Method for accelerating construction of multi-component material thermodynamic database and storage medium
An active learning framework with first-principles calculations iteratively improves the machine learning model for multi-component materials, addressing computational inefficiencies and enhancing accuracy and reliability in thermodynamic database construction.
Patent Information
- Application Number
- CN202410914622.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-07-09
AI Technical Summary
Existing methods for constructing thermodynamic databases of multi-component materials are computationally expensive and lack self-improvement mechanisms, leading to unreliable predictions due to the neglect of magnetism and lack of iterative learning.
An active learning framework coupled with first-principles calculations is used to divide terminal compounds into labeled and unlabeled samples, iteratively improving a machine learning model through an enhanced committee strategy, incorporating macro- and micro-scale features and uncertainty estimation.
This approach accelerates the construction of a multi-component material thermodynamic database by reducing computational resources and time, enhancing model accuracy, and providing uncertainty estimates, while being environmentally friendly and cost-effective.
Smart Images

Figure CN119069042A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computational thermodynamics of metal materials, and in particular to a method for accelerating the construction of a multi-component material thermodynamics database by using an active learning framework coupled with first-principles calculations. Background Art
[0002] Thermodynamic data of alloys and compounds, such as formation enthalpy (also known as standard heat of formation), play an important role in various applications, such as phase diagram calculations and material design, in the exploration of new high melting point materials for advanced coal-fired power plants, heat exchangers, filters and turbines, etc. In the CALculation of PHAse Diagrams (CALPHAD) method, which combines phase-based thermodynamic modeling with equilibrium calculations, the formation enthalpy is usually calculated as a part of the Gibbs free energy, such as the Gibbs free energy of terminal compounds (terminal members for short). The careful quantification and analysis of the Gibbs free energy of each phase helps to intuitively understand the formation and phase transition processes of different phases. This understanding is a key pillar for effective and efficient material design strategies. Density functional theory (DFT) calculations are usually used to accurately calculate the formation energy of terminal compounds, but inevitably lead to huge computational costs, as the number of terminal compounds can reach 10^5 for real alloys containing about 10 different elements. This provides a unique opportunity to apply machine learning (ML) in practice.
[0003] Current research on the use of machine learning for compound energy prediction focuses only on non-magnetic compounds, such as Jean-Claude Crivello from the Sorbonne University in Paris, France, who studied the formation of sigma phase members. When studying alloy compounds in these works, the first-principles calculations used did not consider the contribution of magnetism, which greatly affected the reliability of model predictions. Secondly, for example, the work of Logan Ward's team from Northwestern University in the United States to predict the formation energy of inorganic compounds, the machine learning models used in these works did not have a self-training and self-improvement process, and were often put into use after only one training, which obviously could not meet the needs of machine learning models in the era of big data. Our work not only considers the contribution of magnetism in alloys to the formation energy, but also can continuously iterate the machine learning model as the database continues to grow, improving the model accuracy and generalization ability. Summary of the invention
[0004] The purpose of the present invention is to overcome the blindness of first-principles calculations in establishing thermodynamic databases, provide an active learning framework coupled with first-principles calculations, and accelerate the construction of a multi-component material thermodynamic database. By giving a number of initial compound formation energy samples, the automated process of material thermodynamic database construction can be realized, and the operation is simple and fast.
[0005] The purpose of the present invention can be achieved by the following technical solutions:
[0006] A first aspect of the present invention provides a method for accelerating the construction of a multi-component material thermodynamic database, comprising the following steps:
[0007] S1: Before the iteration begins, all the terminal compounds in the system are divided into three subsets: a labeled sample pool, an unlabeled sample pool, and a test set. The data formation of the labeled sample pool and the test set can be obtained through first-principles calculations, and the unlabeled sample pool consists of terminal compounds that have not been calculated.
[0008] S2: Build a decision model based on the labeled sample pool and observe the prediction accuracy on the test set. If the prediction accuracy reaches the predetermined standard, stop the iteration. Otherwise, select unlabeled samples for calculation optimization based on the enhanced committee strategy.
