Multi-objective adaptive low-dimensional optoelectronic material design method

Abstract

This invention presents a multi-objective adaptive low-dimensional optoelectronic material design method that combines high-throughput computation of semiconductor materials with neural network learning and screening. It optimizes a crystal structure library in two steps and calculates training parameters, then inputs these parameters into a cascaded feedforward neural network for training, thereby classifying and screening target low-dimensional optoelectronic materials. This invention effectively solves the technical problems of high cost and low efficiency in high-throughput design and semiconductor crystal structure screening, achieving the goal of efficient classification and screening to obtain target low-dimensional optoelectronic materials.

Description

Technical Field

[0001] This invention belongs to the field of semiconductors, specifically relating to a multi-objective adaptive low-dimensional optoelectronic material design method. Background Technology

[0002] Currently, there is a vast amount of low-dimensional optoelectronic materials, and high-throughput design of these materials would result in high computational costs. In recent years, with the development of computer technology, machine learning has been able to perform statistical processing and analyze and classify existing features, which has promoted the development of low-dimensional optoelectronic material design methods to some extent.

[0003] Artificial neural networks are composed of numerous neurons connected by adjustable connection weights. They feature large-scale parallel processing, distributed information storage, and self-organizing and self-learning capabilities, and are increasingly widely used in fields such as information processing, pattern recognition, intelligent control, and system modeling.

[0004] Currently, researchers are dedicated to finding a method that can achieve high-throughput design and rapid screening of crystal structures for massive samples, with the advantages of high efficiency, convenience and low cost. Summary of the Invention

[0005] In view of this, the present invention provides a multi-objective adaptive low-dimensional optoelectronic material design method. This method helps to solve the problem of high-throughput design of massive samples and to rapidly screen crystal structures.

[0006] To achieve the above objectives, according to an embodiment of the present invention, the present invention provides a multi-objective adaptive low-dimensional optoelectronic material design method, which includes performing two step-by-step optimizations on a crystal structure file library and calculating training parameters, then inputting the training parameters into a cascaded feedforward neural network for training, and classifying and screening to obtain target low-dimensional optoelectronic materials.

[0007] Optionally, the method includes the following steps:

[0008] Obtain the candidate crystal data file;

[0009] The crystal data file is subjected to dimensionality reduction processing to obtain a dimensionality-reduced crystal data file;

[0010] The reduced-dimensional crystal data file undergoes two structural optimizations to obtain a stable crystal structure file. The first structural optimization fixes the atomic positions while allowing changes in the shape and volume of the unit cell. The second structural optimization maintains the shape and volume of the unit cell while optimizing the atomic positions. (This step-by-step optimization approach allows for faster and more efficient finding of stable crystal structures and enables differentiated adjustments for different systems.)

[0011] The input parameters, including band gap, effective carrier mass, carrier mobility, and material binding energy, are calculated based on the stable crystal structure file.

[0012] The input parameters are fed into a cascaded feedforward neural network, and the target low-dimensional optoelectronic materials are obtained through classification and screening.

[0013] Optionally, the crystal data file is subjected to dimensionality reduction processing to obtain a dimensionality-reduced crystal data file, including: constructing a layer along the cut crystal orientation. The vacuum layer.

[0014] Optionally, the reduced-dimensional crystal data file undergoes two structural optimizations to obtain a stable crystal structure file; wherein, the first structural optimization refers to fixing the atomic positions while allowing changes in the shape and volume of the unit cell; the second structural optimization refers to optimizing the atomic positions while keeping the shape and volume of the unit cell unchanged; including:

[0015] The reduced-dimensional crystal data file is optimized using VASP; the structure calculation parameter file is set, including the calculation function setting file (INCAR), the crystal system coordinate file (POSCAR), the pseudopotential file (POTCAR), and the reciprocal lattice vector space description file (KPOINTS).

