A kind of solid electrolyte ionic conductivity prediction technique based on BP neural network
Technical field
The present invention relates to artificial intelligence fields, specifically provide a kind of solid electrolyte ionic conductance based on BP neural network
Rate prediction technique.
Background technique
For lithium ion battery because of its energy density height, multiplying power property is good and is widely applied.But because electrolyte is there are security risk,
Use scope is subject to certain restrictions.Solid electrolyte material can inhibit the growth of Li dendrite at cathode, and have incombustible
Characteristic, can fundamentally solve lithium battery using security issues, be next-generation lithium battery critical material.Traditional material development
It is main to use " trial and error " experimental method, a large amount of iteration experiment is carried out in the way of " proposing hypothesis-experimental verification ", from
And experimental material is made constantly to approach target material.But this method efficiency is lower, new material development rate is far behind new
Product development speed.Since 20th century, a kind of new material needs about 10 years to practical application since research and development, has been unable to satisfy
Demand of the fields such as lithium battery to new material.
With occurring predicting the technology of material properties using machine learning in recent years, such as closed using SVM prediction
Gold forms glass ability, using conductive capability of logistic regression prediction solid electrolyte etc., has obtained first-stage success.But
Being that above-mentioned technology is appointed has following defects that one, when data volume is very big, and support vector machines training is to hardware device
Demand is usually excessive.Two, Logic Regression Models are easy poor fitting, and nicety of grading may not be high.Three, both the above mathematical model is equal
It is only suitable for predicting specific material properties, does not have general applicability.
Therefore, one kind is found to be able to achieve to multi-class data, the blanket high rate of precision material forecasting model of different hardware
It is the task of top priority.
Summary of the invention
It is an object of the invention to overcome drawbacks described above existing for existing solid electrolyte ionic conductivity prediction technique,
A kind of solid electrolyte ionic conductivity prediction technique based on BP neural network is provided, comprising the following steps:
Step 1: acquisition data, total sample required for BP neural network model is obtained after pretreatment, by total sample set with
The ratio random division of 2:8 is test set sample and training set sample;
It is wherein, pretreated that steps are as follows:
(1.1) chemical formula, atomic coordinates, lattice lengths, volume field are extracted in every group of data of acquisition as input
Parameter;
(1.2) go out 10 features of every group of data according to obtained input Parameters Calculation;10 features include average original
Sub-volume, lithium ion close on the standard deviation of atom number, lithium ion key standard deviation, lithium ion key average value, and lithium ion faces
The average value of nearly atom number, the average chemical key of sublattice is ionic, and average sublattice adjacent atom number, anion structure is matched
Digit, anion volume, and average minimum lithium ion spacing distance;
Step 2: BP neural network model is constructed according to step 1;
Constructing BP neural network model construction, steps are as follows:
(2.1) input layer and output layer are constructed, and calculates the neuron number for forming input layer and output layer;Its
In, the neuron number of input layer and output layer, which calculates, uses formula are as follows:
Tout=f (Tin1, Tin2 ..., Tinn)
In formula: Tout is the data value that neural network needs to predict, Tin1~Tinn is that neural network input layer is defeated respectively
The n data value entered;The neuron of input layer is set as 10 according to the formula, output layer neuron is set as 2;
(2.2) hidden layer is constructed, and hidden layer includes the first hidden layer and the second hidden layer;It calculates hidden for forming first
The neuron number of neuron number containing layer and the second hidden layer, hidden layer is determined using following formula:
In formula, h is hidden layer neuron number, and m is input layer number, and n is output layer neuron number, a 1
Neuron constant between~10;The neuron of the first hidden layer is set as 64 according to the formula, the mind of the second hidden layer
It is set as 32 through member;
(2.3) enter layer, hidden layer, output layer and every layer of neuron according to what step (2.1) and step (2.2) obtained
Number, and by formula Y=Sigmoid [W2*Sigmoid (W1*X-O1)-O2], obtain BP neural network model;
In formula: X is BP neural network input matrix;Y is BP neural network output matrix;W1 and W2 is respectively BP nerve net
Input layer is to hidden layer, the connection weight matrix of hidden layer to output layer in network;O1, O2 be in BP neural network input layer to hidden
The threshold matrix of layer, hidden layer output layer, Sigmoid are each layer excitation function;
Step 3: transmission function, training function and learning function are selected in BP neural network model, and to BP nerve net
Parameters are initialized in network model;
Step 4: choosing learning parameter in initiation parameter, and learning parameter is initialized using Gaussian Profile, obtain
Initial weight and deviation;Wherein, the learning parameter include learning rate, threshold value, frequency of training, training pace, activation primitive,
And learning algorithm;
Step 5: BP neural network model is trained using existing " input-output " training set sample data, and
The training set for the neural network model established is fitted using quadratic function, obtains optimal prediction network;Its quadratic function
Fitting formula is as follows:
In formula: WmaxFor initial training collection, WminFor final training set, KmaxFor maximum number of iterations, k is current iteration time
Number;
Step 6: whether training of judgement collection degree of fitting achieves the desired results;It is no, rebuild BP neural network model;It is,
Execute step 7;
Step 7: being carried out using degree of fitting of existing " input-output " the test set sample data to BP neural network model
Test;
Step 8: inputting sample to be predicted in the BP neural network model after degree of fitting is tested and predicted, obtained
Solid electrolyte ionic conductivity.