[0009] S3: Use an ensemble learning model to predict the formation energy and uncertainty distribution of all unlabeled samples, and use the results to establish a system thermodynamic database.
[0010] Furthermore, S1 specifically includes:
[0011] S1-1: Assume that the formation energy of N terminal compounds of the multi-component system A needs to be calculated. First, the formation energy of terminal compounds of number L and T is calculated using VASP software, where L sample set is the initial labeled sample set, T sample set is the test set, and the remaining N-(L+T) samples are the unlabeled sample set U;
[0012] S1-2: perform data preprocessing on N samples and construct 11 macro-scale features;
[0013] S1-3: Perform data preprocessing on N samples and construct 110 micro-scale features.
[0014] Further, in S1-1, each terminal compound is calculated using the following formula:
[0015]
[0016] Where n represents the number of elements in each terminal group, and k represents the different element properties. i and They correspond to the composition and specified element properties of the i-th element, respectively.
[0017] Further, in S1-3, the microscopic scale feature is divided into two parts:
[0018] The first part is e*p features, where e represents the number of inequivalent sites in the Wyckoff position of the compound, and p is the 11 element properties from Table 1, using one-hot encoding of the Wyckoff position of the compound;
[0019] The second part is e*p features, which describe the local environment characteristics of each equivalent site. The Voronoi technique is used to calculate the nearest neighbor Voronoi polyhedron around each atom in the unit cell to weight the energy contribution of the surrounding atoms to the central atom, which is calculated by the following formula:
[0020]
[0021] Where P i represents the properties of the i-th central atom, P h represents the property of the nth atom nearest to the central atom i, A n Represents the area of the Voronoi polyhedron corresponding to the nth nearest atom of the central atom i.
[0022] Furthermore, S2 specifically includes:
[0023] S2-1: Modeling is performed on the L data set using a verified support vector machine algorithm as a decision model; the support vector machine regression algorithm uses the sciki-learn library based on the Python language, and uses a grid search and 10-fold cross validation method to screen the best model parameters and model D;
[0024] S2-2: Calculate the mean absolute error (MAE) of the decision model D on the test set T. The formula is as follows:
[0025]
[0026] in is the predicted value of the decision model D in the i-th test set sample, y i represents the true value of the i-th test set sample, and n represents the total number of test set samples;
[0027] If the indicator MAE is less than the threshold, the iteration is stopped, otherwise based on the enhanced committee strategy, unlabeled samples are selected for calculation optimization.
[0028] Furthermore, the selection of unlabeled samples for computational optimization based on the enhanced committee strategy includes:
[0029] Select some terminal compounds from the unlabeled sample pool as candidates, and then perform first-principles calculations on these candidates. After the calculations are completed, add them to the labeled sample pool and repeat S2 until the decision model reaches the expected accuracy on the test set and stop the iteration.
[0030] Furthermore, the selection of unlabeled samples for computational optimization based on the enhanced committee strategy specifically includes:
[0031] A-1: Based on the labeled sample set L, five independent training sets are obtained using bootstrap sampling for the subsequent training of five committee models;
[0032] A-2: Use the support vector machine algorithm as the base model, train 5 committee models on 5 independent training sets, and use grid search and 10-fold cross validation methods to screen the best model parameters and models;
[0033] A-3: Use the Enhanced-Query-By-Committee (EQBC) strategy to calculate the uncertainty index on the unlabeled sample set U;
[0034] A-4: Select several candidates with the largest uncertainty indicators from the unlabeled sample pool, then perform first-principles calculations on these candidates, and add them to the labeled sample pool L after the calculations are completed.
[0035] Furthermore, the calculation formula of the uncertainty index is as follows:
[0036]
[0037] Where m represents the mth sample in the unlabeled sample set U, represents the uncertainty of the sample, P represents the number of committee models, represents the predicted value of the sample on the p-th committee model, Represents the mean value predicted by the m-th sample on P committee models.
