A design method for nanowire materials with giant resistance effect based on multi-head self-attention deep learning

Abstract

The present invention discloses a method for designing nanowire materials with giant resistance effect based on multi-head self-attention deep learning, comprising: for crystal structures in a material database, based on quotient graph representation, encoding the crystal structures into character strings through an effective matching method to obtain crystal training data; inputting a Transformer model, and based on a multi-head self-attention mechanism, deeply learning the grammatical features of crystal encoding to obtain a universal crystal neural network; selecting nanowire material data with giant resistance effect that meet set conditions to form a data set, inputting a universal crystal neural network, learning topological and chemical composition features, and fine-tuning the universal crystal neural network to obtain a specialized crystal neural network; using the trained specialized crystal neural network to generate a character string that meets the set conditions; reverse decoding to reconstruct the crystal structure; and performing multi-step screening to obtain nanowire materials with giant resistance effect that meet the uniqueness of chemical components, structural novelty, stability, and band gap range.

Description

Technical Field

[0001] The present invention relates to a material design method based on deep learning in the computer field, and in particular to a giant resistance effect nanowire material design method based on multi-head self-attention deep learning. Background Art

[0002] The current silicon-based MEMS membrane is the core component of a variety of sensor sensitive elements. With the advancement of science and technology, the requirements for sensor sensitivity and performance are getting higher and higher. In order to obtain higher sensitivity, researchers have begun to explore the use of new nanowire materials for design. Nanowire materials have excellent electrical and giant piezoresistive effects, which can significantly improve the sensitivity of sensors, while taking into account the miniaturization and high integration of devices. However, the traditional laboratory method of preparing new materials is time-consuming and resource-intensive, and it is difficult to quickly meet market demand. At the same time, traditional numerical simulation methods are inefficient in dealing with their complex structures and characteristics, and lack standardized processes. Therefore, the use of generative deep learning and inverse crystal coding system methods provides new ideas and possibilities.

[0003] At present, there is no method in the world that combines generative deep learning and reversible crystal coding system to study new materials of giant resistance nanowires.

[0004] It is important to note that when using deep learning methods based on multi-head self-attention, it is necessary to choose appropriate architectures, hyperparameters, and data preprocessing methods to ensure the accuracy and reliability of the model. In addition, complex problems in model training may require large-scale computing resources, so high-performance computing and distributed computing resources need to be considered. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a design method for giant resistance effect nanowire materials based on multi-head self-attention deep learning, which combines generative deep learning and reversible crystal coding system to study the method of new giant resistance effect nanowire materials.

[0006] In order to achieve the above object, the present invention proposes a method for designing nanowire materials with giant resistance effect based on multi-head self-attention deep learning, which is used to discover new nanomaterials. The method comprises:

[0007] Step 1) for the crystal structure in the material database, a quotient graph-based representation is used, and the crystal structure is encoded into a string through an effective matching method to obtain crystal training data;

[0008] Step 2) Input the crystal training data into the Transformer model, and based on the multi-head self-attention mechanism, deeply learn the grammatical features of the crystal encoding to obtain a universal crystal neural network;

[0009] Step 3) selecting giant resistance effect nanowire material data that meets the set conditions from the material database to form a data set, inputting the universal crystal neural network, learning the topological and chemical composition characteristics, fine-tuning the universal crystal neural network, and obtaining a specialized crystal neural network;

[0010] Step 4) using the trained specialized crystal neural network to generate a character string that meets the set conditions;

[0011] Step 5) Reverse decoding is achieved through graph theory topology, general force field structure optimization and restructure refinement to reconstruct the crystal structure;

[0012] Step 6) The decoded crystal structure is subjected to multi-step screening to obtain a giant resistance effect nanowire material that meets the uniqueness of chemical components, structural novelty, stability and band gap range.

[0013] Preferably, the step 1) comprises:

[0014] Parse crystal structures as structure objects from crystallographic information files in material databases;

[0015] Construct structure graphs based on structure objects using the effective coordination number method;

[0016] Extract chemical composition, bond connectivity, and translation vectors from structure graphs to generate corresponding strings.

