired and analyzed, and the data is output for secondary processing information.
[0064] Based on characteristic two, and according to the correlation between the indicators, the indicators in the initial prediction results of crystal properties are grouped into groups based on the mutual correlation of errors, resulting in correlated error groups.
[0065] The ratio of the measured error to the historical data reference value is used to obtain the preprocessing error coefficient. Based on the preprocessing error coefficient and the measured error, error compensation is performed on the predicted values of all indicators in the initial prediction results of crystal properties, and the final prediction results of crystal properties are output.
[0066] Specifically, the preset Poisson's ratio is set to 0.25 based on the actual characteristics of NaCl crystals (ionic crystals, cubic crystal system). (The actual Poisson's ratio range for NaCl crystals is 0.24-0.26, and the preset value is within a reasonable range, used to compare the ratio of bulk modulus to shear modulus.) The preset energy comparison threshold one (for formation energy) is set to -405 kJ / mol based on historical formation energy data for NaCl crystals (judgment rule: formation energy ≤ threshold one, which is considered acceptable; since formation energy is negative, the smaller the value, the more stable it is, consistent with the characteristics of ionic crystals). The preset crystal structure type to be matched is set to "cubic crystal system, ionic bonding, face-centered cubic structure" based on the cubic crystal system characteristics of NaCl crystals mentioned above (used to match the crystal type corresponding to the band gap value, ensuring the predicted result is consistent with the actual crystal structure). The preset energy comparison threshold two (for total energy) is set to -3470 kJ / mol based on historical total energy data for NaCl unit cells. eV / cell (judgment rule: total energy ≤ threshold two, i.e., meets the criteria; the total energy is negative, and the smaller the value, the more stable the unit cell); preset energy comparison threshold three (for electron affinity): based on the electron affinity characteristics of Cl atoms in NaCl crystals, the preset threshold three is 3.4-3.8 eV (judgment rule: electron affinity within this range, i.e., meets the criteria); preset error classification threshold: based on industry standards for crystal property prediction, the preset error classification threshold is 2% (i.e., the error to be judged < 2% is a small deviation, and the error to be judged ≥ 2% is a large deviation). For example, judgment result one (ratio of bulk modulus to shear modulus vs. Poisson's ratio): calculate the ratio of bulk modulus to shear modulus in the initial prediction result, i.e., 24 GPa ÷ 10 GPa = 2.4. The judgment result one is "compliant". Judgment result two (if the predicted formation energy is ≤ preset energy comparison threshold one, then it is "compliant"), judgment result three (if the band gap completely matches the crystal structure type to be matched, then it is "compliant"), judgment result four, and judgment result five. Based on judgment result two, and so on, the error to be measured (using the relative error formula: error to be judged = |initial predicted value - historical reference value| ÷ historical reference value × 100%) is calculated for each index. If all judgment results are compliant and all errors to be judged are <2%, then situation characteristic one is output; if any judgment result in the result set is "non-compliant", regardless of the error size, situation characteristic two is also output; if there is a situation that does not belong to either situation one or situation two, such as some judgment results being non-compliant and some errors exceeding the threshold range, then situation characteristic three is output, and the original data needs to be collected again. Correlation error groups (e.g., formation energy - total energy and band gap are correlated: formation energy is a derived value of total energy, and inaccurate total energy will lead to formation energy error; at the same time, formation energy reflects crystal stability, and stability will affect the band gap size (for example, in unstable crystals, band gap prediction is prone to deviation).Bulk modulus – shear modulus, total energy: Bulk modulus and shear modulus are directly related (together they determine Poisson's ratio). A large error in either one will lead to an abnormal Poisson's ratio. At the same time, total energy reflects the internal bonding force of the crystal, and the strength of the bonding force will affect the values of the two moduli. Eh (electron affinity) – band gap, formation energy: Both Eh and band gap are related to the electronic structure of the crystal. Band gap anomalies are often accompanied by Eh deviations. Formation energy reflects crystal stability and also indirectly affects the