[0038] Furthermore, S3 specifically includes:
[0039] S3-1: Using the p committee models trained in the previous iteration, output the distribution of the mean absolute percentage error (MAPE) estimate of the remaining unlabeled samples, as follows:
[0040]
[0041] where y i represents the ith actual value, represents the predicted value of the i-th one, and n represents all data points;
[0042] S3-2: The decision model D and the p committee models completed in the previous round of iterative training, a total of (1+p) models, are combined into an ensemble learning model. The stacked ensemble method is used, and the grid search and 10-fold cross-validation methods are used to select the hyperparameters of the best ensemble learning model M. The prediction process of model M is as follows:
[0043]
[0044] Where m represents the number of integrated models, M m (x) represents the predicted value of the mth model for sample x, M meta Represents the meta-model, which takes the combined feature H(x) as input and outputs the final prediction result;
[0045] S3-3: Use the integrated learning model M to output the formation energy of all unlabeled samples and give the uncertainty distribution map of the samples for the establishment of the system thermodynamic database.
[0046] A second aspect of the present invention provides a storage medium comprising computer executable instructions, which, when executed by a computer processor, is used to execute the method for accelerating the construction of a multi-component material thermodynamic database as described above.
[0047] Compared with the prior art, the present invention has the following technical advantages:
[0048] 1. The method of the present invention overcomes the blindness of traditional first-principles calculations, saves resources and time, couples the active learning framework with the first-principles calculations, and can automatically establish a material thermodynamics database given an initial sample set.
[0049] 2. The method of the present invention does not involve experiments or use chemicals in the entire process, does not produce chemical pollution, and is in line with the concept of green environmental protection; it is simple to operate, low in cost, easy to implement, and suitable for promotion and application.
[0050] 3. The method of the present invention can predict the formation energy of materials in advance through the calculation results of the machine learning model, and select samples with lower uncertainty for thermodynamic modeling, which can save modeling time and resources, improve modeling efficiency, play a guiding role, and avoid blindness.
[0051] 4. A multi-scale feature set that takes into account both macroscopic and microscopic scales is given, which can more intuitively demonstrate the contribution of different scale features to the formation energy of compounds.
[0052] 5. The method of the present invention takes into account the contribution of magnetism to energy in the process of machine learning modeling, and calculates the magnetic moments of some pure components through first principles, which ensures the rationality of modeling and increases the accuracy of the model.
[0053] 6. The present invention provides uncertainty estimates for all uncalculated samples, increasing the interpretability of the machine learning model. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 Figure 2 is a flowchart of the active learning framework for coupling the first principles calculation of this method.
[0055] Figure 2 It is a sigma phase crystal structure.
[0056] Figure 3 The Voronoi polyhedron for each atom.
[0057] Figure 4 MAPEE on unlabeled samples for the fifth iteration. DETAILED DESCRIPTION
[0058] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms and other features not clearly described in this technical solution are all considered to be common technical features disclosed in the prior art.
[0059] In specific implementation, the technical solution adopted by the present invention is: a method for accelerating the construction of a multi-component material thermodynamic database by coupling an active learning framework of first-principles calculations, see Figure 1 , including the following steps:
[0060] Step 1: Before the iteration begins, all the terminal compounds in the system are divided into three subsets: a labeled sample pool, an unlabeled sample pool, and a test set. The data of the labeled sample pool and the test set can be formed by first-principles calculations, and the unlabeled sample pool consists of terminal compounds that have not been calculated. ;
[0061] Step 2: Build a decision model based on the labeled sample pool and observe the prediction accuracy on the test set. If the expected accuracy is achieved, stop the iteration, skip step 3, and proceed to step 4. If the expected accuracy is not achieved, proceed to step 3;
[0062] Step 3: Use an active learning framework based on the enhanced committee strategy to select some candidates (i.e., terminal compounds) from the unlabeled sample pool, and then perform first-principles calculations on these candidates. After the calculations are completed, add them to the labeled sample pool and repeat step 2 until the decision model reaches the expected accuracy on the test set, and then stop the iteration;
[0063] Step 4: Use the integrated learning model to output the formation energy and uncertainty distribution map of all unlabeled samples for the establishment of the system thermodynamic database.