[0017] Preferably, the string includes: atomic symbols, node indexes and edge labels; wherein,

[0018] The string starts with an atomic symbol, which is used to represent the chemical composition of the crystal structure; each chemical bond consists of 5 characters, and the edge is explicitly represented in the form of "uvabc", where "uv" is the node index and "abc" indicates the position of the cell to which it is connected. The edge label is used to define the characteristics and specifies the translation periodicity of the edge.

[0019] Preferably, the setting conditions in step 3) include simultaneously satisfying the following conditions:

[0020] (1) There are 1-5 atoms in the unit cell;

[0021] (2) Formation energy E form <0;

[0022] (3) containing atoms with an atomic number not exceeding 86;

[0023] (4) No low-dimensional structural units

[0024] (5) Band gap satisfy:

[0025] (6) Piezoelectric modulus d>0.1C / m 2 .

[0026] Preferably, the step 1) further includes performing data enhancement on the generated character string using a randomization method, so that the atomic order and edge order in the character string can be arranged arbitrarily.

[0027] Preferably, the step 5) comprises:

[0028] Step 5-1) Generate an initial guess structure using a topology-based method based on graph theory;

[0029] Step 5-2) Predicting the optimized geometry with physicochemical significance based on the modified general force field GFN-FF;

[0030] Step 5-3) Structural refinement of the geometry-optimized structure is performed using a universal material interatomic potential based on a graph neural network and including three-body interactions.

[0031] Preferably, the step 5-1) comprises:

[0032] The string is converted into the corresponding labeled graph G(V,E), which contains n nodes and m edges, v i =(s i )∈V,v i Represents an atomic symbol s i Node i, e j =(u j ,v j ,x j ,y j ,z j )∈E,e j Indicates that the node u in the original cell j Connect to via x j a+y j b+z j Node v in the cell after c is moved j , where a, b, c are lattice basis vectors; V represents the point set, and E represents the edge set;

[0033] Based on the graph theory method, according to the given e j , calculate the metric tensor Z and the edge vector Ω expressed in cell fractional coordinates, satisfying the following formula:

[0034]

[0035] where Z consists of the dot product of the lattice basis vectors and Ω is the eutectic lattice vector L * The function of L * is the embedded lattice vector, quantifying its total deviation from the original centroid embedding; when L *When the matrix is ​​zero, the combination of Z and Ω(0) defines the barycentric embedding of G in Cartesian coordinates. Z and Ω(L * ) defines a non-barycentric Cartesian embedding of G.

[0036] Preferably, the step 5-2) comprises:

[0037] The modified general force field GFN-FF takes the adjacency list of the finite cluster constructed based on the labeled quotient graph as input and outputs the equilibrium bond length of the finite cluster. Balanced bond angle and the associated force constants;

[0038] Construct the inner product matrix g of the edge vectors as follows:

[0039] g=Ω(L * )·Z·(Ω(L * )) t

[0040] g diagonal elements g jj Represents edge l j The square length satisfies the following formula: g jj =(l j ) 2 , the off-diagonal elements g jk Contains edge l j and l k The angle component θ between jk , satisfying the following formula: g jk = l j l k cosθ jk ;

[0041] The target matrix T is constructed as and The objective function O is defined as follows:

[0042] O=∑(T jk -g jk ) 2

[0043] By manipulating Z and L * By minimizing O, chemically meaningful non-gravity-centered embedding is obtained.

[0044] Preferably, the step 5-3) comprises:

[0045] Z and L obtained in step 5-2) * The optimized structure was further refined using a universal material interatomic potential based on a graph neural network and including three-body interactions to obtain the IAP refined structure.

[0046] Preferably, the step 6) comprises:

[0047] Remove candidate materials that already exist in the material database;

[0048] Select the crystal material with the highest symmetry among the candidate materials with the same composition;

[0049] Selecting candidate materials that have a different structure from the crystal training data in step 1) and meet a set threshold;

[0050] Use atomic wire graph neural networks to evaluate band gaps and eliminate candidate materials that do not meet semiconductor characteristics;

[0051] The remaining candidate materials are subjected to structural relaxation and band structure verification at the PBE level to obtain giant resistive effect nanowire materials that meet the uniqueness of chemical components, structural novelty, stability and band gap range.