[0016] Optionally, the reduced-dimensional crystal data file undergoes two structural optimizations to obtain a stable crystal structure file; wherein, the first structural optimization refers to fixing the atomic positions while allowing changes in the shape and volume of the unit cell; the second structural optimization refers to optimizing the atomic positions while keeping the shape and volume of the unit cell unchanged; including:

[0017] The first optimization sets the KPOINTS file to 7×7×1, and the second optimization increases the KPOINTS file setting to 13×13×1. This setting helps improve the calculation accuracy of the reciprocal space, so as to find a more stable crystal structure. The convergence criterion is that the change in the structural parameters is less than 1% of the absolute value of the structural parameters; when the convergence criterion is reached, the optimization ends.

[0018] Optionally, the reduced-dimensional crystal data file undergoes two structural optimizations to obtain a stable crystal structure file; wherein, the first structural optimization refers to fixing the atomic positions while allowing changes in the shape and volume of the unit cell; the second structural optimization refers to optimizing the atomic positions while keeping the shape and volume of the unit cell unchanged; including:

[0019] When optimizing the INCAR settings for the first time, add the Selective dynamics option to the POSCAR file and add an FFF marker after each atom in each line; the purpose is to ensure that the atom positions are fixed and only the crystal lattice parameters and volume are optimized.

[0020] During the second optimization of the INCAR settings, the F FF after each atom in the POSCAR file was changed to TTT markers. The purpose was to optimize the position of the atoms in the lattice in order to further achieve a more stable crystal structure.

[0021] Optionally, the reduced-dimensional crystal data file is subjected to two structural optimizations to obtain a stable crystal structure file; wherein, the first structural optimization refers to fixing the atomic positions while allowing changes in the shape and volume of the unit cell; the second structural optimization refers to optimizing the atomic positions while keeping the shape and volume of the unit cell unchanged.

[0022] The initial INCAR settings are as follows: ENCAT = 1.2~1.5×ENMAX; PREC = Normal; EDIFF = 1E-6; EDIFFG = -1E-2; ISTART = 0; ICHARG = 2; ISIF = 3; NSW = 200; NELM = 60.

[0023] The second INCAR parameter setting changes the ISIF parameter to ISIF=2.

[0024] Optionally, input parameters including band gap, effective carrier mass, carrier mobility, and material binding energy are calculated based on the stable crystal structure file, wherein the binding energy is calculated by the following formula:

[0025]

[0026] Among them, E b E(A) represents the binding energy of a single atom in a material. m B n A represents the total energy of the material; m Represents the energy of atom A; B n The energy of atom B is represented by m and n, respectively, which represent the number of atoms A and B.

[0027] Optionally, the input parameters are input into a cascaded feedforward neural network to classify and filter target low-dimensional optoelectronic materials, including:

[0028] After inputting the first set of parameters from the training set into the first hidden layer for training, the output result and another set of parameters are used as the input to the next hidden layer. The output result and parameters of the upper layer are used as the input to the lower layer in sequence. All parameters are input, and then the output result of each hidden layer is trained together with the original data. The classification and screening result is obtained through the output layer. Based on the classification and screening result, the target low-dimensional optoelectronic material is obtained.

[0029] Optionally, the input parameters are input into a cascaded feedforward neural network to classify and filter target low-dimensional optoelectronic materials, including:

[0030] After inputting the first set of parameters from the training set into the first hidden layer for training, the output result and another set of parameters are used as the input to the next hidden layer. The output of the previous hidden layer and the input of the next hidden layer are then combined and considered sequentially. Finally, the output results of the three hidden layers and the original data obtained from high-throughput computation are used as the input to the output layer. The classification and screening results are obtained through the output layer. Based on the classification and screening results, the target low-dimensional optoelectronic material is obtained. The size of the cascaded feedforward neural network is 3×1, and the size of the neurons in the hidden layer is 72×48×3.

[0031] The beneficial effects of this invention are:

[0032] To rapidly and efficiently design low-dimensional materials, this invention combines high-throughput computation with neural network learning and screening. First, a crystal structure library is optimized twice in stages. Multiple staged structure optimization calculations are then performed to quickly and effectively find stable crystal structures, with differentiated adjustments made for different systems. Based on the optimization results, training parameters are further calculated and then input into a self-developed cascaded feedforward neural network for training. This cascaded feedforward neural network, applied in the field of semiconductor material screening, possesses advantages such as fast learning speed, high degree of freedom in the number and depth of neurons, the ability to maintain the original structure after changes in the training set, and the elimination of the need for backpropagation of error signals. It can efficiently classify and screen target low-dimensional optoelectronic materials. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the cascaded feedforward neural network structure in a multi-objective adaptive low-dimensional optoelectronic material design method according to Embodiment 1 of the present invention.