[0064] Wherein, the step 1 comprises:
[0065] Step 1.1, assuming that the multi-component A system needs to calculate the formation energy of N terminal compounds, first use VASP software to calculate the formation energy of terminal compounds with numbers L and T. The L sample set is called the initial labeled sample set, and the T sample set is called the test set. The remaining N-(L+T) samples are called the unlabeled sample set U.
[0066] Step 1.2, preprocess the data of N samples and construct 11 macro-scale features, the element properties involved are provided in Table 1. Each terminal compound is calculated using the following formula:
[0067]
[0068] Where n represents the number of elements in each terminal group member, and k represents the different element properties. i and They correspond to the composition and the specified element properties of the i-th element respectively;
[0069] Step 1.3, perform data preprocessing on N samples and construct 110 micro-scale features, the element properties involved will be provided in Table 1. The micro-scale features are divided into two parts: the first part has a total of e*p features, where e represents the number of inequivalent sites in the Wyckoff position of the compound, and p is the 11 element properties from Table 1, using one-hot encoding of the Wyckoff position of the compound; the second part also has e*p features, describing the local environmental characteristics of each equivalent site. The Voronoi technique is used to calculate the nearest neighbor Voronoi polyhedron around each atom in the unit cell to weight the energy contribution of the surrounding atoms to the central atom. Calculated by the following formula:
[0070]
[0071] Where P i represents the properties of the i-th central atom, P n represents the property of the nth atom nearest to the central atom i, A n Represents the area of the Voronoi polyhedron corresponding to the nth nearest atom of the central atom i.
[0072] The step 2 comprises:
[0073] Step 2.1, using a machine learning algorithm, here we recommend the proven support vector machine algorithm as the decision model, and modeling on the L data set. The support vector machine regression algorithm uses the sciki-learn library based on the Python language, and uses grid search and 10-fold cross validation methods to screen the best model parameters and model D;
[0074] Step 2.2, calculate the mean absolute error (MAE) of the decision model D on the test set T, the formula is as follows:
[0075]
[0076] in is the predicted value of the decision model D in the i-th test set sample, y i represents the true value of the i-th test set sample, and n represents the total number of test set samples.
[0077] If the indicator MAE is less than the threshold value of 500 J / mol*atom, skip step 3 and proceed to step 4.
[0078] The step 3 comprises:
[0079] Step 3.1: Based on the labeled sample set L, bootstrap sampling is used to obtain 5 independent training sets for the subsequent training of 5 committee models.
[0080] Step 3.2, using the support vector machine algorithm as the base model, train 5 committee models on 5 independent training sets, and use grid search and 10-fold cross validation methods to screen the best model parameters and models;
[0081] In step 3.3, the Enhanced-Query-By-Committee (EQBC) strategy [1] is used to calculate the uncertainty index on the unlabeled sample set U. The index calculation formula is as follows:
[0082]
[0083] Where m represents the mth sample in the unlabeled sample set U, represents the uncertainty of the sample, P represents the number of committee models, here it is 5, represents the predicted value of the sample on the p-th committee model, Represents the mean value predicted by the m-th sample on P committee models.
[0084] Step 3.4, select several candidates with the largest uncertainty index (ie, terminal compounds) from the unlabeled sample pool, then perform first-principles calculations on these candidates, and add them to the labeled sample pool L after the calculations are completed.
[0085] The step 4 comprises:
[0086] Step 4.1, using the p committee models trained in the previous iteration, output the distribution of the mean absolute percentage error (MAPE) estimate of the remaining unlabeled samples, as follows:
[0087]
[0088] where y i represents the ith actual value, represents the predicted value of the ith one, and n represents all data points.
[0089] Step 4.2: The decision model D and the p committee models completed in the previous round of iterative training, a total of (1+p) models, are combined into an ensemble learning model. The stacked ensemble method is used to select the hyperparameters of the best ensemble learning model M using grid search and 10-fold cross validation methods. The prediction process of model M is as follows:
[0090]
[0091] Where m represents the number of integrated models. Assuming the number of committees p is 5, there are 6 in total, M m (x) represents the predicted value of the mth model for sample x, M meta Represents the meta-model, which takes the combined feature H(x) as input and outputs the final prediction result.