[0052] Compared with the prior art, the advantages of the present invention are:

[0053] 1. Three-step reconstruction of crystal structure: Previous methods have not yet achieved the reversibility and invariance of crystal reconstruction at the same time. The crystal coding system used in the present invention realizes the crystal representation based on the string, and realizes the reconstruction of the string to the crystal structure through a three-step process, which meets the reversibility and invariance.

[0054] 2. Reduce experimental resource requirements: Compared with traditional laboratory methods for preparing new materials, the method of the present invention provides a new way to generate the crystal structure of new materials, which can effectively reduce time and resource consumption and quickly meet market demand.

[0055] 3. Extension to other application fields: Although the method of the present invention takes the giant resistance effect nanowire new material as an example, similar methods can be extended to other fields, such as direct bandgap semiconductor new materials, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is the overall embodiment of the present invention;

[0057] Figure 2 It is a schematic diagram of the representation principle of the crystal coding string of the present invention;

[0058] Figure 3 This is a case process diagram of the present invention using the NdSiRu corresponding string to reconstruct the crystal structure in three steps;

[0059] Figure 4 This is a workflow diagram of the invention for generating a new material of giant resistance effect nanowire based on Transformer training and crystal encoding system;

[0060] Figure 5The present invention provides two promising candidate crystal structures of nanowire materials with giant piezoresistive effect and their band gaps, E hull and PBE band structure. DETAILED DESCRIPTION

[0061] This invention provides an innovative method for studying giant piezoresistance nanowire materials by combining generative deep learning methods and reversible crystal coding systems, aiming to discover new nanowire materials. Nanowire materials have excellent electrical and giant piezoresistance effects, which can significantly improve the sensitivity of sensors, taking into account the miniaturization and high integration of devices. Therefore, efficient and novel research methods make it possible to discover new materials.

[0062] The key technical feature of the present invention is that it combines a generative deep learning method based on multi-head self-attention to generate a character string that conforms to the giant resistance effect nanowire material, and reversely solves the crystal structure through a crystal coding system, maintaining reversibility and invariance.

[0063] It is important to note that when using deep learning methods based on multi-head self-attention, it is necessary to choose appropriate architectures, hyperparameters, and data preprocessing methods to ensure the accuracy and reliability of the model. In addition, complex problems in model training may require large-scale computing resources, so high-performance computing and distributed computing resources need to be considered.

[0064] The general embodiment of the present invention is as follows Figure 1 shown.

[0065] The following is a detailed explanation starting from a single unit. The crystal encoding system of the present invention first realizes the encoding of the crystal structure into a string. The crystal has an infinite periodic three-dimensional structure, so only one repeating unit of the crystal, i.e., the unit cell, is considered to represent the crystal graph. For this purpose, a finite graph representation method of an infinite periodic structure is introduced, which is called the quotient graph. Figure 2 In the quotient graph, atoms and chemical bonds in a crystal structure unit cell are represented by nodes and edges, respectively. The quotient graph itself cannot provide a unique representation of the crystal structure. In order to ensure a one-to-one correspondence between the crystal structure and its quotient graph, labels representing the periodicity and direction of the translation are added to the edges of the quotient graph to represent the offset vector.

[0066] The string representation consists of three parts. First, the string begins with the atomic symbol in the cell (see Figure 2 ), encoding the chemical composition of the corresponding crystal structure. Each chemical bond then consists of 5 characters. Edges are explicitly represented in the form of "uva bc", where "uv" is the node index and "abc" indicates the location of the cell it is connected to. Essentially, the edge label specifies the translation vector required to connect the cells. For example, in Figure 2In the figure, the label of edge e4 is "001", which means that e4 connects node A0 to the unit cell on the periodic mirror of A1 translated by one unit along the z-axis. Edge labels are defining features that specify the translational periodicity of edges, thereby enabling the construction of a suitable initial guess structure derived from graph theory. Therefore, a string representation of the crystal structure is successfully constructed through three parts: encoding element symbols, node index numbers, and edge labels. In particular, in order to eliminate the ambiguity between node indices and edge labels in the string representation, "o", "+" and "-" are used to represent "0", "1" and "-1" in edge labels, respectively. This encoding ensures that the '0' and '1' in the string refer specifically to the node index number, avoiding the appearance of "0" and "1" in both the node index number and the edge label, eliminating potential confusion during model training, and making it easier to understand the different meanings of these different parts when learning the model.