[0034] Figure 2 : This is a flowchart of a multi-objective adaptive low-dimensional optoelectronic material design method in Embodiment 1 of the present invention. Detailed Implementation

[0035] The embodiments described are provided to better illustrate the present invention, but are not intended to limit the scope of the invention to the embodiments described. Non-essential improvements and adjustments made to the embodiments by those skilled in the art based on the above description are still within the scope of protection of the present invention.

[0036] The present application will now be described in detail with reference to the accompanying drawings, which will enable those skilled in the art to readily implement the invention. However, the invention can be implemented in many different ways and is not limited to the exemplary embodiments described herein. The endpoints and any values ​​of the ranges disclosed herein are not limited to the precise ranges or values, and these ranges or values ​​should be understood to include values ​​close to these ranges or values. For numerical ranges, the endpoint values ​​of the various ranges, the endpoint values ​​of the various ranges and individual point values, and individual point values ​​can be combined with each other to obtain one or more new numerical ranges, which should be considered as specifically disclosed herein.

[0037] The objectives and advantages of the exemplary embodiments of the present invention will become clearer and more apparent from the description below, but the objectives and advantages of the exemplary embodiments of the present invention are not limited to the following description. In the description of the exemplary embodiments, detailed explanations of the prior art will be omitted where it is believed that such detailed explanations might unnecessarily obscure the focus of the present invention.

[0038] Basic principle explanation:

[0039] This invention combines a high-throughput computing module and a neural network module to perform large-scale, high-speed, and efficient screening of crystal structure files from different crystal systems. Multi-objective refers to the ability to perform calculations on more objectives (a large number of samples) through machine learning algorithms, thus enabling high-throughput computing; adaptive refers to a more intelligent algorithm that can adjust matching conditions to obtain more reasonable results. The high-throughput computing module employs a two-step optimization approach: the first optimization fixes atomic positions while allowing changes in the shape and volume of the unit cell; the second optimization maintains the shape and volume of the unit cell while optimizing the atoms. This process allows for faster and more efficient identification of stable crystal structures and enables differentiated adjustments for different systems. Once the crystal structure has converged and stabilized after optimization (e.g., ...), ... Figure 2 As shown, unstable structures are discarded, and only the parameter values ​​of stable structures are calculated. Input parameters, including band gap, effective carrier mass, mobility, and material binding energy, are calculated based on the stable crystal structure file. These selected parameters can be accurately calculated in subsequent high-throughput calculations and, through training with a self-developed neural network, provide good classification and selection results. The self-developed neural network has a cascaded feedforward neural network structure. This structure has advantages such as fast learning speed, high degree of freedom in the number and depth of neurons, the ability to maintain the original structure after changes in the training set, and the elimination of the need for backpropagation of error signals. It can efficiently adapt to the input of multiple parameters and, through learning, allocate the weights of the parameters corresponding to different hidden layers. Then, the target low-dimensional optoelectronic material is obtained through classification and selection using the trained neural network.

[0040] Example 1, as Figure 1-2 As shown, taking a multi-objective adaptive low-dimensional optoelectronic material structure search and design method as an example, the specific implementation of the present invention is explained:

[0041] Step S1: Obtain candidate crystal data files: Use the Materials Project crystal database to obtain crystal structure files of different crystal systems, and use VESTA software to convert the structure data files into .vasp files;

[0042] Crystal structure files for different crystal systems were obtained using the Materials Project crystal database. All the crystal structures provided in the database are stable ground-state compound crystal structures.

[0043] The crystal data file is dimensionality reduced by cutting along the crystal orientation using Material Studio software. When constructing the low-dimensional crystal structure, a layer should be built on the material surface along the cut crystal orientation. The vacuum layer. The low-dimensional crystal structure data file was converted into a .vasp file using VESTA software.

