CN111798940A - Method and device for predicting superconducting material based on deep neural network algorithm - Google Patents
Method and device for predicting superconducting material based on deep neural network algorithm Download PDFInfo
- Publication number
- CN111798940A CN111798940A CN202010600263.1A CN202010600263A CN111798940A CN 111798940 A CN111798940 A CN 111798940A CN 202010600263 A CN202010600263 A CN 202010600263A CN 111798940 A CN111798940 A CN 111798940A
- Authority
- CN
- China
- Prior art keywords
- neural network
- superconducting material
- deep neural
- superconducting
- predicting
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 239000000463 material Substances 0.000 title claims abstract description 186
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 95
- 238000000034 method Methods 0.000 title claims abstract description 89
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 54
- 238000012549 training Methods 0.000 claims abstract description 50
- XEEYBQQBJWHFJM-UHFFFAOYSA-N Iron Chemical compound [Fe] XEEYBQQBJWHFJM-UHFFFAOYSA-N 0.000 claims description 30
- 230000006870 function Effects 0.000 claims description 28
- 238000010606 normalization Methods 0.000 claims description 17
- 229910052742 iron Inorganic materials 0.000 claims description 15
- 238000003062 neural network model Methods 0.000 claims description 15
- RYGMFSIKBFXOCR-UHFFFAOYSA-N Copper Chemical compound [Cu] RYGMFSIKBFXOCR-UHFFFAOYSA-N 0.000 claims description 14
- 229910052802 copper Inorganic materials 0.000 claims description 14
- 239000010949 copper Substances 0.000 claims description 14
- 230000004913 activation Effects 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 11
- 238000004590 computer program Methods 0.000 claims description 9
- 210000002569 neuron Anatomy 0.000 claims description 9
- 238000005457 optimization Methods 0.000 claims description 8
- 238000011478 gradient descent method Methods 0.000 claims description 7
- 238000012216 screening Methods 0.000 claims description 6
- 239000006185 dispersion Substances 0.000 claims description 5
- 238000010801 machine learning Methods 0.000 abstract description 11
- 238000012360 testing method Methods 0.000 abstract description 11
- 238000002474 experimental method Methods 0.000 abstract description 6
- 230000008569 process Effects 0.000 abstract description 6
- 230000005484 gravity Effects 0.000 description 10
- 230000000052 comparative effect Effects 0.000 description 9
- 238000013507 mapping Methods 0.000 description 6
- 238000007637 random forest analysis Methods 0.000 description 6
- 239000002887 superconductor Substances 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 239000002609 medium Substances 0.000 description 4
- 238000011160 research Methods 0.000 description 3
- 238000012706 support-vector machine Methods 0.000 description 3
- 241001137251 Corvidae Species 0.000 description 2
- 229910008561 TiTe2 Inorganic materials 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 239000013078 crystal Substances 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000000611 regression analysis Methods 0.000 description 2
- 239000000126 substance Substances 0.000 description 2
- 101100012466 Drosophila melanogaster Sras gene Proteins 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000002844 melting Methods 0.000 description 1
- 230000008018 melting Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000012120 mounting media Substances 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000009401 outcrossing Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C60/00—Computational materials science, i.e. ICT specially adapted for investigating the physical or chemical properties of materials or phenomena associated with their design, synthesis, processing, characterisation or utilisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention provides a method for predicting a novel superconducting material based on a deep neural network algorithm, which predicts the performance of the superconducting material by using the deep neural network, optimizes training errors and testing errors more rapidly by using a machine learning method based on the big data of the existing material and using the deep neural network and combines an evolutionary algorithm to optimize characteristic quantities related to superconducting temperature, improves the prediction accuracy, can process a large amount of sample data, better predicts the superconducting performance of unknown materials and provides a basis for experiments and material selection.
Description
Technical Field
The invention relates to the technical field of superconducting materials, in particular to a method and a device for predicting a superconducting material based on a deep neural network algorithm.
Background
The material science has entered the development era of the fourth scientific paradigm of big data + artificial intelligence, and especially, the accurate prediction of the characteristics of the material before the synthesis of a novel material has very important scientific research and economic significance. Because of the characteristics of zero resistance, diamagnetism and the like, the superconducting material has wide application prospects in the fields of information, energy and communication, however, at present, few superconducting materials are suitable for commercial application.
In the field of superconducting materials, a great deal of experimental data has been accumulated in the last century, and the method is an excellent model for predicting novel materials by utilizing machine learning. At present, the method for predicting the novel superconducting material by utilizing machine learning mainly comprises regression analysis methods such as random forests, support vector machines and the like. Both methods are based on a mathematical least squares approach to reduce the error and thus predict new material data. Wherein, the former tends to saturate training errors and testing errors over a certain data volume (about several hundreds) as the sample data volume increases; the latter can only process small sample space, and can not perform induction analysis on large-scale data, so the latter is only used in iron-based superconducting materials, and has larger error.
Therefore, it is desirable to develop a method and apparatus for predicting new materials by performing machine learning on a large number of data samples.