[0067] In addition, encoding all atoms within a unit cell in the system eliminates the need to introduce crystal symmetry groups and simplifies the construction rules. Although this leads to a less compact representation, this trade-off is reasonable because the generative deep learning method based on multi-head self-attention can handle long sequence strings well, making compactness less important. Due to its simple and clear definition, the crystal encoding system can be effectively applied to the chemical description and informatics of solid-state materials.

[0068] The next step is to encode the specific crystal structure into a string. In order to analyze the connectivity of the crystal structure and identify the crystallographic neighbor environment, the effective coordination number method (EconNN) is used, which provides a relatively good compromise between speed, accuracy, and robustness to atomic perturbations. Therefore, the analysis of the chemical environment is performed by defining the edges (bonds) of the string using the effective coordination number algorithm implemented in a high-throughput materials computing package (Pymatgen). Specifically, encoding the crystal structure into a string includes three steps: (1) parsing the crystal structure from the crystallographic information file into a structure object; (2) constructing a structure graph based on the structure object using the effective coordination number algorithm; (3) extracting the chemical composition, bond connectivity, and translation vector from the structure graph to generate the corresponding string.

[0069] The encoding of the crystal structure into the corresponding string, followed by a three-step reconstruction scheme from the string to the crystal structure, is implemented. An example of the reconstruction process for NdSiRu is Figure 3 As shown. It includes the following steps:

[0070] (I) Generate an initial guess structure using a topology-based approach based on graph theory

[0071] The string is converted into its corresponding labeled graph G(V,E), which contains n nodes and m edges. i =(s i)∈V represents an atomic symbol s i Node i, and e j =(u j ,v j ,x j ,y j ,z j )∈E means that the node u in the original cell is j Connect to via x j a+y j b+z j Node v in the cell after c is moved j , where a, b, c are lattice basis vectors.

[0072] Based on the graph theory method, if e is given j , we can calculate the metric tensor Z and the edge vector Ω expressed in cell fractional coordinates. The forms of Z and Ω are as follows:

[0073]

[0074] where Z consists of the dot product of the lattice basis vectors, and Ω is the co-lattice vector L * The embedded eutectic lattice vector L * quantifies the total deviation from the original centroid embedding. * When Z is a zero matrix, the combination of Z and Ω(0) defines the barycentric embedding of G in Cartesian coordinates. * ) defines a non-barycentric Cartesian embedding of G. The barycentric embedding generated by the corresponding string of the crystal structure of NdSiRu is shown in Figure 3 As shown in the structure 1.

[0075] (II) Optimization of chemically meaningful geometries based on modified GFN-FF predictions

[0076] In step (I), the centroid embedding is constructed using a purely mathematical approach without considering the node s i Therefore, the length of the lattice basis vector has no chemical meaning. For example, the center-of-gravity embedding of NdSiRu has very small lattice parameters a, b and c. Therefore, the chemical information in the string should be used to find the chemically meaningful optimized structure of the center-of-gravity embedding.

[0077] The general force field GFN-FF (Geometries, Frequencies, and Non-covalent interactions Force Field) can effectively simulate a variety of systems, from organic molecules to inorganic crystals, with atomic numbers up to 86. GFN-FF accepts Cartesian coordinates and elemental composition as input, and then generates a covalent topology encoded in a neighbor list based on interatomic distance criteria. Based on this topology, GFN-FF calculates a set of topology-based electronegativity balance (EEQ) charges q t . Subsequently, all potential energy terms can be based mainly on the neighbor list and q t The topological properties of GFN-FF allow us to develop a modified GFN-FF capable of estimating bond lengths and angles from strings.

[0078] GFN-FF is implemented in the XTB (Extended Tight Binding) package. A new module, xtb_io_reader_top, was built to accept an adjacency list as input and store the adjacency list in topo%nb. In addition, xtb_gfnff_ini, xtb_gfnff_ini2, xtb_gfnff_rab, xtb_gfnff_setup were modified to initialize newGFFCalculator using topo%nb (without relying on the Cartesian coordinates of the atoms (mol%xyz)). In addition, instead of using the Cartesian coordinates of the atoms to calculate the coordination number of the atoms (cn(i)), cn(i) is set to the normal coordination number for that atom type (param%normcn(mol%at(i))).