value of Eh. The final prediction result of crystal properties (if the formation energy has an error (3%–8%): first correct the total energy (because the formation energy is derived from the total energy), then, based on the corrected formation energy, reverse-correct the band gap (eliminating the band gap error caused by the deviation in crystal stability judgment). If the bulk modulus and shear modulus have errors, or Poisson's ratio is close to the critical value (0 or 0.5): simultaneously correct the two moduli, combine the deviation of the total energy, adjust the modulus values to ensure that the corrected Poisson's ratio is still within the range of 0–0.5, and the errors of the two moduli decrease. If the band gap has an error (after excluding the influence of formation energy): correct it in combination with the value of Eh (e.g., the band gap is too high, and Eh…) If the value is too low, simultaneously fine-tune both the band gap and the formation energy to ensure they conform to the electronic structure of the corresponding crystal. If the total energy has an error: first correct the total energy, then simultaneously correct the formation energy, bulk modulus, and shear modulus (because these three indicators are all related to the total energy). If Eh has an error: combine the correction results of the band gap and the formation energy, fine-tune the Eh value to ensure it conforms to the normal range of the corresponding crystal type, and matches the values of the band gap and the formation energy. The preprocessing error coefficient = error ÷ historical data reference value. Taking the total energy error compensation as an example: total energy correction value = predicted value − error × preprocessing error coefficient, and so on.
[0067] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.
Claims
1. A method of crystal property prediction, characterized by, Includes the following steps: Step S1: Obtain the original crystal structure data, normalize the original crystal structure data to obtain the crystal feature vector, obtain the encoded atomic-level input embedding vector based on the Fourier analysis network FAN layer and the crystal feature vector, train the Transformer encoder according to historical crystal structure data, preset the simulated atomic-level input embedding vector, input the simulated atomic-level input embedding vector into the trained Transformer encoder for feature learning, and obtain the preprocessing loss feature value of the atomic coordinates, atomic classification and lattice parameters. According to the preprocessing loss feature value and the atomic-level input embedding vector to be tested, obtain the pre-trained periodic embedding. Step S2: Based on the physical laws of crystal periodicity in the original crystal structure data, establish a pseudo-period prediction model. Based on the pre-trained periodic embedding and pseudo-period prediction model, combined with feedforward network processing, obtain the initial prediction results of crystal properties. Step S3: Compare the ratio between bulk modulus and shear modulus in the initial crystal property prediction results with the preset Poisson's ratio; compare the formation energy in the initial crystal property prediction results with the preset energy comparison threshold one; match the band gap value with the preset crystal structure type to be matched; compare the total energy with the preset energy comparison threshold two; and compare the electron affinity energy with the preset energy comparison threshold three. A result set is obtained by combining these results. Based on the test error between the predicted values of all indicators in the initial crystal property prediction results and the historical data reference values in the result set, error compensation is performed on the predicted values of all indicators in the initial crystal property prediction results, and the final crystal property prediction result is output.
2. The method of claim 1, wherein, Step S1 includes: The original crystal structure data is obtained, and the atomic type, atomic fraction coordinates and lattice parameters in the original crystal structure data are normalized to obtain normalized feature information. The atomic features, coordinate features and lattice parameters in the normalized feature information are concatenated to obtain the crystal feature vector. Perform a linear projection transformation on the crystal feature vector to obtain the projected features; Based on the FAN layer of the Fourier analysis network, the periodic crystal structure features in the projection features are explicitly encoded, and the atomic-level input embedding vector to be tested is output. Historical crystal structure data is acquired, and the Transformer encoder is pre-trained based on the general spatial distribution and periodic structure characteristics of the crystals in the historical crystal structure data to obtain a trained Transformer encoder.