Disclosure of Invention
In view of the problems in the prior art, the invention provides a method for predicting a novel superconducting material based on a deep neural network algorithm, which is characterized in that a deep neural network is established through existing superconducting material big data, the performance of the superconducting material is predicted by utilizing the established deep neural network, an evolutionary algorithm is adopted to optimize characteristic quantities related to superconducting temperature while the deep neural network is established, training errors and testing errors are optimized more rapidly, unknown superconducting materials can be well predicted, and a basis is provided for experiments and material selection.
In order to achieve the purpose, the invention adopts the following technical scheme:
in a first aspect, the present invention provides a method for predicting a superconducting material based on a deep neural network algorithm, the method comprising: and predicting the performance of the superconducting material by using the deep neural network.
Preferably, the method comprises the steps of:
(1) training and establishing a deep neural network according to the superconducting material sample set;
(2) and (3) predicting the performance of the superconducting material by using the deep neural network in the step (1).
Preferably, the superconducting material sample set in step (1) includes elemental data of the superconducting material sample.
Preferably, the elemental data of the superconducting material sample includes elemental properties.
The invention provides a method for predicting a novel superconducting material based on a deep neural network algorithm, wherein the deep neural network algorithm is more similar to a thinking mode of a human brain compared with a traditional machine learning method, and a superconducting material sample set is collected firstly. And then, establishing a deep neural network model by utilizing the model, wherein the model comprises an input layer, a hidden layer and an output layer. The input layer is made of superconducting material and characteristic quantity thereof; the function of the hidden layer is to map known features to unknown, complex quantities of features; the output layer is the superconducting temperature of the superconducting material, so that the performance of the superconducting material, such as the superconducting temperature, is predicted by using the deep neural network model; compared with the traditional regression analysis methods such as random forest and support vector machine, the method has the advantages of small test error, large sample space and the like.
The element properties comprise basic element properties determined by the molecular formula of the superconducting material, and specifically comprise atomic number, periodicity, s-valent electron number, total-valent electron number, unfilled f-electron number, absolute zero degree, lower space group, Mendeleev's number, family number, p-valent electron number, unfilled s-electron number, total unfilled electron number, energy band gap at absolute zero degree, atomic mass, covalent radius, d-valent electron number, unfilled p-electron number, specific volume at absolute zero degree, melting point, electronegativity, f-valent electron number, unfilled d-electron number and magnetic moment of each atom at absolute zero degree.
The source of the element property is not particularly limited, and any source capable of providing related data can be adopted, for example, related data provided by a Materials organization Platform for information and Exploration (Magpie) data Platform can be used.
Preferably, the elemental data of the superconducting material sample further includes: and (4) generating element characteristic quantities by using a dispersion square sum method according to element properties.
The characteristic quantities of the elements in the invention include the method of using the sum of squared deviations according to the properties of the elements6 calculated LPNorm, occupancy of 4 valence orbitals (s, p, d, f), ionic strength, and 132 element statistics (min, max, range, mean, absolute deviation, and mean absolute deviation), all or part, preferably all, of these characteristic quantities.
Preferably, the superconducting material sample set includes: the sample set of the conventional superconducting material, the sample set of the iron-based superconducting material or the sample set of the copper-based superconducting material or the combination of at least two of the two, wherein typical non-limiting combinations are the combination of the conventional superconducting material sample set and the iron-based superconducting material sample set, the combination of the conventional superconducting material sample set and the sample set of the copper-based superconducting material, the combination of the sample set of the copper-based superconducting material and the sample set of the iron-based superconducting material and the combination of the conventional superconducting material sample set, the sample set of the copper-based superconducting material and the sample set of the iron-based superconducting material.
When the combination of different types of material sample sets is adopted, the superconducting materials can be classified and then trained and learned respectively, and then different types of superconducting materials can be predicted respectively, or the superconducting materials can be directly mixed together for training and learning without classification, and the method is not limited specially.
The superconducting material sample set is already data, and related data which is already disclosed can be adopted, for example, 7002 superconducting materials with the superconducting temperature of more than 10K extracted from a superconducting material database of the Japan national Material science research institute can be adopted.
Because the properties of the copper-based superconducting material, the conventional superconducting material and the iron-based superconducting material are greatly different, the existing superconducting material data can be divided into three sample sets to construct a deep neural network respectively, and the number of superconductors in each sample set is more than 1000 to realize classification and respective prediction.
The conventional superconducting material in the invention refers to a superconducting material which can describe the superconducting behavior of the material by using Bardeen-Cooper-Schrieffer (BCS) theory or deduction thereof.
Preferably, the deep neural network in step (1) comprises: an input layer, a hidden layer, and an output layer.
The invention has no special limit to the number of input layers, hidden layers and output layers and the number of neurons in each layer, and the number of layers and the number of neurons of the neural network can be constructed according to the training and learning result.
Preferably, the deep neural network is established by adopting any one or a combination of at least two methods of a batch processing normalization method, a random gradient descent method, a grid search method or an evolutionary algorithm of MaxMin.
In the invention, the evolutionary algorithm is preferably adopted to construct the deep neural network, because in a plurality of element characteristic quantities, not all characteristic quantities have related influence or great influence on the superconducting property of the superconducting material, such as superconducting temperature, redundant and irrelevant characteristic quantities in the characteristic quantities can be removed through the evolutionary algorithm, so that important characteristic quantities are extracted by combining the deep neural network, and the redundant characteristic quantities can be eliminated after a plurality of iterations. Compared with the traditional machine learning method, the method can more rapidly optimize the training error and the testing error along with the increase of the sample space, improve the prediction accuracy and screen more key physical parameters.