[0079] Since GFN-FF implemented in XTB is designed for finite systems such as molecules or clusters, we use a finite cluster (3×3×3 supercell) extracted from a crystal to simulate periodic boundary conditions. Specifically, the modified GFN-FF takes as input the adjacency list of a finite cluster (3×3×3 supercell) constructed based on the labeled quotient graph and outputs the equilibrium bond lengths / angles and the associated force constants of the finite cluster. Then, the equilibrium bond lengths / angles and the associated force constants corresponding to the central cell of the 3×3×3 supercell cluster are extracted and used to optimize the non-gravity-centered embedding in step (II) of the reconstruction process (see Figure 3 ).

[0080] In summary, we modified GFN-FF to accept s-based i and e j The constructed neighbor list takes as input and outputs the equilibrium bond length and equilibrium bond angle This modified GFN-FF is able to transform composition and connectivity into physicochemically meaningful geometries.

[0081] Equilibrium bond lengths predicted by the modified GFN-FF The barycentric embedding constructed in step (I) is used to rescale. For simplicity, this rescaled barycentric embedding of G in Cartesian coordinates is called the rescaled structure. The rescaled NdSiRu structure, such as Figure 3 Structure 2 shows a significantly improved lattice parameter.

[0082] In order to compare the bond lengths and angles of a specific non-center-of-gravity embedding with and The inner product matrix g of the edge vectors is constructed as follows:

[0083] g=Ω(L * )·Z·(Ω(L * )) t

[0084] Its diagonal elements represent the square length of the side g jj =(l j ) 2 , the off-diagonal elements contain the angle component g between sides j and k jk = l j l k cosθ jk .

[0085] The target matrix T is constructed as and The objective function is defined as follows:

[0086] O=∑(T jk -g jk ) 2

[0087] This is the sum of the squared differences between the inner products of the target matrix and the edge vectors in a particular non-barycentric embedding. * By minimizing O, a chemically meaningful non-gravity embedding of G is obtained. For simplicity, we refer to this chemically meaningful non-gravity embedding of G as ZL * Optimized structure. ZL of NdSiRu * Optimize the structure such as Figure 3 As shown in Structure 3, the bond lengths are closer to the values ​​of the original structure.

[0088] (III) Using M3GNet IAP for ZL * Optimize the structure for structural refinement.

[0089] Although GFN-FF is able to model a wide range of systems, its discontinuous potential energy surface makes it unsuitable for direct optimization of crystal structures. The universal material interatomic potential (M3GNet IAP) based on graph neural networks and including three-body interactions is able to perform high-fidelity structural optimization of materials with different chemical compositions. To this end, the ZL obtained in step (II) * The optimized structure is further refined using M3GNet IAP, and the resulting IAP refined structure (see Figure 3 The structure in step (4) is highly consistent with the original structure. However, the IAP-refined rescaled structure obtained by directly using M3GNet IAP to refine the rescaled structure in step (II) deviates greatly from the original structure, highlighting the importance of ZL for GFN-FF geometry. * The key role of optimization (step II).

[0090] In summary, through the above three-step scheme, the reconstruction of the string into the crystal structure was achieved with high performance and matching rate. Among them, the modified GFN-FF potential used in step (II) limits the applicability of the reconstruction method to crystal structures with atomic number up to 86.

[0091] The next step is to use the crystal coding system to reverse design new nanowire materials with giant resistance effect. The reverse design workflow includes four stages (such as Figure 4 (shown in Figure 3): (1) Use the Materials Project database to train a multi-head self-attention-based Transformer model to learn the grammar of crystal-encoded strings; (2) Use the giant-resistance nanowire material dataset to fine-tune the general Transformer and further develop the Transformer; (3) Use the Transformer to generate a large number of strings and reconstruct them into crystal structures; (4) Screen these crystal structures to identify new nanowire materials with giant-resistance effects.