Further, 643 groups of material datas are acquired in the step 1, every group of data include 10 features, 1 label;It is right
643 groups of material datas obtain 643 groups of total sample sets after being normalized, 624 groups of total sample sets are random with the ratio of 2:8
It is divided into test set sample and training set sample;
The parameters of BP neural network model include: using sigmoid as activation primitive, cross entropy in the step 3
As loss function, gradient descent method as loss function.
Further, the learning rate in the step 4 is set as 0.0003, and threshold value is set as 0.5, frequency of training and sets
Be set to 2500 times, training pace be set as 100, learning algorithm is declined using cross entropy and gradient.
The degree of fitting of BP neural network model is tested specifically includes the following steps: by test set in the step 6
Sample is as input sample, and network layer connection weight and neuron threshold values condition after the training adjustment by step 5
Under, by forward-propagating, being calculated by the-the second hidden layer of the-the first hidden layer of input layer-output layer, output valve is all label,
The reality output label value and the label value in training sample for comparing output layer, can calculate trained BP neural network mould
Degree of fitting of the type to test set sample.
The present invention compared with the prior art, have the following advantages that and the utility model has the advantages that
The present invention can accurately predict the solid electrolyte ionic conductivity performance containing elements such as Li, Na, Mg, Al, then
According to prediction as a result, being prepared to its corresponding ingredient and structure, it is possible to reduce the blindness in electrode material design process
Property, save a large amount of time and cost.
Detailed description of the invention
Fig. 1 is artificial neural network structure's schematic diagram of the invention.
Fig. 2 is method flow schematic diagram of the invention.
Specific embodiment
The present invention is described in further detail below with reference to embodiment and its attached drawing, but embodiments of the present invention are simultaneously
It is without being limited thereto.
Embodiment
As shown in Fig. 2, a kind of solid electrolyte ionic conductivity prediction technique based on BP neural network of the invention, packet
Include following steps:
Step 1: 643 groups of material datas are acquired from material database, every group of data include 10 features, 1 label;It is right
643 groups of material datas be normalized pretreatment after obtain 643 groups of total sample sets, by 624 groups of total sample sets with the ratio of 2:8 with
Machine is divided into test set sample and training set sample;
Specifically, one after the material data of acquisition changes pretreatment including extracting chemical formula in every group of data, atom seat
Mark, lattice lengths, volume field are used as input parameter, and go out 10 features of every group of data according to input Parameters Calculation;This 10
A feature includes Average atomic volumes, and lithium ion closes on the standard deviation of atom number, lithium ion key standard deviation, lithium ion key
Average value, lithium ion close on the average value of atom number, and the average chemical key of sublattice is ionic, average sublattice adjacent atom
Number, anion structure ligancy, anion volume, and average minimum lithium ion spacing distance.Wherein, 10 features are as people
The input of artificial neural networks, label value are then the output of artificial neural network.
Step 2: BP neural network model is constructed according to step 1;
Specifically, as shown in Figure 1, building input layer and output layer, and the mind of determining composition input layer and output layer first
Through first number;Wherein, the neuron number of input layer and output layer, which calculates, uses formula are as follows: Tout=f (Tin1, Tin2 ...,
Tinn) in the formula: Tout is the data value that neural network needs to predict, Tin1~Tinn is neural network input layer input respectively
N data value.Wherein, by pre-training, compare BP neural network convergence precision and convergence rate, determine optimal imply
Layer neuronal quantity.The input layer includes 10 neurons of 10 features input, and output layer includes 2 nerves of label value
Member.
Secondly building hidden layer, and hidden layer is set as two layers, i.e. hidden layer includes that the first hidden layer and second are implicit
Layer, composition includes the first hidden layer and the neuron number of the second hidden layer uses formulaTo determine: the formula
In, h is hidden layer neuron number, and m is input layer number, and n is output layer neuron number, and a is between 1~10
Neuron constant.Wherein, by pre-training, compare BP neural network convergence precision and convergence rate, determine optimal imply
Layer neuronal quantity.As shown in Figure 1, wherein the neuron of the first hidden layer is set as 64, and the neuron of the second hidden layer is set
It is set to 32.