The batch processing normalization method of MaxMin refers to dispersion normalization, which is linear transformation of original data, so that a result value is mapped between 0 and 1.
Preferably, the step (1) of establishing the deep neural network comprises the following steps:
(1') establishing neurons and neural networks of each layer;
(2') carrying out normalization processing on each layer by adopting a batch processing normalization method, determining an activation function of each layer, and searching parameters by adopting a first optimization method;
and (3') searching an optimal hyper-parameter by adopting a second optimization method to obtain the deep neural network.
In the invention, the scimit-spare software package matched with the Pythrch is preferably used for establishing the deep neural network model.
Preferably, the batch normalization method in step (2') is a batch normalization method of MaxMin.
Preferably, the activation function of each layer is a LeakyReLU function and/or a Log-Sigmoid function.
In the invention, the activation function of each layer can be the same or different, and preferably, a LeakyReLU function is adopted for each layer.
Preferably, the loss function for each layer is chosen to be the mean square error.
Preferably, the first optimization method comprises a random gradient descent method.
The invention preferably adopts a random gradient descent method to optimize parameters and searches the optimal specific gravity and bias of each layer.
Preferably, the parameters include parameters and parameter values.
Preferably, the parameter comprises a characteristic quantity associated with the superconducting temperature.
Preferably, the parameter values include a specific gravity and a bias of the characteristic amount.
Preferably, the second optimization method in step (3') includes a lattice search method.
Preferably, the hyper-parameters include the dimensionality of the hidden layer, the number of hidden layers, the learning rate, the number of training data, the size of each batch, and the iteration period.
In the invention, the searching performance of searching the hyper-parameters is optimized by preferably utilizing a grid searching method, and the hyper-parameters comprise neurons of each layer, the dimension of each hidden layer, the number of the hidden layers, the learning rate, the number of training data, the size of batch normalized data, the number of iterations and the like. With R2(value range is 0-1) to represent the learning effect of machine learning, R2The closer to 1, where R is2Is represented by formula (1):
wherein, yiIs an experimental value of the superconducting temperature of the material,
is the predicted superconducting temperature of the material,
is the average of all experimental values of the superconducting temperature of the material.
Preferably, between step (2 ') and step (3'), further comprising: and (5) screening parameters by using an evolutionary algorithm, and circularly iterating.
Preferably, the evolutionary algorithm is a differential evolutionary algorithm.
As a preferred technical scheme of the invention, the method comprises the following steps:
(1') dividing the superconducting material sample set into three different sample sets according to the superconducting material sample set, wherein the three different sample sets are respectively a conventional superconducting material sample set, an iron-based superconducting material sample set and a copper-based superconducting material sample set, and training the three sample sets respectively and establishing a deep neural network respectively; the superconducting material sample set comprises the element properties of the sample and element characteristic quantities generated according to the element properties by using a dispersion square sum method;
(2') establishing neurons of each layer and a neural network with adjustable activation functions;
(3') performing normalization processing on each layer by adopting a MaxMin batch processing normalization method, wherein an activation function of each layer is a LeakyReLU function and/or a Log-Sigmoid function, a loss function of each layer is a mean square error, and a random gradient descent method is adopted to search characteristic quantities related to superconducting temperature and proper specific gravity and bias thereof;
(4') the characteristic quantities related to the superconducting temperature, the specific gravity and the bias thereof are provided for a differential evolution algorithm, a combination of the characteristic quantities is randomly selected by the differential evolution algorithm, the combination of the characteristic quantities is crossed and varied on the basis of training and testing scores, and is circulated to the step (1) for multiple iterations to screen the characteristic quantities related to the superconducting temperature;
(5') searching for an optimal hyper-parameter by adopting a grid searching method to obtain a deep neural network;
(6 ') predicting the performance of the superconducting material using the deep neural network of step (5').
Preferably, the characteristic amount a in each layeri (j)Is as shown in equation (2):
wherein, wj (i)Is the specific gravity of the j parameter of the i layer, bj (i)Is the bias of the j parameter of the i layer, g is the activation function, i represents the number of layers, and j represents the subscript of the number of the neurons of each layer; a isj (i)A characteristic quantity representing the i-layer j parameter; i and j are natural numbers greater than or equal to 1 respectively, the input layer is recorded as the 1 st layer, i is equal to 1, and the output layer is recorded as the last layer, i is equal to imax(ii) a From the input layer XjTo aj (2)Also using a mapping method similar to equation (2), where XjCorresponding to a in formula (2)j (1)The rest is the same as the formula (2); last and output layers hθ(x) The mapping relation of (2) is also referred to, wherein hθ(x) Is equivalent to a(imax)。
Preferably, the differential evolution algorithm comprises the following steps:
step 1: initializing a characteristic population based on the characteristic quantity related to the superconducting temperature, the specific gravity and the bias thereof searched in the step (3') and determining a difference algorithm parameter;
step 2: inputting the population characteristics into a deep neural network model to judge whether the error and the number of the characteristics are the lowest, if so, determining the combination of the optimal characteristic quantities, otherwise, entering the step 3;
and step 3: calculating the fitness of individual characteristics in the population, and performing variation characteristic operation, cross characteristic operation and characteristic selection operation on the individual characteristics;
and 4, step 4: and (4) inputting the selected features obtained in the step (3) into the deep neural network model, and circulating to the step (2) to iterate until the combination of the optimal feature quantity is found.