[0092] Step 4.3, use the integrated learning model M to output the formation energy of all unlabeled samples and give the uncertainty distribution map of the samples for the establishment of the system thermodynamic database.
[0093] Example 1
[0094] In this embodiment, an active learning framework coupled with first-principles calculations is used to accelerate the construction of a multi-component material thermodynamic database, and the steps are as follows:
[0095] Step 1: Taking the sigma phase in the Cr-Fe-Co-Ni quaternary system as an example, a total of 1020 (excluding pure component energy) sigma phase terminal compounds need to be calculated by first principles. Before the iteration begins, all sigma phase terminal compounds are divided into three subsets: a labeled sample pool, an unlabeled sample pool, and a test set.
[0096] Step 1.1, the Cr-Fe-Co-Ni system needs to calculate the formation energies of 1020 terminal compounds. First, the VASP software is used to calculate the formation energies of 180 and 168 terminal compounds. The initial labeled sample set L and the test set T contain a total of 348 samples, and the remaining 672 samples are called the unlabeled sample set U.
[0097] Step 1.2, data preprocessing was performed on 1020 samples to construct 11 macro-scale features, the element properties involved are provided in Table 1. Each terminal compound was calculated using the following formula:
[0098]
[0099] Where n represents the number of elements in each endpoint group member, and k represents different attribute types. i and They correspond to the composition and specified attributes of the i-th element respectively;
[0100] Table 1. Properties of elements used to calculate multi-scale feature sets
[0101]
[0102]
[0103] *From reference [2]
[0104] Integrate the values of the density of states below the Fermi level
[0105] The data comes from the website [http: / / phases.imet-db.ru / elements / main.aspx]
[0106] Step 1.3, data preprocessing was performed on 1020 samples to construct 110 micro-scale features, the element properties involved are provided in Table 1. The micro-scale features are divided into two parts: the first part has 55 features, which are based on the atomic microstructure and atomic local environment features. Figure 2 Represents the microstructure of the sigma phase. One-hot encoding is used to encode the Wyckoff position of the sigma phase; the second part has 55 features, describing the local environment characteristics of atoms. The Voronoi technique is used to calculate the nearest neighbor Voronoi polyhedron around each atom in the unit cell to weight the energy contribution of the surrounding atoms to the central atom. Figure 3 As shown. It is calculated by the following formula:
[0107]
[0108] Where P i represents the properties of the i-th central atom, P n represents the property of the nth atom nearest to the central atom i, A n Represents the area of the Voronoi polyhedron corresponding to the nth nearest atom of the central atom i.
[0109] Step 2: Set the prediction accuracy threshold of formation energy to 500 J / mol*atom, build a decision model based on the labeled sample pool, and observe the prediction accuracy on the test set. If it is reached, stop the iteration, skip step 3, and proceed to step 4. If the expected accuracy is not achieved, proceed to step 3;
[0110] Step 2.1, use the support vector machine regression algorithm from the scikit-learn library to perform modeling on the labeled sample data set of this iteration, and use the grid search and 10-fold cross validation method to screen the best model parameters and model D;
[0111] Step 2.2, calculate the mean absolute error (MAE) of the decision model D on the test set T, the formula is as follows:
[0112]
[0113] in is the predicted value of the decision model D in the i-th test set sample, y i represents the true value of the i-th test set sample, and n represents the total number of test set samples.
[0114] If the indicator MAE is less than the threshold value of 500 J / mol*atom, skip step 3 and proceed to step 4.
[0115] Step 3: Use the enhanced committee strategy to select some candidates (i.e., terminal compounds) from the unlabeled sample pool, and then perform first-principles calculations on these candidates. After the calculations are completed, add them to the labeled sample pool and repeat step 2 until the decision model reaches the expected accuracy on the test set, and then stop the iteration;
[0116] Step 3.1: Based on the labeled sample set L, bootstrap sampling is used to obtain 5 independent data sets for the subsequent training of the 5 committee members.