[0044] Step S2: Set up VASP structure calculation parameter files, including calculation function settings file (INCAR), crystal system coordinate file (POSCAR), pseudopotential file (POTCAR), and reciprocal lattice vector space description file (KPOINTS), and optimize the crystal structure;

[0045] Analyze the material from the pseudopotential file (POTCAR) to obtain the ENMAX parameter, and set the ENCUT parameter in the INCAR to 1.2 to 1.5 times that of ENMAX.

[0046] First-principles structural optimization calculations were performed on the initially constructed low-dimensional optoelectronic material to obtain stable crystal structure information, including atomic coordinates and lattice constant data. The direct optimization method was then selected to optimize the structural information of the low-dimensional optoelectronic material.

[0047] The optimization process requires the use of two-dimensional material optimization methods, which means that a fixed vacuum layer is needed for optimization.

[0048] The specific settings are as follows:

[0049] The structural optimization was carried out in two stages. The first optimization used a K-point of 7×7×1, and the second optimization increased the K-point to 13×13×1. The first optimization kept the atomic positions in the unit cell fixed, allowing only the shape and volume of the unit cell to change.

[0050] The first step is to set the INCAR parameters as follows:

[0051] ENCAT = 1.2 to 1.5 × ENMAX;

[0052] PREC = Normal;

[0053] EDIFF = 1E-6;

[0054] EDIFFG = -1E-2;

[0055] ISTART = 0;

[0056] ICHARG = 2;

[0057] ISIF = 3;

[0058] NSW = 200;

[0059] NELM = 60.

[0060] Additionally, the Selective dynamics option needs to be added to the POSCAR file, and an FFF marker needs to be added after each atom in the line, which means the position of the atom is fixed.

[0061] The second step of optimization involves the relaxation of ion positions, which means optimizing the atoms while keeping the size and shape of the crystal material unchanged.

[0062] In the second step, the INCAR settings parameters need to be modified to ISIF=2, and the FFF after each atom in the POSCAR file needs to be changed to TTT, meaning the fixed atom position is released. Then, open the KPOINTS file and modify the number of K points to 13×13×1. The convergence criterion is that the change in structural parameters is less than 1% of the absolute value of the structural parameters. During the calculation, a PBE exchange-correlated functional is used.

[0063] Obtain the final optimized crystal structure file with a stable configuration from the CONTCAR file in the optimization calculation output file. After the structure optimization calculation converges, the subsequent calculations will begin.

[0064] Step S3: Obtain the final optimized crystal structure file with a stable configuration from the CONTCAR file in the structure optimization calculation output file;

[0065] Step S4: Perform first-principles calculations on the optimized crystal structure with a stable configuration to obtain relevant parameters such as the material's band gap, effective carrier mass, mobility, and binding energy.

[0066] In this embodiment, the final optimized structure file is subjected to first-principles electronic structure calculations using the VASP simulation software package based on plane wave method density functional theory.

[0067] Calculations and analyses yielded relevant parameters of the material, including band gap, effective carrier mass, mobility, and binding energy. A composition-structure-performance parameter database was then established based on the calculation results. The binding energy of the material is calculated using the following formula:

[0068]

[0069] In the formula E b E(A) represents the binding energy of a single atom in a material. m B n A represents the total energy of the material; m Represents the energy of atom A; B n The energy of atom B is represented by m and n, respectively, which represent the number of atoms A and B.

[0070] Step S5 uses the calculated relevant parameters as input to the cascaded feedforward neural network, and randomly divides the calculated relevant parameters into three parts: training set, validation set, and test set.

[0071] The ratio of training set, validation set, and test set is 70%, 15%, and 15%, respectively. The training method used is the Scaled Conjugate Gradient method.

[0072] The photoelectric properties of materials are classified using a three-layer neural network. The input parameters are the material's bandgap width, carrier mobility, and effective mass (the number of hidden layers can be varied depending on the number of parameters; three layers are used as an example here). The neural network is a cascaded feedforward neural network, specifically consisting of three hidden layers. During training, information is cascaded and passed between the three layers. The first set of parameters is trained in the first hidden layer. The output of the first hidden layer is used as input to the second set of parameters for training the second hidden layer. Similarly, the third set of parameters, along with the outputs of the first and second hidden layers, is used to train the third hidden layer. Finally, the outputs of the three hidden layers, along with the original data, are used as input to the output layer, which then outputs the final classification result.

[0073] The above-described structural classification has the following advantages: First, compared to training the three parameters separately and then cascading them, the complexity of the network can be greatly reduced, and the computational resource complexity will also be reduced exponentially. Therefore, by combining the output of the previous hidden layer with the input of the next hidden layer, the complexity of a multi-layered neural network can be simplified to the complexity of a single layer. Second, to avoid ignoring information from the original data during training, the original data, along with the results of the three hidden layers, is used as the input to the output layer. This prevents the neural network system from forgetting the original data and avoids misjudging the final output result.

[0074] A 3×1 cascaded feedforward neural network was initialized, with the neuron sizes of the three hidden layers being 72, 48, and 3 respectively. The neural network system contains a total of 10368 neurons. The constructed cascaded feedforward neural network is as follows: Figure 1 As shown.

[0075] The MATLAB code for this neural network is as follows:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081] The training set is used to train the cascaded feedforward neural network, serving as the data samples for model fitting. The weight parameters of each layer in the network are adjusted by training the network on the training set.

[0082] To prevent the cascaded feedforward neural network from overfitting, a validation set is used to validate the neural network.

[0083] Finally, a test set was selected to test the network's classification performance.

[0084] The predicted classification results are as follows:

[0085] Label describe 0 Inefficient 1 medium efficiency 2 High efficiency

[0086] Among them, high efficiency indicates that the material's properties are very suitable for making photoelectric detectors, medium efficiency indicates that the material's properties are suitable for making photoelectric detectors, and low efficiency indicates that the material's properties are not suitable for making photoelectric detectors.

[0087] The method described in the above embodiments has the following advantages:

[0088] 1. It can construct multi-objective phase diagrams of the chemical composition, structure and function of low-dimensional optoelectronic materials, which is beneficial for establishing a database of composition-structure-performance parameters.

[0089] 2. To address the issue of high computational costs for high-throughput design of massive samples, an empirical model potential is introduced and combined with first-principles methods based on quantum mechanics to realize an adaptive structure search strategy.

[0090] 3. Utilize machine learning to achieve high-throughput design on massive samples and complete rapid screening of crystal composition, structure, and performance.

[0091] In summary, this invention, by constructing a multi-objective phase diagram of the chemical composition, structure, and function of low-dimensional optoelectronic materials, facilitates the efficient establishment of a composition-structure-performance parameter database, enabling high-concurrency screening of massive amounts of low-dimensional materials. It has the advantages of high efficiency, convenience, and low cost.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-objective adaptive low-dimensional optoelectronic material design method, characterized in that: The process involves two steps: optimizing the crystal structure library in two stages and calculating training parameters. These training parameters are then input into a cascaded feedforward neural network for training. The target low-dimensional optoelectronic materials are then classified and screened. The process includes the following steps: Obtain the candidate crystal data file; The crystal data file is subjected to dimensionality reduction processing to obtain a dimensionality-reduced crystal data file; The reduced-dimensional crystal data file is subjected to two structural optimizations to obtain a stable crystal structure file. The first structural optimization refers to fixing the atomic positions while allowing changes in the shape and volume of the unit cell. The second structural optimization refers to optimizing the atomic positions while keeping the shape and volume of the unit cell unchanged. The input parameters, including band gap, effective carrier mass, carrier mobility, and material binding energy, are calculated based on the stable crystal structure file. The input parameters are fed into a cascaded feedforward neural network, and the target low-dimensional optoelectronic materials are obtained through classification and screening.

2. The multi-objective adaptive low-dimensional optoelectronic material design method as described in claim 1, characterized in that, The crystal data file is subjected to dimensionality reduction processing to obtain a dimensionality-reduced crystal data file, including: constructing a 10Å~20Å vacuum layer along the cut crystal orientation.