In a second aspect, the present invention provides an apparatus for predicting a superconducting material based on a deep neural network algorithm, the apparatus comprising: the system comprises a sample set module, a deep neural network training and establishing module, a material input module and a superconducting material prediction output module;
the sample set module is used for storing or collecting sample data of the superconducting material;
the deep neural network training and establishing module is used for training and establishing a deep neural network model;
the material input module is used for inputting element data of the superconducting material to be predicted to the deep neural network training and establishing module;
and the superconducting material prediction output module is used for outputting superconducting material performance data obtained by predicting a deep neural network model in the deep neural network training and establishing module.
The device for predicting the superconducting material based on the deep neural network algorithm can learn and train the superconducting material sample set in the sample set module through the deep neural network training and establishing module, and can construct a reasonable deep neural network model, so that the performance of the superconducting material is predicted according to input data of the material input module, and finally predicted data is output from the superconducting material prediction output module.
In a third aspect, the present invention provides a computer-readable storage medium having stored thereon a computer program which, when being executed by a processor, implements the method for predicting a superconducting material based on a deep neural network algorithm according to the first aspect.
In a fourth aspect, the present invention provides an electronic device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method for predicting a superconducting material based on a deep neural network algorithm according to the first aspect when executing the computer program.
Compared with the prior art, the invention has at least the following beneficial effects:
(1) according to the method for predicting the superconducting material based on the deep neural network algorithm, the deep neural network is established through the existing large data of the superconducting material, the performance of the superconducting material is predicted by utilizing the established deep neural network, the sample handling capacity is large, and the prediction result is accurate;
(2) the method for predicting the superconducting material based on the deep neural network algorithm combines the evolutionary algorithm with the deep neural network, removes the characteristic quantity which has low or no correlation with the superconducting temperature, greatly improves the training speed and reduces the testing accuracy;
(3) the method for predicting the superconducting material based on the deep neural network algorithm preferentially divides the existing superconducting material sample set into three classes, respectively trains aiming at materials with different properties, and improves the effectiveness of the training and the accuracy of the final prediction result.
Drawings
Fig. 1 is a general flowchart of a method for predicting a superconducting material based on a deep neural network algorithm provided in example 1.
Fig. 2 is a flowchart of the construction of the deep neural network in the method for predicting the superconducting material based on the deep neural network algorithm provided in embodiment 1.
Fig. 3 is a schematic flowchart of a differential evolution algorithm in the method for predicting a superconducting material based on a deep neural network algorithm provided in embodiment 1.
FIG. 4 is a schematic diagram of a deep neural network of the iron-based superconducting material and the conventional superconducting material constructed in example 1.
FIG. 5 is a schematic view of a deep neural network of the copper-based superconducting material constructed in example 1.
FIG. 6 is a schematic diagram of a deep neural network of the hybrid superconducting material constructed in example 1.
FIG. 7 is a graph showing the number of training samples as a function of the number of training samples for the conventional superconducting materials in example 1 and comparative example 1.
FIG. 8 is a graph showing the number of training samples as a function of the number of training samples for the iron-based superconducting materials in example 1 and comparative example 1.
FIG. 9 is a graph of the number of training samples as a function of the number of training samples for copper-based superconducting materials in example 1 and comparative example 1.
FIG. 10 is a graph of experimental values versus the results of the training and prediction for the hybrid superconducting material of example 2.
Detailed Description
The technical scheme of the invention is further explained by the specific implementation mode in combination with the attached drawings.
The present invention is described in further detail below. The following examples are merely illustrative of the present invention and do not represent or limit the scope of the claims, which are defined by the claims.
First, an embodiment
Example 1
The embodiment provides a method for predicting a superconducting material based on a deep neural network algorithm, as shown in fig. 1, the method includes the following steps:
(1) establishing a superconducting material sample set according to existing superconducting material data which are derived from a superconducting material database of the national Material science research institute, and according to related element data provided by a Materials analytical Platform for information and expression (Magpie) data Platform, wherein the superconducting material sample set comprises the element properties of the sample and element characteristic quantities generated according to the element properties by using a dispersion square sum method; dividing the superconducting material sample set into three different sample sets, namely a conventional superconducting material sample set, an iron-based superconducting material sample set and a copper-based superconducting material sample set, respectively training the three sample sets and respectively establishing a deep neural network;
the deep neural network construction steps are shown in fig. 2, and specifically include:
firstly, establishing neurons of each layer and a neural network with adjustable activation functions;
secondly, performing normalization processing on each layer by adopting a MaxMin batch processing normalization method, selecting a LeakyReLU function as an activation function of each layer, selecting a mean square error as a loss function of each layer, and searching characteristic quantity related to superconducting temperature and proper specific gravity and bias thereof by adopting a random gradient descent method;
thirdly, providing the characteristic quantity associated with the superconducting temperature, the specific gravity and the bias thereof to a differential evolution algorithm, randomly selecting a plurality of combinations of the characteristic quantity by using the differential evolution algorithm, carrying out crossing and variation on the combinations on the basis of larger training and testing scores, circulating the combinations to the first step, carrying out N times (N >2000) of iteration, screening the characteristic quantity associated with the superconducting temperature, and removing redundant and irrelevant characteristic quantities;
fourthly, searching for an optimal hyper-parameter by adopting a grid search method to obtain a deep neural network;
the hyper-parameters comprise dimension of each hidden layer, number of hidden layers, training rate, batch normalization data size, iteration number and the like, and R is used2(0-1) to characterize the learning effect of machine learning, R2The closer to 1, the better the machine learning effect;
(2) and (3) predicting the performance of the superconducting material by using the deep neural network in the step (1), and screening out the superconducting material with the superconducting temperature higher than 10K.