Finally, according to obtained input layer, hidden layer, output layer and every layer of neuron number, and pass through formula Y=
Sigmoid [W2*Sigmoid (W1*X-O1)-O2], obtains BP neural network model;In the formula: X is that BP neural network inputs square
Battle array;Y is BP neural network output matrix;W1 and W2 be respectively in BP neural network input layer to hidden layer, hidden layer to output layer
Connection weight matrix;O1, O2 are input layer in BP neural network to hidden layer, the threshold matrix of hidden layer output layer, Sigmoid
For each layer excitation function.
Step 3: transmission function, training function and learning function are selected in BP neural network model, and to BP nerve net
Parameters are initialized in network model.Wherein, the parameters of BP neural network model include: using sigmoid as sharp
Work function, cross entropy are as loss function, gradient descent method as loss function.
Step 4: after creating BP neural network model, activation primitive appropriate and study need to be selected to calculate for the network model
Method, to activate study, feedback and the final forecast function of whole network, it is therefore desirable to initialize each of BP neural network model
Item parameter, and learning parameter is chosen in initiation parameter.Wherein, the learning parameter includes learning rate, threshold value, training time
Number, training pace, activation primitive and learning algorithm.Specifically, the present embodiment is using sigmoid function as activation primitive, friendship
Entropy is pitched as loss function, gradient descent method as loss function, weight and deviation, learning rate are initialized using Gaussian Profile
Be set as 0.0003, threshold value be set as 0.5, frequency of training be set as 2500 times, training pace be set as 100.
Step 5: BP neural network model is trained using existing " input-output " training set sample data, and
The training set for the neural network model established is fitted by quadratic function, obtains optimal prediction network.Wherein, secondary letter
Number fitting formula are as follows:In formula: WmaxFor initial training collection, WminFinally to train
Collection, KmaxFor maximum number of iterations, k is current iteration number.Wherein initial inertia weight wmax=0.6, final inertia weight
Wmin=0.8, maximum number of iterations kmax=2000.
Wherein, existing " input-output " is utilized after BP neural network model buildings are good and parameters initialize
Training set sample data is trained BP neural network model, obtains optimal prediction network.The instruction of BP neural network model
Practice process, is that the variable forward direction calculating of input is constantly repeatedly adjusted with the weight matrix in layer each when error back propagation, makes
The process that degree of fitting is higher than precision prescribed is obtained, specific training method is as follows:
The input sample that training pace number is randomly selected in training set sample data, the ginseng initialized in step 3
Under said conditions, by forward-propagating when initial, calculated by the-the second hidden layer of the-the first hidden layer of input layer-output layer, it is defeated
Value is all label out, and the reality output label value and the label value in training set sample for comparing output layer can calculate the secondary instruction
Experienced artificial nerve network model judges whether network degree of fitting meets the requirements to the degree of fitting of training set sample;Work as fitting
When degree reaches the maximum times of default precision or frequency of training greater than setting, then terminate to train;When the not up to default essence of degree of fitting
When spending, error back propagation is adjusted weight according to the learning rate of setting by system, and repeats input sample and counted
It calculates, the maximum frequency of training until meeting the requirements or reaching setting.In this example, BP neural network model training have passed through
After 2400 times, reach 97% or more training set degree of fitting, illustrate artificial nerve network model fast convergence rate of the invention,
Training effect is good.
Step 6: whether training of judgement collection degree of fitting achieves the desired results, and the training set degree of fitting in the present invention reaches 95%
Left and right;If training set sample data is not up to 95% to the collection degree of fitting of practicing of BP neural network model, BP mind is just rebuild
Through network model;If training set sample data reaches 95% to the collection degree of fitting of practicing of BP neural network model, 7 are entered step
Degree of fitting of the test set sample data to BP neural network model is tested.
Step 7: after BP neural network model buildings train, every weight has the numerical value determined, using now
" input-output " the test set sample data having tests the degree of fitting of BP neural network model, specific test method
It is as follows: network layer connection weight and mind using test set sample as input sample, and after the training adjustment by step 5
Under the conditions of first threshold values, by forward-propagating, calculated by the-the second hidden layer of the-the first hidden layer of input layer-output layer, it is defeated
Value is all label out, and the reality output label value and the label value in training sample for comparing output layer can be calculated and be trained
BP neural network model to the degree of fitting of test set sample.
Step 8: inputting sample to be predicted in the BP neural network model after degree of fitting is tested and predicted, obtained
Solid electrolyte ionic conductivity.In this example, BP neural network model reaches 95% or more test set after training
Degree of fitting, the high ionic conductivity material of necessary being and the high ionic conductivity material predicted by BP neural network almost phase
Together, illustrate that BP neural network model of the invention can be very good the ionic conductivity of prediction solid electrolyte, and process is quick
Stablize, by changing the neuron number of BP network, the data of various structures can be matched, solve existing machine study prediction
The low problem of model general applicability.
As described above, the present invention can be implemented well.