[0117] Step 3.2, using the support vector machine algorithm as the base model, training 5 committee members on 5 independent data sets, and using grid search and 10-fold cross validation methods to screen the best model parameters and models;
[0118] In step 3.3, the Enhanced-Query-By-Committee (EQBC) strategy [1] is used to calculate the uncertainty index on the unlabeled sample set U. The index calculation formula is as follows:
[0119]
[0120] Where m represents the mth sample in the unlabeled sample set U, represents the uncertainty of the sample, P represents the number of committee models, here it is 5, represents the predicted value of the sample on the p-th committee model, Represents the mean value predicted by the m-th sample on P committee models.
[0121] Step 3.4, select 50 candidates with the largest uncertainty indicators (i.e., terminal compounds) from the unlabeled sample pool, then perform first-principles calculations on these candidates, and add them to the labeled sample pool L after the calculations are completed.
[0122] Step 4: Use the integrated learning model to output the formation energy and uncertainty distribution map of all unlabeled samples for the establishment of the system thermodynamic database.
[0123] Step 4.1, using the five committee models trained in the previous iteration, output the distribution of the mean absolute percentage error estimate of the remaining unlabeled samples. The MAPE calculation formula is as follows:
[0124]
[0125] where y i represents the ith actual value, represents the predicted value of the ith one, and n represents all data points.
[0126] Figure 4 The MAPE-estimation indicator for estimating the uncertainty distribution of unlabeled samples after the fifth iteration is shown.
[0127] Step 4.2, the decision model D and the five committee models completed in the previous round of iterative training, a total of six models, are combined into an ensemble learning model. The stacked ensemble method is used, and the grid search and 10-fold cross-validation methods are used to screen the parameters of the best ensemble learning model M. The prediction process of model M is as follows:
[0128]
[0129] Where m represents the number of integrated models, here p = 5, a total of 6, M m (x) represents the predicted value of the mth model for sample x, M meta Represents the meta-model, which takes the combined feature H(x) as input and outputs the final prediction result.
[0130] Step 4.3, use the integrated learning model M to output the formation energy of all unlabeled samples for the establishment of the system thermodynamic database.
[0131] The above-mentioned embodiment method can overcome the blindness of high-throughput first-principles calculations, save resources and time, and based on the active learning coupled first-principles framework, only a small amount of initial compound formation energy is needed to obtain the formation energy of all terminal compounds of the complete alloy system, which is convenient and fast. This method can minimize the computing resources of the first principles through the enhanced committee query strategy. Each iteration selects the most valuable candidate samples for first-principles calculations, which can save a lot of computing time and resources, improve the efficiency of calculations, play a guiding role, and avoid blindness.
[0132] In summary, the above embodiment is based on the active learning framework coupled with the first principle calculation, and the method for constructing a thermodynamic database of multi-component materials. Before the iteration starts, all terminal compounds of the system are divided into three subsets: a labeled sample pool, an unlabeled sample pool, and a test set. The formation energy of the labeled sample pool and the test set data is calculated by the first principle, and the unlabeled sample pool is composed of terminal compounds that have not been calculated by the first principle. Then, a decision model is constructed based on the labeled sample pool, and the prediction accuracy is observed on the test set. If it is reached, the iteration is stopped. The model can directly output the formation energy of all unlabeled samples for the establishment of a subsequent thermodynamic database. If the expected accuracy is not achieved, an active learning algorithm based on the enhanced committee strategy is used to select some candidates (i.e., terminal compounds) from the unlabeled sample pool, and then these candidates are calculated by the first principle. After the calculation is completed, the labeled sample pool is added until the decision model reaches the expected accuracy on the test set, and the iteration is stopped. At this time, the decision model is combined with the committee model generated by the previous round of iteration, and the formation energy of all unlabeled samples is output using the integrated learning model for the establishment of the system thermodynamic database.
[0133] This embodiment also proposes a computer-readable storage medium, the storage medium stores computer instructions, the computer instructions are used to enable the computer to execute the above-mentioned active learning framework of coupled first-principles calculations, and accelerate the method of building a multi-component material thermodynamic database. The storage medium can be an electronic medium, a magnetic medium, an optical medium, an electromagnetic medium, an infrared medium, or a semiconductor system or a propagation medium. The storage medium can also include semiconductor or solid-state memory, a magnetic tape, a removable computer disk, a random access memory (RAM), a read-only memory (ROM), a hard disk, and an optical disk. The optical disk can include a compact disk-read only memory (CD-ROM), a compact disk-read and write (CD-RW), and a DVD.