A specific flowchart of the differential evolution algorithm in embodiment 1 is shown in fig. 3, and specifically includes the following steps:
step 1: initializing a characteristic population based on the characteristic quantity related to the superconducting temperature searched in the second step and the specific gravity and bias thereof, and determining a difference algorithm parameter;
step 2: inputting the population characteristics into a deep neural network model to judge whether the error and the number of the characteristics are the lowest, if so, determining the combination of the optimal characteristic quantities, otherwise, entering the step 3;
and step 3: calculating the fitness of individual characteristics in the population, and performing variation characteristic operation, cross characteristic operation and characteristic selection operation on the individual characteristics;
and 4, step 4: and (4) inputting the selected features obtained in the step (3) into the deep neural network model, and circulating to the step (2) to iterate until the combination of the optimal feature quantity is found.
The neural network constructed based on the iron-based superconducting material sample set and the conventional superconducting material sample set in the embodiment 1 is shown in fig. 4, and as can be seen from the figure, the deep neural network comprises an input layer, a hidden layer and an output layer; a deep neural network constructed based on a sample set of copper-based superconducting materials is shown in fig. 5, and as can be seen from the figure, the deep neural network includes an input layer and a layer,hidden layer three layers and output layer one layer, in FIGS. 4 and 5, ai (j)Is as shown in equation (2):
wherein, wj (i)Is the specific gravity of the j parameter of the i layer, bj (i)Is the bias of the i layer j parameter, g is the activation function, i represents the number of layers, j represents the index of the number of neurons in each layer, XjRepresenting the superconducting material and its characteristic quantities, constituting the input layer; a isj (i)The characteristic quantity representing the parameter of the i layer j, i and j are natural numbers which are respectively more than or equal to 1, the input layer is marked as the 1 st layer, i is equal to 1, and the output layer is marked as the last layer, i is equal to imax(ii) a From the input layer XjTo aj (2)Also using a mapping method similar to equation (2), where XjCorresponding to a in formula (2)j (1)The rest is the same as the formula (2); last and output layers hθ(x) The mapping relation of (2) is also referred to, wherein hθ(x) Is equivalent to a(imax). In FIG. 4, the known characteristic quantity a is represented by the hidden layerj (2)Mapping to unknown, complex characteristic quantities aj (3)(ii) a In FIG. 5, aj (3)Will also map to unknown, more complex characteristic quantities aj (4)Superconducting temperature h of superconducting materialθ(x) Forming an output layer.
Example 2
The embodiment provides a method for predicting a superconducting material based on a deep neural network algorithm, which is the same as that in embodiment 1 except that a superconducting material sample set is not classified, training is directly performed on a mixed sample set, and a deep neural network is established.
The neural network constructed based on training and learning of the mixed superconducting material sample set in the embodiment 2 is shown in fig. 6, and as can be seen from the figure, the deep neural network comprises an input layer, an implied layer and an output layer, and the mapping relationship between each layer refers to the embodiment 1.
Example 3
The embodiment provides an apparatus for predicting a superconducting material based on a deep neural network algorithm, which may be implemented by hardware or software, and the apparatus includes:
the system comprises a sample set module, a deep neural network training and establishing module, a material input module and a superconducting material prediction output module;
the sample set module is used for storing or collecting sample data of the superconducting material;
the deep neural network training and establishing module is used for training and establishing a deep neural network model;
the material input module is used for inputting element data of the superconducting material to be predicted to the deep neural network training and establishing module;
and the superconducting material prediction output module is used for outputting superconducting material performance data obtained by predicting a deep neural network model in the deep neural network training and establishing module.
Example 4
Embodiments of the present invention also provide a computer-readable storage medium, where the computer-executable instructions are executed by a computer processor to perform the method for predicting a superconducting material based on a deep neural network algorithm provided by an embodiment of the present invention. Storage media refers to any of various types of memory devices or storage devices. The term "storage medium" is intended to include: mounting media such as CD-ROM, floppy disk, or tape devices; in addition, the storage medium may be located in a first computer system in which the program is executed or may be located in a second, different computer system, the second computer system being connected to the first computer system via a network (such as the Internet) Implemented as a computer program).
Example 5
The embodiment of the invention also provides electronic equipment, wherein the electronic equipment is a computer, a computer system or an electronic machine capable of performing calculation based on the method for predicting the superconducting material by the deep neural network algorithm, and the like, the electronic equipment comprises a memory, a processor and a computer program which is stored on the memory and can be run on the processor, and the processor realizes the method for predicting the superconducting material based on the deep neural network algorithm when executing the computer program.