3. The method of claim 2, wherein, Step S1 also includes: Based on the atomic-level input embedding vector to be tested, a simulated atomic-level input embedding vector is preset, and the simulated atomic-level input embedding vector is input into the trained Transformer encoder for feature learning to obtain the predicted atomic position, the pre-determined atomic type, and the predicted lattice parameters; The difference between the predicted atomic positions and the actual atomic positions in the historical crystal structure data is used to obtain the coordinate reconstruction loss feature value. The predicted atom type is compared with the actual atom type in the historical crystal structure data, and the number of incorrect judgments is counted to obtain the atom classification loss feature value. The difference between the predicted lattice parameters and the actual lattice parameters in the historical crystal structure data is used to obtain the lattice parameter loss characteristic value. The coordinate reconstruction loss feature value, the atom classification loss feature value, and the lattice parameter loss feature value are combined to form the preprocessing loss feature value.
4. The method of predicting crystal properties according to claim 3, wherein, Step S1 also includes: Based on the preprocessed loss feature values, the trained Transformer encoder is adjusted to obtain the adjusted Transformer encoder. Based on the adjusted Transformer encoder, which employs a combination of multi-head attention, layer normalization, and feedforward network structure, feature learning is performed on the atomic-level input embedding vector to extract the spatial arrangement, bonding relationships, and periodic general features of atoms in the crystal, and output a pre-trained periodic embedding with common laws of crystal structure.
5. The method for predicting crystal properties according to claim 4, characterized in that, Step S2 includes: Based on the periodic physical laws of crystals in the original crystal structure data, a cell period offset term and a periodic mirror mapping rule are preset. Based on the cell period offset term, the periodic mirror mapping rule and historical crystal structure data, a pseudo-finite periodic attention mechanism is trained. The traditional multi-head attention inside the pseudo-trained Transformer encoder is replaced with the pseudo-finite periodic attention mechanism to establish a pseudo-periodic prediction model. The pre-trained period is embedded into the input pseudo-period prediction model to extract long-range crystal interaction features. The long-range crystal interaction features are normalized and processed by a feedforward network to obtain initial prediction results of crystal properties. The initial prediction results of crystal properties include bulk modulus, shear modulus, formation energy, band gap value, total energy, and electron affinity.
6. The method for predicting crystal properties according to claim 5, characterized in that, Step S3 includes: Preset Poisson's ratio, energy comparison threshold one, crystal structure type to be matched, energy comparison threshold two, and energy comparison threshold three; The ratio of the bulk modulus to the shear modulus is compared with Poisson's ratio to obtain judgment result one; the formation energy is compared with the energy to threshold one to obtain judgment result two; the band gap value is matched with the crystal structure type to be matched to obtain judgment result three; the total energy is compared with the energy to threshold two to obtain judgment result four; and the electron affinity energy is compared with the energy to threshold three to obtain judgment result five. The judgment result 1, judgment result 2, judgment result 3, judgment result 4 and judgment result 5 are combined to form a result set; Historical data reference values are obtained, and the predicted values and historical data reference values of all indicators in the initial prediction results of crystal properties are statistically analyzed to obtain the error to be judged. An error division threshold is preset. If the judgment results of all indicators in the result set are in compliance and the error to be judged is less than the error division threshold, then the first case feature is output. Based on characteristic one of the aforementioned situations, the initial prediction result of the crystal properties is determined to be the final prediction result, and the verification result is output.
7. The method for predicting crystal properties according to claim 6, characterized in that, Step S3 also includes: If any result in the result set is not compliant, or if any error in the error to be judged is within the error division threshold, then output characteristic two. The remaining case features after removing case features one and case features two from the result set and the error to be judged are determined to be case feature three. Based on case feature three, the original crystal structure data is re-acquired and analyzed, and secondary data processing information is output. Based on characteristic two of the situation, and according to the correlation between the indicators, the indicators in the initial prediction results of crystal properties are grouped into groups based on the correlation between their errors, thus obtaining the correlated error groups; The ratio of the measured error to the historical data reference value is used to obtain the preprocessing error coefficient. Based on the preprocessing error coefficient and the measured error, error compensation is performed on the predicted values of all indicators in the initial prediction results of crystal properties, and the final prediction results of crystal properties are output.