[0134] References
[0135] [1] D. Wu, VJ Lawhern, S. Gordon, BJ Lance, and C.-T. Lin, "Offline EEG-based driver drowsinessestimation using enhanced batch-mode active learning (EBMAL) for regression," in 2016 IEEEInternational Conference on Systems, Man, and Cybernetics (SMC), Budapest, Hungary: IEEE, Oct. 2016, pp.000730–000736.doi:10.1109 / SMC.2016.7844328.
[0136] [2] Y.Zhang, JRGEvans, and S.Yang, "Corrected Values for BoilingPoints and Enthalpies ofVaporization of Elements in Handbooks," J.Chem.Eng.Data, vol.56, no.2, Art.no.2, Feb.2011, doi:10.1021 / je1011086.
[0137] The above describes the method embodiments of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made according to the purpose of the invention. Any parameter changes or calculation simplifications made according to the principle of the technical solution of the present invention, as long as they meet the purpose of the invention and do not deviate from the principles and concepts of the method for constructing a multi-component material thermodynamic database in an active learning framework coupled with first-principles calculations of the present invention, all fall within the scope of protection of the present invention.
Claims
1. A method to accelerate the construction of a multi-component material thermodynamic database. It is characterized in that The following steps are involved: S1: Before the iteration begins, all the terminal compounds in the system are divided into three subsets: a labeled sample pool, an unlabeled sample pool, and a test set. The data formation of the labeled sample pool and the test set can be obtained through first-principles calculations, and the unlabeled sample pool consists of terminal compounds that have not been calculated. S2: Build a decision model based on the labeled sample pool and observe the prediction accuracy on the test set. If the prediction accuracy reaches the predetermined standard, stop the iteration. Otherwise, select unlabeled samples for calculation optimization based on the enhanced committee strategy. S3: Use an ensemble learning model to predict the formation energy and uncertainty distribution of all unlabeled samples, and use the results to establish a system thermodynamic database.
2. A method for accelerating the construction of a multi-component material thermodynamic database according to claim 1, It is characterized in that S1 specifically includes: S1-1: Assume that the formation energy of N terminal compounds of the multi-component system A needs to be calculated. First, the formation energy of terminal compounds of number L and T is calculated using VASP software, where L sample set is the initial labeled sample set, T sample set is the test set, and the remaining N-(L+T) samples are the unlabeled sample set U; S1-2: perform data preprocessing on N samples and construct 11 macro-scale features; S1-3: Perform data preprocessing on N samples and construct 110 micro-scale features.
3. A method for accelerating the construction of a multi-component material thermodynamic database according to claim 2, It is characterized in that In S1-1, each terminal compound is calculated using the following formula: Where n represents the number of elements in each terminal group, and k represents the different element properties. i and They correspond to the composition and specified element properties of the i-th element, respectively.
4. A method for accelerating the construction of a multi-component material thermodynamic database according to claim 2, It is characterized in that In S1-3, the microscopic scale features are divided into two parts: The first part is e*p features, where e represents the number of inequivalent sites in the Wyckoff position of the compound, and p is the 11 element properties from Table 1, using one-hot encoding of the Wyckoff position of the compound; The second part is e*p features, which describe the local environment characteristics of each equivalent site. The Voronoi technique is used to calculate the nearest neighbor Voronoi polyhedron around each atom in the unit cell to weight the energy contribution of the surrounding atoms to the central atom, which is calculated by the following formula: Where P i represents the properties of the i-th central atom, P n represents the property of the nth atom nearest to the central atom i, A n Represents the area of the Voronoi polyhedron corresponding to the nth nearest atom of the central atom i.