Second, comparative example
Comparative example 1
This comparative example provides a method for predicting superconducting Materials based on a random forest algorithm using the random forest method described in "mechanical learning modulation of superconducting detailed temperature", Valentinstainov, et al, NPJ computerized Materials, (2018)4:29.
Third, application and results
Respectively carrying out sample training on a conventional superconducting material sample set, an iron-based superconducting material sample set and a copper-based superconducting material sample set by using the methods provided in example 1 and comparative example 1, and using R2(0-1) to represent the training learning effect of machine learning, and to compare R2The relationship with the number of samples is plotted as shown in FIGS. 7 to 9, respectively, and it is apparent from FIGS. 7 to 9 that the training method provided in example 1 is significantly better than the method provided in comparative example 1 in the case where the number of samples reaches about 300 for a conventional superconducting material body and an iron-based superconductor on the basis of a large number of samples, where R is R2The predicted value is closer to 1, and the result of the predicted value is closer to the result of the actual experiment value; for copper-based superconductors, the deep neural network method provided by the embodiment 1 is always superior to the random forest method provided by the comparative example 1 in test result, and therefore, the deep neural network prediction method provided by the invention is shownThe method for the performance of the superconducting material is more suitable for a large number of data samples, and the prediction accuracy is high.
At the same time, the invention trains the machine to R using the method provided in example 12And (3) predicting a plurality of known materials and unknown materials after the temperature is more than or equal to 0.9, wherein the prediction results are respectively shown in the table 1 and the table 2.
TABLE 1
As can be seen from table 1, the method for predicting a superconducting material based on a deep neural network algorithm provided in example 1 has a better prediction effect for the conventional superconductor, the copper-based superconductor and the iron-based superconductor, is closer to an experimental value, has an absolute error of less than or equal to 3.38K for the prediction of the material, and has high prediction accuracy.
TABLE 2
| Chemical formula (II) | Crystal structure | Superconducting temperature (K) | Chemical formula (II) | Crystal structure | Superconducting temperature (K) |
| Li(TiTe2)5 | Three oblique lines | 36.46 | LiCa2Ga | Cube | 30.90 |
| Li(YGe)4 | Orthogonal | 36.26 | K23Na8(CdIn4)12 | Hexagonal shape | 30.80 |
| K3(Mg10In7)2 | Cube | 35.44 | CsTi2(CuTe4)2 | Monocline | 30.67 |
| Rb(Zr3Te4)4 | Hexagonal shape | 35.03 | Li(YbGe)4 | Orthogonal | 30.49 |
| K(Zr3Te4)4 | Three oblique lines | 34.99 | Li(CeGe)4 | Orthogonal | 29.99 |
| Na(SrAs)4 | Orthogonal | 33.51 | LiCa2In | Cube | 29.79 |
| ZrZn2Ge | Square block | 32.74 | Sr2LiTl | Cube | 29.71 |
| Li(TiTe2)2 | Monocline | 32.10 | NaCa2In | Cube | 29.65 |
| Ba9In4H | Square block | 32.08 | Mg(ScGa)2 | Square block | 29.39 |
| LiCa2Ge | Cube | 31.58 | TiAl2Zn | Square block | 29.12 |
As can be seen from table 2: the method provided by the invention can predict the superconducting temperature of different materials, so that the experiment can be well guided to be carried out, the superconducting material with the superconducting temperature higher than 10K can be screened out, and a good foundation is provided for material screening and experiments.
The method provided by embodiment 2 is used for carrying out sample training on a mixed superconducting material sample set, and predicting the superconducting temperature of a known material by using a trained model, wherein a relation graph of the superconducting temperature and the experimental temperature obtained in the training and predicting process is shown in FIG. 10, the results of the training and predicting process without classification are also very good, wherein R in the training process2Can reach 0.98, R is adopted2R between predicted value and experimental value when prediction is carried out on the deep neural network model after 0.982The prediction result is accurate up to 0.94.
In summary, compared with the existing random forest method, the method for predicting the superconducting material based on the deep neural network algorithm provided by the invention is based on the big data of the existing material and utilizes the machine learning method of the deep neural network, the test error is small, the support vector machine method is difficult to process the large-scale data, and the error is relatively larger; by combining the differential evolution algorithm on the basis of the deep neural network to optimize the characteristic quantity related to the superconducting temperature, the training error and the testing error are optimized more quickly, the prediction accuracy is improved, a large amount of sample data can be processed, the characteristics of the material are predicted accurately, and a foundation is provided for experiments and material screening.
The applicant declares that the present invention illustrates the detailed structural features of the present invention through the above embodiments, but the present invention is not limited to the above detailed structural features, that is, it does not mean that the present invention must be implemented depending on the above detailed structural features. It should be understood by those skilled in the art that any modifications of the present invention, equivalent substitutions of selected components of the present invention, additions of auxiliary components, selection of specific modes, etc., are within the scope and disclosure of the present invention.
Claims (10)
1. A method for predicting superconducting materials based on a deep neural network algorithm, the method comprising: and predicting the performance of the superconducting material by using the deep neural network.