[0092] The following design criteria were set: target band gap, piezoelectric modulus, stability, compositional novelty, and structural uniqueness. Specifically: (1) Band gap (at the Perdew-Burke-Ernzerhof (PBE) level); (2) piezoelectric modulus d>0.1C / m 2 ; (3) Energy E higher than the convex hull hull<50meV / atom; (4) the composition must be different from the entries in the Materials Project database; (5) the structure should show low similarity to the training set structures. Due to the finite temperature effect and the errors of density functional theory, it is assumed that E hull Crystals with <50meV / atom are synthetically feasible.

[0093] The discovery of new materials consists of two aspects: learning the syntax of the string representation of the crystal encoding system and learning the topological and compositional features for key properties. First, a general Transformer was trained on the "General Dataset", which includes crystal structures from the Materials Project that meet the following four conditions: (1) there are 1-5 atoms in the unit cell; (2) the formation energy E form <0; (3) contains atoms with atomic numbers not exceeding 86; (4) has no low-dimensional structural units. Then, by and piezoelectric modulus d>0.1C / m 2 The Transformer is further fine-tuned on a generic dataset of crystal structures to learn to predict the features of giant resistive nanowires.

[0094] In addition, in order to improve the performance of the crystal coding system model, when learning the topological and compositional features of grammatical features and key characteristics, randomization methods are used for data enhancement in the process of generating strings, so that the atomic order and edge order in the string can be arbitrarily arranged.

[0095] The developed Transformer model was used to generate about 50,000 strings. About 43,000 of them were decoded into crystal structures, and the remaining strings could not be reconstructed due to repeated edges. The crystal structure was then screened. First, candidate materials that already existed in the database were removed, narrowing the number of candidate materials to about 38,000. Next, in order to avoid duplication in data enhancement, the most symmetric representatives among the candidate materials with the same composition were selected, reducing the number of candidates to about 21,000. Then, the structures with a structural dissimilarity value lower than 0.75 (the threshold of the Materials Project material library) and E were removed. hull >50meV / atom candidate materials, and the atomic wire graph neural network (ALIGNN) model was used to evaluate the band gap, excluding candidate materials that were unlikely to be semiconductors, leaving 17 candidate materials. Finally, the Perdew-Burke-Ernzerhof (PBE) level structural relaxation and band structure verification were performed on the materials, and the crystal structures of 7 new nanowire materials with giant piezoresistive effect that met the design criteria were identified. The structure, band gap, E hull and the PBE band structure as Figure 5shown.

[0096] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0097] Example 1

[0098] The embodiment of the present invention proposes a method for designing nanowire materials with giant resistance effect based on multi-head self-attention deep learning. The method comprises the following steps:

[0099] Collect topological and chemical composition characteristics: Determine the topological and chemical composition characteristics of giant resistance nanowire materials for training purposes, such as band gap, piezoelectric modulus, etc.

[0100] Train the Transformer model: Train the Transformer model to learn the grammatical features of the crystal coding system string and generate a universal crystal neural network; further train the Transformer to learn the giant resistance effect nanowire material data set, generate a specialized crystal neural network, and obtain a string that meets the conditions.

[0101] Reverse decoding: The string is reversely decoded in a three-step process including graph theory topology, general force field structure optimization and restructure refinement to reconstruct the crystal structure.

[0102] Dataset processing and screening: The decoded crystal structures are screened in multiple steps to meet the criteria of uniqueness of chemical composition, structural novelty, stability, band gap range, etc.

[0103] Result analysis: Users can analyze the crystal structure of the new nanowire material to analyze the properties of the new material.

[0104] The analysis can include examining the structural similarity of the new material to existing materials and the possibility of giant drag effects.

[0105] Summary:

[0106] The crystal coding system used in the present invention can realize crystal representation based on character strings, while satisfying reversibility and invariance, and has the hope and prospect of discovering new nanowire materials.

[0107] At the same time, Transformer, a generative deep learning method based on multi-head self-attention, has attracted much attention in natural language processing and generation tasks because it can significantly improve the accuracy, efficiency and flexibility of the model.

[0108] Transformer uses a multi-head self-attention mechanism and feedforward neural networks to efficiently process sequence data and is widely used in machine translation, text generation and other fields. The multi-head self-attention mechanism can capture the relationship between different positions in the input sequence, thereby better understanding the contextual information. In addition, Transformer does not need to rely on recurrent neural networks (RNNs), so it has better performance and parallelization capabilities when processing long sequences. In this way, Transformer has become a standard model for many generative deep learning tasks.

[0109] The crystal coding system string can better learn the structure-activity relationship of the crystal structure. Therefore, combined with the generative deep learning method, after training, the model can better learn the grammatical features of the crystal coding string by learning different crystal structures. Then, through the transfer learning method, the characteristics of the nanowire crystal material with giant resistance effect are learned, and the expected string is generated in a directional manner.

[0110] In this way, the character string can be reversely decoded into the crystal material through the crystal coding system, and then through continuous correction and multi-step screening, the decoded crystal structure can be obtained, thereby realizing the prediction of new nanowire materials that meet the giant resistance effect.

[0111] Through the above steps, the method of the present invention can realize the rapid generation of potential new nanowire materials with giant resistance effect, providing engineers and researchers with the possibility of discovering new materials and crystal structures, thereby promoting the development and application of MEMS nanowire technology. This method also has a wide range of applicability and can be used to generate new semiconductor crystal structure materials under different conditions to meet the needs of various application fields.

[0112] Key technical points of this application:

[0113] 1. Crystal representation based on crystal coding system;

[0114] 2. A three-step reverse decoding method based on the crystal coding system;

[0115] 3. Design methods for generating new materials by combining Transformer and crystal coding systems.

[0116] The method crystal encoding system in the present invention can allow users to achieve reversible crystal representation to design new materials, while satisfying reversibility and invariance. In the context of generative reversible molecular representation, the research field of crystal representation methods for solid materials has made certain progress in recent years. Previous representation methods, such as crystal graph convolutional neural networks, satisfy symmetry invariance but do not satisfy reversibility; such as 3D image-based crystal representation methods, which achieve reversibility but do not satisfy symmetry invariance and have high computational costs. In short, there is no reversible crystal representation that satisfies complete invariance.

[0117] The present invention introduces a generative deep learning method based on multi-head self-attention and a crystal coding system based on strings, and applies them to discover nanowire materials with giant resistance effect. Through the generative deep learning method, a new nanowire crystal material data set that meets the giant resistance effect is trained and represented by a string. And through the crystal coding system, the reverse crystal structure representation based on the string is successfully realized, and the invariance is guaranteed. The system maintains invariance through encoding topology and combination, achieves reversibility through graph theory topology method and deep learning standpoint, and achieves wide applicability, high matching rate and reconstruction performance.

[0118] The method of the present invention can be used to generate new materials under different conditions: when training, material properties under different conditions, such as structure, material properties, band gap, etc., are given, and by training the corresponding character string through Transformer and then decoding it, possible crystal structures and new materials that meet the material properties can be generated.

[0119] 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 the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A method for designing nanowire materials with giant resistance effect based on multi-head self-attention deep learning, which is used to discover new nanomaterials, the method comprising: Step 1) for the crystal structure in the material database, a quotient graph-based representation is used, and the crystal structure is encoded into a character string through an effective matching method to obtain crystal training data; Step 2) Input the crystal training data into the Transformer model, and based on the multi-head self-attention mechanism, deeply learn the grammatical features of the crystal encoding to obtain a universal crystal neural network; Step 3) selecting giant resistance effect nanowire material data that meets the set conditions from the material database to form a data set, inputting the universal crystal neural network, learning the topological and chemical composition characteristics, fine-tuning the universal crystal neural network, and obtaining a specialized crystal neural network; Step 4) using the trained specialized crystal neural network to generate a character string that meets the set conditions; Step 5) Reverse decoding is achieved through graph theory topology, general force field structure optimization and restructure refinement to reconstruct the crystal structure; Step 6) The decoded crystal structure is subjected to multi-step screening to obtain a giant resistance effect nanowire material that meets the uniqueness of chemical components, structural novelty, stability and band gap range.

2. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 1, characterized in that: The step 1) comprises: Parse crystal structures as structure objects from crystallographic information files in material databases; Construct structure graphs based on structure objects using the effective coordination number method; Extract chemical composition, bond connectivity, and translation vectors from structure graphs to generate corresponding strings.

3. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 2, characterized in that: The string includes: atomic symbols, node indexes and edge labels; wherein, The string starts with an atomic symbol and is used to represent the chemical composition of the crystal structure; each chemical bond consists of 5 characters, and the edge is explicitly represented in the form of "uvabc", where "uv" is the node index and "abc" indicates the position of the cell to which it is connected. The edge label is used to define the characteristics and specifies the translation periodicity of the edge.

4. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 1, characterized in that: The setting conditions in step 3) include simultaneously satisfying the following conditions: (1) There are 1-5 atoms in the unit cell; (2) Formation energy E form <0; (3) containing atoms with an atomic number not exceeding 86; (4) No low-dimensional structural units (5) Band Gap satisfy: (6) Piezoelectric modulus d>0.1C / m 2 .

5. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 1, characterized in that: The step 1) also includes performing data enhancement on the generated character string using a randomization method, so that the atomic order and edge order in the character string can be arranged arbitrarily.

6. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 1, characterized in that: The step 5) comprises: Step 5-1) Generate an initial guess structure using a topology-based method based on graph theory; Step 5-2) Predicting the optimized geometry with physicochemical significance based on the modified general force field GFN-FF; Step 5-3) Structural refinement of the geometry-optimized structure is performed using a universal material interatomic potential based on a graph neural network and including three-body interactions.

7. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 6, characterized in that: The step 5-1) comprises: The string is converted into the corresponding labeled graph G(V,E), which contains n nodes and m edges, v i =(s i )∈V,v i Represents an atomic symbol s i Node i, e j =(u j ,v j ,x j ,y j ,z j )∈E,e j Indicates that the node u in the original cell j Connect to via x j a+y j b+z j Node v in the cell after c is moved j , where a, b, c are lattice basis vectors; V represents the point set, and E represents the edge set; Based on the graph theory method, according to the given e j , calculate the metric tensor Z and the edge vector Ω expressed in cell fractional coordinates, satisfying the following formula: where Z consists of the dot product of the lattice basis vectors and Ω is the eutectic lattice vector L * The function of L * is the embedded lattice vector, quantifying its total deviation from the original centroid embedding; when L * When the matrix is ​​zero, the combination of Z and Ω(0) defines the barycentric embedding of G in Cartesian coordinates. Z and Ω(L * ) defines a non-barycentric Cartesian embedding of G.

8. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 7, characterized in that: The step 5-2) comprises: The modified general force field GFN-FF takes the adjacency list of the finite cluster constructed based on the labeled quotient graph as input and outputs the equilibrium bond length of the finite cluster. Balanced bond angle and the associated force constants; Construct the inner product matrix g of the edge vectors as follows: g=Ω(L * )·Z·(Ω(L * )) t g diagonal elements g jj Represents edge l j The square length satisfies the following formula: g jj =(l j ) 2 , the off-diagonal elements g jk Contains edge l j and l k The angle component θ between jk , satisfying the following formula: g jk = l j l k cosθ jk ; The target matrix T is constructed as and The objective function O is defined as follows: O=∑(T jk -g jk ) 2 By manipulating Z and L * By minimizing O, chemically meaningful non-gravity-centered embedding is obtained.

9. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 8, characterized in that: The step 5-3) comprises: Z and L obtained in step 5-2) * The optimized structure was further refined using a universal material interatomic potential based on a graph neural network and including three-body interactions to obtain the IAP refined structure.

10. The method for designing giant resistance nanowire materials based on multi-head self-attention deep learning according to claim 1, characterized in that: The step 6) comprises: Remove candidate materials that already exist in the material database; Select the crystal material with the highest symmetry among the candidate materials with the same composition; Selecting candidate materials that have a different structure from the crystal training data in step 1) and meet a set threshold; Use atomic wire graph neural networks to evaluate band gaps and eliminate candidate materials that do not meet semiconductor characteristics; The remaining candidate materials are subjected to structural relaxation and band structure verification at the PBE level to obtain giant resistive effect nanowire materials that meet the uniqueness of chemical components, structural novelty, stability and band gap range.

Images