5. A method for accelerating the construction of a multi-component material thermodynamic database according to claim 1, It is characterized in that S2 specifically includes: S2-1: Modeling is performed on the L data set using a verified support vector machine algorithm as a decision model; the support vector machine regression algorithm uses the sciki-learn library based on the Python language, and uses a grid search and 10-fold cross validation method to screen the best model parameters and model D; S2-2: Calculate the mean absolute error (MAE) of the decision model D on the test set T. The formula is as follows: in is the predicted value of the decision model D in the i-th test set sample, y i represents the true value of the i-th test set sample, and n represents the total number of test set samples; If the indicator MAE is less than the threshold, the iteration is stopped, otherwise based on the enhanced committee strategy, unlabeled samples are selected for calculation optimization.
6. A method for accelerating the construction of a multi-component material thermodynamic database according to claim 1, It is characterized in that The method of selecting unlabeled samples for computational optimization based on the enhanced committee strategy includes: Select some terminal compounds from the unlabeled sample pool as candidates, and then perform first-principles calculations on these candidates. After the calculations are completed, add them to the labeled sample pool and repeat S2 until the decision model reaches the expected accuracy on the test set and stop the iteration.
7. A method for accelerating the construction of a multi-component material thermodynamic database according to claim 6, It is characterized in that The method of selecting unlabeled samples for computational optimization based on the enhanced committee strategy specifically includes: A-1: Based on the labeled sample set L, five independent training sets are obtained using bootstrap sampling for the subsequent training of five committee models; A-2: Use the support vector machine algorithm as the base model, train 5 committee models on 5 independent training sets, and use grid search and 10-fold cross validation methods to screen the best model parameters and models; A-3: Use the Enhanced-Query-By-Committee (EQBC) strategy to calculate the uncertainty index on the unlabeled sample set U; A-4: Select several candidates with the largest uncertainty indicators from the unlabeled sample pool, then perform first-principles calculations on these candidates, and add them to the labeled sample pool L after the calculations are completed.
8. A method for accelerating the construction of a multi-component material thermodynamic database according to claim 7, It is characterized in that The calculation formula of the uncertainty index is as follows: Where m represents the mth sample in the unlabeled sample set U, represents the uncertainty of the sample, P represents the number of committee models, represents the predicted value of the sample on the p-th committee model, Represents the mean value predicted by the m-th sample on P committee models.
9. A method for accelerating the construction of a multi-component material thermodynamic database according to claim 1, It is characterized in that S3 specifically includes: S3-1: Using the p committee models trained in the previous iteration, output the distribution of the mean absolute percentage error (MAPE) estimate of the remaining unlabeled samples, as follows: where y i represents the ith actual value, represents the predicted value of the i-th one, and n represents all data points; S3-2: The decision model D and the p committee models completed in the previous round of iterative training, a total of (1+p) models, are combined into an ensemble learning model. The stacked ensemble method is used, and the grid search and 10-fold cross-validation methods are used to select the hyperparameters of the best ensemble learning model M. The prediction process of model M is as follows: Where m represents the number of integrated models, M m (x) represents the predicted value of the mth model for sample x, M meta Represents the meta-model, which takes the combined feature H(x) as input and outputs the final prediction result; S3-3: Use the integrated learning model M to output the formation energy of all unlabeled samples and give the uncertainty distribution map of the samples for the establishment of the system thermodynamic database.
10. A storage medium containing computer executable instructions, It is characterized in that When the storage medium of the computer executable instructions is executed by a computer processor, it is used to execute the method for accelerating the construction of a multi-component material thermodynamic database according to any one of claims 1 to 9.
Citation Information
Patent Citations
High-entropy alloy mechanical property calculation method based on atom occupying behavior in sublattice
CN112395762A
Construction method of material performance prediction model, material performance prediction method and system
CN116562159A
Method and device for accelerating material structure search, electronic equipment and medium
CN116646026A
CALPHAD method for predicting PCT curve of hydrogen storage alloy
CN117275626A
Machine Learning to Accelerate Alloy Design
US20200257933A1
Cited By
AI-driven X-ray spectroscopy experiment active optimization and decision-making system and method
CN122656000A
AI-driven x-ray spectroscopy experiment active optimization and decision system and method
CN122656000B