2. Method according to claim 1, characterized in that it comprises the following steps:
(1) training and establishing a deep neural network according to the superconducting material sample set;
(2) and (3) predicting the performance of the superconducting material by using the deep neural network in the step (1).
3. The method according to claim 2, wherein the set of superconducting material samples in step (1) includes elemental data of the superconducting material samples;
preferably, the elemental data of the superconducting material sample includes elemental properties;
preferably, the elemental data of the superconducting material sample further includes: according to the element property, the element characteristic quantity is generated by using a dispersion square sum method;
preferably, the superconducting material sample set includes: any one or a combination of at least two of the conventional superconducting material sample set, the iron-based superconducting material sample set or the copper-based superconducting material sample set.
4. The method of claim 2 or 3, wherein the deep neural network in step (1) comprises: an input layer, a hidden layer and an output layer;
preferably, the deep neural network is established by adopting any one or a combination of at least two methods of a batch processing normalization method, a random gradient descent method, a grid search method or an evolutionary algorithm of MaxMin.
5. The method according to claim 2 or 3, wherein the establishing of the deep neural network in step (1) comprises the steps of:
(1') establishing neurons and neural networks of each layer;
(2') carrying out normalization processing on each layer by adopting a batch processing normalization method, determining an activation function of each layer, and searching parameters by adopting a first optimization method;
and (3') searching an optimal hyper-parameter by adopting a second optimization method to obtain the deep neural network.
6. The method of claim 5, wherein the batch normalization method in step (2') is a batch normalization method of MaxMin;
preferably, the activation function of each layer is a LeakyReLU function and/or a Log-Sigmoid function, and preferably a LeakyReLU function;
preferably, the loss function of each layer is selected from mean square error;
preferably, the first optimization method comprises a random gradient descent method;
preferably, the second optimization method in step (3') includes a lattice search method.
7. The method of claim 5, further comprising, between step (2 ') and step (3'): screening parameters by using an evolutionary algorithm, and circularly iterating;
preferably, the evolutionary algorithm is a differential evolutionary algorithm.
8. An apparatus for predicting a superconducting material based on a deep neural network algorithm, the apparatus comprising: the system comprises a sample set module, a deep neural network training and establishing module, a material input module and a superconducting material prediction output module;
the sample set module is used for storing or collecting sample data of the superconducting material;
the deep neural network training and establishing module is used for training and establishing a deep neural network model;
the material input module is used for inputting element data of the superconducting material to be predicted to the deep neural network training and establishing module;
and the superconducting material prediction output module is used for outputting superconducting material performance data obtained by predicting a deep neural network model in the deep neural network training and establishing module.
9. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, implements the method for predicting a superconducting material based on a deep neural network algorithm according to any one of claims 1 to 7.
10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method for predicting a superconducting material based on a deep neural network algorithm according to any one of claims 1 to 7 when executing the computer program.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010600263.1A CN111798940A (en) | 2020-06-28 | 2020-06-28 | Method and device for predicting superconducting material based on deep neural network algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010600263.1A CN111798940A (en) | 2020-06-28 | 2020-06-28 | Method and device for predicting superconducting material based on deep neural network algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111798940A true CN111798940A (en) | 2020-10-20 |
Family
ID=72804299
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010600263.1A Pending CN111798940A (en) | 2020-06-28 | 2020-06-28 | Method and device for predicting superconducting material based on deep neural network algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111798940A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112349357A (en) * | 2020-11-05 | 2021-02-09 | 大连英达士智能科技有限公司 | Long-term batching method and system for production of non-ferrous copper |
| CN115035966A (en) * | 2022-08-09 | 2022-09-09 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Superconductor screening method, device and equipment based on active learning and symbolic regression |
| WO2024011979A1 (en) * | 2022-07-11 | 2024-01-18 | 浙江大学杭州国际科创中心 | Pre-trained deep neural network based superconducting material prediction method and apparatus |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| FR2688080A1 (en) * | 1992-02-28 | 1993-09-03 | Air Liquide | Analog neural network |
| CN104965969A (en) * | 2015-04-20 | 2015-10-07 | 广西大学 | Inversion method for surrounding rock mechanical parameters of large cavern group |
| US20170024642A1 (en) * | 2015-03-13 | 2017-01-26 | Deep Genomics Incorporated | System and method for training neural networks |
| CN108985376A (en) * | 2018-07-17 | 2018-12-11 | 东北大学 | It is a kind of based on convolution-Recognition with Recurrent Neural Network rotary kiln sequence operating mode's switch method |