3. The multi-objective adaptive low-dimensional optoelectronic material design method according to claim 1, characterized in that, The reduced-dimensional crystal data file undergoes two structural optimizations to obtain a stable crystal structure file. The first structural optimization fixes the atomic positions while allowing changes in the shape and volume of the unit cell. The second structural optimization maintains the shape and volume of the unit cell while optimizing the atomic positions. This includes: The reduced-dimensional crystal data file was optimized using VASP; the structure calculation parameter files included the calculation function setting file INCAR, the crystal system coordinate file POSCAR, the pseudopotential file POTCAR, and the reciprocal lattice vector space description file KPOINTS.

4. The multi-objective adaptive low-dimensional optoelectronic material design method as described in claim 1, characterized in that, The reduced-dimensional crystal data file undergoes two structural optimizations to obtain a stable crystal structure file. The first optimization fixes the atomic positions while allowing changes in the shape and volume of the unit cell. The second optimization maintains the shape and volume of the unit cell while optimizing the atomic positions. This includes: The first optimization sets the KPOINTS file to 7. 7 1. The second optimization increases the KPOINTS file setting to 13. 13 1. The convergence criterion is that the change in the structural parameters is less than 1% of the absolute value of the structural parameters; the optimization ends when the convergence criterion is met.

5. The multi-objective adaptive low-dimensional optoelectronic material design method as described in claim 1, characterized in that, The reduced-dimensional crystal data file undergoes two structural optimizations to obtain a stable crystal structure file. The first structural optimization fixes the atomic positions while allowing changes in the shape and volume of the unit cell. The second structural optimization maintains the shape and volume of the unit cell while optimizing the atomic positions. This includes: When optimizing the INCAR settings for the first time, add the Selective dynamics option to the POSCAR file and add the FFF marker after each atom in the line. During the second optimization of the INCAR settings, the FFF after each atom in the POSCAR file was changed to the TTT marker.

6. The multi-objective adaptive low-dimensional optoelectronic material design method as described in claim 5, characterized in that, The initial INCAR settings are as follows: ENCAT = 1.2~1.5 ENMAX; PREC = Normal; EDIFF = 1E-6; EDIFFG = -1E-2; ISTART = 0; ICHARG = 2; ISIF = 3; NSW = 200; NELM = 60; The second INCAR parameter setting changes the ISIF parameter to ISIF=2.

7. The multi-objective adaptive low-dimensional optoelectronic material design method as described in claim 1, characterized in that, Based on the stable crystal structure file, input parameters including band gap, effective carrier mass, carrier mobility, and material binding energy are calculated, wherein the binding energy is calculated by the following formula: in, It represents the binding energy of a single atom in a material. This represents the total energy of the material; This represents the energy of atom A; The energy of atom B is represented by m and n, respectively, which represent the number of atoms A and B.

8. The multi-objective adaptive low-dimensional optoelectronic material design method as described in claim 1, characterized in that, The input parameters are fed into a cascaded feedforward neural network to classify and filter target low-dimensional optoelectronic materials, including: After inputting the first set of parameters from the training set into the first hidden layer for training, the output result and another set of parameters are used as the input to the next hidden layer. The output result and parameters of the upper layer are used as the input to the lower layer in sequence. All parameters are input, and then the output result of each hidden layer is trained together with the original data. The classification and screening result is obtained through the output layer. Based on the classification and screening result, the target low-dimensional optoelectronic material is obtained.

9. The multi-objective adaptive low-dimensional optoelectronic material design method as described in claim 1, characterized in that, The input parameters are fed into a cascaded feedforward neural network to classify and filter target low-dimensional optoelectronic materials, including: After inputting the first set of parameters from the training set into the first hidden layer for training, the output result and another set of parameters are used as the input to the next hidden layer. The output of the previous hidden layer and the input of the next hidden layer are then combined and considered sequentially. Finally, the output results of the three hidden layers and the original data obtained from high-throughput computation are used as the input to the output layer. The classification and screening results are obtained through the output layer. Based on the classification and screening results, the target low-dimensional optoelectronic material is obtained. The size of the cascaded feedforward neural network is 3×1, and the size of the neurons in the hidden layer is 72×48×3.

Images