| CN109101638A (en) * | 2018-08-20 | 2018-12-28 | 河海大学 | A kind of Dam Deformation Monitoring continuity missing data complementing method |
| KR20190081218A (en) * | 2017-12-29 | 2019-07-09 | 중앙대학교 산학협력단 | Terminal device and Method for setting hyperparameter of convolutional neural network |
| CN110443364A (en) * | 2019-06-21 | 2019-11-12 | 深圳大学 | A kind of deep neural network multitask hyperparameter optimization method and device |
| US20200104715A1 (en) * | 2018-09-28 | 2020-04-02 | Xilinx, Inc. | Training of neural networks by including implementation cost as an objective |
-
2020
- 2020-06-28 CN CN202010600263.1A patent/CN111798940A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| FR2688080A1 (en) * | 1992-02-28 | 1993-09-03 | Air Liquide | Analog neural network |
| US20170024642A1 (en) * | 2015-03-13 | 2017-01-26 | Deep Genomics Incorporated | System and method for training neural networks |
| CN104965969A (en) * | 2015-04-20 | 2015-10-07 | 广西大学 | Inversion method for surrounding rock mechanical parameters of large cavern group |
| KR20190081218A (en) * | 2017-12-29 | 2019-07-09 | 중앙대학교 산학협력단 | Terminal device and Method for setting hyperparameter of convolutional neural network |
| CN108985376A (en) * | 2018-07-17 | 2018-12-11 | 东北大学 | It is a kind of based on convolution-Recognition with Recurrent Neural Network rotary kiln sequence operating mode's switch method |
| CN109101638A (en) * | 2018-08-20 | 2018-12-28 | 河海大学 | A kind of Dam Deformation Monitoring continuity missing data complementing method |
| US20200104715A1 (en) * | 2018-09-28 | 2020-04-02 | Xilinx, Inc. | Training of neural networks by including implementation cost as an objective |
| CN110443364A (en) * | 2019-06-21 | 2019-11-12 | 深圳大学 | A kind of deep neural network multitask hyperparameter optimization method and device |
Non-Patent Citations (2)
| Title |
|---|
| CAMILA MARTINS SAPORETTI ET AL.: "Extreme Learning Machine combined with a Differential Evolution algorithm for lithology identification", 《REVISTA DE INFORM´ATICA TE ´ORICA E APLICADA》, vol. 25, no. 4, 19 November 2018 (2018-11-19) * |
| THANH DUNG LE ET AL.: "Critical Temperature Prediction for a Superconductor: A Variational Bayesian Neural Network Approach", 《IEEE TRANSACTIONS ON APPLIED SUPERCONDUCTIVITY》, vol. 30, no. 4, pages 1 - 5, XP011774846, DOI: 10.1109/TASC.2020.2971456 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112349357A (en) * | 2020-11-05 | 2021-02-09 | 大连英达士智能科技有限公司 | Long-term batching method and system for production of non-ferrous copper |
| CN112349357B (en) * | 2020-11-05 | 2021-07-30 | 大连英达士智能科技有限公司 | Long-term batching method and system for production of non-ferrous copper |
| WO2024011979A1 (en) * | 2022-07-11 | 2024-01-18 | 浙江大学杭州国际科创中心 | Pre-trained deep neural network based superconducting material prediction method and apparatus |
| CN115035966A (en) * | 2022-08-09 | 2022-09-09 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Superconductor screening method, device and equipment based on active learning and symbolic regression |
| CN115035966B (en) * | 2022-08-09 | 2022-11-04 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Superconductor screening method, device and equipment based on active learning and symbolic regression |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Young et al. | Optimizing deep learning hyper-parameters through an evolutionary algorithm | |
| CN111798940A (en) | Method and device for predicting superconducting material based on deep neural network algorithm | |
| CN108985335B (en) | Integrated learning prediction method for irradiation swelling of nuclear reactor cladding material | |
| CN109242223B (en) | Quantum support vector machine evaluation and prediction method for urban public building fire risk | |
| CN113256066A (en) | PCA-XGboost-IRF-based job shop real-time scheduling method | |
| CN112382352A (en) | Method for quickly evaluating structural characteristics of metal organic framework material based on machine learning | |
| CN109241278A (en) | Scientific research knowledge management method and system | |
| Singh et al. | Performance Measure of Similis and FPGrowth Algo rithm | |
| Huang et al. | Multi-objective multi-generation Gaussian process optimizer for design optimization | |
| Wang et al. | Objective-hierarchy based large-scale evolutionary algorithm for improving joint sparsity-compression of neural network | |
| Weiguo et al. | Variable precision rough set based decision tree classifier | |
| CN109840551B (en) | Method for optimizing random forest parameters for machine learning model training | |
| Kumar et al. | Particle swarm optimization: a study of variants and their applications | |
| Wang et al. | Kernel-based fuzzy C-means clustering based on fruit fly optimization algorithm | |
| CN114021425A (en) | Power system operation data modeling and feature selection method and device, electronic equipment and storage medium | |
| Ahmed et al. | Predicting and analysis of students’ academic performance using data mining techniques | |
| Zhang et al. | BP-PSO-based intelligent case retrieval method for high-rise structural form selection | |
| CN117114184A (en) | Urban carbon emission influence factor feature extraction and medium-long-term prediction method and device | |
| Guo | Application of computer technology in optimal design of overall structure of special machinery | |
| CN115035966B (en) | Superconductor screening method, device and equipment based on active learning and symbolic regression | |
| CN111523573A (en) | Bridge structure state evaluation method and system based on multi-parameter fusion | |
| CN107766887A (en) | A kind of local weighted deficiency of data mixes clustering method | |
| Guimaraes et al. | Local learning and search in memetic algorithms | |
| Fu et al. | Self-stacking random weight neural network with multi-layer features fusion | |
| Wu | Evaluation model of product shape design scheme based on fuzzy genetic algorithm mining spatial association rules |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |